Supply Chain Planning

Supply Chain Planning
Hans-Otto Günther • Herbert Meyr

Editors
Supply Chain Planning

Quantitative Decision Support
and Advanced Planning Solutions
Prof. Dr. Hans-Otto Günther Prof. Dr. Herbert Meyr
TU Berlin H 95 TU Darmstadt
Department of Production Management Department of Production
Straße des 17. Juni 135 and Supply Chain Management
10623 Berlin Hochschulstr. 1
Germany 64289 Darmstadt
[email protected] Germany
[email protected]
ISBN 978-3-540-93774-6 e-ISBN 978-3-540-93775-3
DOI 10.1007/978-3-540-93775-3
Library of Congress Control Number: 2008943995
© 2009 Springer-Verlag Berlin Heidelberg
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permissions for use must always be obtained from Springer-Verlag. Violations
are liable for prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.
Cover design: WMXDesign GmbH, Heidelberg, Germany
Printed on acid-free paper
987654321
springer.com
Preface
In recent years, supply chain planning has emerged as one of the most challenging
problems in the industry. As a consequence, the planning focus is shifting from the
management of plant-specific operations to a holistic view of the various logistics
and production stages, that is an approach in which suppliers, production plants and
customers are considered as constituents of an integrated network. A major driv-
ing force behind this development lies in the globalization of the world economy,
which has facilitated the co-operation between different partners working together
in world-wide logistics networks. Hence, considerable cost savings can be gained
from optimizing the structure and the operations of complex supply networks link-
ing plants, suppliers, distribution centres and customers. Consequently, to improve
the performance of the entire logistic chain, more sophisticated planning systems
and more effective decision support are needed.
Clearly, successful applications of supply chain management have driven the
development of advanced planning systems (APS), which are concerned with sup-
porting decision-making activities at the strategic, tactical and operational decision
level. These software packages basically rely on the application of quantitative
methods, which are used to model the underlying complex decision problems con-
sidering the limited availability of resources and the need to react on time to
customer orders. The core module at the mid-term level of APS comprises oper-
ational supply chain planning. In many industries, production stages are assigned to
different plants and distribution centres have been established at geographically dis-
persed locations. Supply chain planning aims at coordinating production activities
within such multi-site logistics networks to avoid excessive inventories, inefficient
capacity utilization and poor customer service. In APS, advanced optimization tech-
niques are provided to determine the quantities to be produced, stored, transported
and procured in the supply network.
This book is divided into five parts. The first one is concerned with decision-
making in supply chains. The paper by Hartmut Stadtler addresses issues of
collaboration in inter-organizational supply chains. It analyzes and clusters the ap-
proaches that have been developed since the advent of supply chain management
about two decades ago and gives an overview on the state-of-the-art. Moreover,
the paper provides a framework for collaborative planning in supply chains, with
v
vi Preface
special emphasis on model-based decision support at the operational level of inter-

organizational supply chains.
In their paper, Christian Almeder, Margaretha Preusser and Richard F. Hartl
present a general framework for combining an optimization model and discrete
event simulation to support supply chain planning. Both models are applied in an
iterative fashion until convergence is reached. It is shown that this approach deliv-
ers competitive results much faster compared to mixed-integer linear programming
models in a stochastic environment.
Part 2 comprises five papers on demand management and order fulfillment in
supply chains. Revenue management is a topic that has received great attention
in service and recently also in manufacturing industries. In their paper, Rainer
Quante, Herbert Meyr and Moritz Fleischmann analyze the underlying supply chain
processes of revenue management and demand fulfilment in different business envi-
ronments. In particular, they discuss and clarify the relationships between available
software solutions and applications as well as scientific models in this field and
highlight directions of future research.
A systematic mathematical programming approach for active demand manage-
ment in process industries is presented by Aaron A. Levis and Lazaros Papageorgiou.
They develop an iterative algorithm for supporting decision-making on pricing
strategies as well as on output levels for substitute products. Several case studies
are used to obtain game-theoretical insights for a duopolistic market situation.
The paper by Herbert Meyr focuses on the segmentation of customers and the
allocation of available quantities. It is shown that the current practice of rule-based
order promising can be improved by exploiting information on customer hetero-
geneity and customer demand. As a practical tool, deterministic linear programming
models to support available-to-promise decisions in a make-to-stock environment
are presented.
In their paper, Thomas R. Ervolina, Markus Ettl, Young M. Lee and Daniel J.
Peters propose a novel management process for determining marketable product
alternatives in a supply chain. Their approach aims at better integrating the supply
chain horizontally by connecting the activities of customers, business partners and
sales teams to procurement and manufacturing capabilities of a firm. The proposed
mathematical optimization based approach has contributed to substantial business
improvements at IBM.
The paper by Richard Pibernik and Prashant Yadav focuses on real-time order
promising in a make-to-stock environment. An integrated approach is developed
that exploits the structure of order arrivals and material receipts to determine inven-
tory reservations for high-priority orders in real time. In a comprehensive numerical
study, the impact of inventory reservation and order promising is investigated under
varying system parameters.
The three papers included in Part 3 of this book focus on inventory management.
In supply chains, safety stocks are needed to be more responsive to customer orders
and to meet the target service levels. Youssef Boulaksil, Jan C. Fransoo and Ernico
N. G. van Halm propose a combination of an optimization and a simulation model.
They apply this hybrid modeling approach to determine safety stock levels in a
Preface vii
multi-item multi-stage inventory system. As a case study, the supply chain of a

biopharmaceutical company is considered.
The paper by Pieter L.M. van Nyen, J. Will M. Bertrand, Henny P.G. van Ooijen
and Nico J. Vandaele investigates the impact of different variants of supplier-
managed inventory on costs in a supply chain consisting of a parts supplier and an
original equipment manufacturer’s assembly plant. Numerical experiments revealed
that substantial system-wide cost savings could be gained compared to a situation
in which the assembly plant manages the inventories.
In the subsequent paper on vendor-managed inventory, Bogdan C. Bichescu and
Michael J. Fry analyze decentralized supply chains under different degrees of chan-
nel power, that is the agents’ ability to control the decision-making process in a
supply chain. Game-theoretical models are used to compare the effectiveness of
vendor-managed inventory and to analyze different channel power relationships un-
der a variety of environmental conditions.
Applications of supply chain planning in the chemical industry are presented
in Part 4. The concept of value chain management can be seen as another chal-
lenging extension of classical supply chain management. Considering a case study
from the chemical industry, Matthias Kannegiesser, Hans-Otto Günther, Paul van
Beek, Martin Grunow and Christoph Habla present an operative planning model
for coordinating sales, distribution and production activities throughout a global
value chain. Specifically, the optimization model addresses spot demand for chem-
ical commodities with volatile and uncertain sales prices and evaluates the impact
of the respective price-quantity elasticity.
An optimization model based on mixed-integer linear programming to schedule
campaigns in a specialty chemicals plant is presented by Marcus Brandenburg and
Franz-Josef Tölle. They focus on a real-world problem, which is characterized by
a variety of chemical processes with sequence-dependent setup conditions, com-
plex material flows, flexible use of resources and facility-dependent batch sizes. To
solve this complex scheduling problem, a two-stage solution procedure is applied,
which determines near-optimal schedules even for large-sized real-world problem
instances within reasonable CPU time.
Finally, two papers with applications in the automotive industry are presented
in Part 5. The paper by Herbert Meyr, originally published in 2004, first gives an
overview of short- and mid-term approaches for supply network planning in the
automotive industry and specifically discusses the application of OR methodology
to support the various planning tasks involved. Afterwards, the author discusses the
impact of the ongoing change in strategy, namely the change from a built-to-stock
oriented to a customized built-to-order production, on the future application of OR
methods.
The final paper by Ralf Bihlmaier, Achim Koberstein and René Obst considers
strategic flexibility and capacity planning under uncertain demand in production
networks of automobile manufacturers. Their solution approach is integrated into a
decision support system, which determines minimum-cost product allocations and
develops tactical workforce plans. The practicality of their approach is demonstrated
by the use of an industrial case study.
viii Preface
The primary objective of this book is to reflect the recent developments of sup-
ply chain planning and to examine new research issues. It presents recent research
results on collaborative planning in supply chains, demand and inventory manage-
ment in logistics networks as well as industrial applications. The specific focus of
this book is on the application of quantitative methods, which also form the ba-
sis of commercial advanced planning software systems. Fourteen papers previously
published in “OR Spectrum – Quantitative Approaches in Management” have been
selected for publication in this volume. All papers have been peer-reviewed accord-
ing to the standards of the journal.
This book has greatly benefited from the cooperation among the authors, review-
ers and editors. We express our sincere thanks to the reviewers for their excellent
and timely refereeing. Last, but not least, we thank all authors for their contribu-
tions, which made this book possible.
Berlin and Darmstadt, Hans-Otto Günther

February 2009 Herbert Meyr
Contents
Part I: Decision Making in Supply Chains

A framework for collaborative planning and state-of-the-art ........................... 3
H. Stadtler
Simulation and optimization of supply chains: alternative

or complementary approaches? .......................................................................... 29
C. Almeder, M. Preusser, R.F. Hartl
Part II: Demand Management

Revenue management and demand fulfillment: matching applications,
models and software............................................................................................. 57
R. Quante, H. Meyr, M. Fleischmann
Active demand management for substitute products through price

optimisation........................................................................................................... 89
A.A. Levis, L.G. Papageorgiou
Customer segmentation, allocation planning and order promising

in make-to-stock production ............................................................................. 117
H. Meyr
Managing product availability in an assemble-to-order supply

chain with multiple customer segments ........................................................... 145
T.R. Ervolina, M. Ettl, Y.M. Lee, D.J. Peters
ix
x Contents
Inventory reservation and real-time order promising

in a make-to-stock system.................................................................................. 169
R. Pibernik, P. Yadav
Part III: Inventory Management

Setting safety stocks in multi-stage inventory systems under rolling horizon
mathematical programming models................................................................. 199
Y. Boulaksil, J.C. Fransoo, E.N.G. van Halm
Supplier managed inventory in the OEM supply chain: the impact

of relationship types on total costs and cost distribution................................ 219
P.L.M. van Nyen, J.W.M. Bertrand, H.P.G. van Ooijen, N.J. Vandaele
Vendor-managed inventory and the effect of channel power ........................ 247

B.C. Bichescu, M.J. Fry
Part IV: Applications in the Chemical Industry

Value chain management for commodities: a case study
from the chemical industry................................................................................ 283
M. Kannegiesser, H.-O. Günther, P. van Beek, M. Grunow, C. Habla
MILP-based campaign scheduling in a specialty chemicals plant:

a case study.......................................................................................................... 315
M. Brandenburg, F.-J. Tölle
Part V: Applications in the Automotive Industry

Supply chain planning in the German automotive industry.......................... 343
H. Meyr
Modeling and optimizing of strategic and tactical production planning

in the automotive industry under uncertainty................................................. 367
R. Bihlmaier, A. Koberstein, R. Obst
A framework for collaborative planning
and state-of-the-art
Hartmut Stadtler
Originally published in:

OR Spectrum (2009) 31:5–30
DOI 10.1007/s00291-007-0104-5
Abstract Inter-organizational supply chain management incurs the challenge to

align the activities of all members which contribute to the value creation of a product
or service offered to customers. In general, a supply chain faces the “problem” of
information asymmetry, members having their own objectives and constraints which
may be in conflict with those of the other members. Still, activities have to be aligned
in such a way that the supply chain as a whole stays or becomes competitive while
each member wins by cooperating. A number of collaborative planning schemes have
been put forward in the last two decades with different assumptions and different
areas of application. This paper intends to provide a framework and an overview on
the state-of-the-art of collaborative planning. The criteria of the framework will allow
us to position existing concepts and to identify areas where more research is needed.
The focus of the literature reviewed here will be on model-based decision support at
the operational planning level.
Keywords Collaborative planning · Supply chain management · State-of-the-art
1 Introduction and definitions
Supply chain management (SCM) is concerned with the coordination of material,

information and financial flows within and across legally separated organizational
units (Christopher 1998). One important way to achieve coordination in an inter-orga-
nizational supply chain (SC) is the alignment of future activities of SC members,
hence the coordination of plans. The aim of this paper is to present a new framework
H. Stadtler (B)
Institute for Logistics and Transportation, University of Hamburg,
Von-Melle-Park 5, 20146 Hamburg, Germany
e-mail: [email protected]
H.O. Günther, H. Meyr, Supply Chain Planning 3

c Springer-Verlag Berlin Heidelberg 2009
4 H. Stadtler
of collaborative planning (CP) together with a state-of-the-art overview of concepts

for CP from the literature with special emphasis on a model-based decision support
at the operational planning level of an inter-organizational SC.
Coordination of flows requires adapting plans of supply chain members at various
levels of a planning hierarchy. Planning is regarded as an activity to choose, sequence
and evaluate future activities for a specific decision making unit (e.g. a company). A
procedure for aligning plans of two or more decision-making units is called a coor-
dination scheme . The terms procedure and process are regarded synonyms here for
describing the interaction of activities of at least two SC members (e.g. a negotiation
process). In order to become a CP scheme we further require that individual plans are
adapted in an effort of joint decision making, i.e. a willingness to cooperate and to
contribute to the generation of a plan which will be accepted by these SC members
(which may well be a subset of the overall SC). In other words, we exclude pure cen-
tral planning by a single (focal) SC member without the active contribution of other
SC members. This is usually required to overcome information asymmetry, where no
SC member possesses all the information and preferences of the other SC members
(see Schneeweiss 2003, p. 29 for a similar definition). This private information may be
revealed to the other SC members in the course of joint decision making provided ade-
quate incentives for true information providing exist (e.g. Feldmann and Müller 2003).
Note that for central planning at least one SC member must possess all the information
relevant for generating an overall SC plan that is accepted by all members.
In summary, we define collaborative planning as a joint decision making pro-
cess for aligning plans of individual SC members with the aim of achieving
coordination in light of information asymmetry.
But when does CP result in a state of SC coordination? The most stringent answer
refers to the contract literature: a contract coordinates the SC if (and only if) “the set of
supply chain optimal actions is a Nash equilibrium, i.e., no firm has a profitable unilat-
eral deviation from the set of supply chain optimal actions” (Cachon 2003, p. 230 see
Myerson 1991 for the definition of a Nash equilibrium). Here, coordination requires
a solution which represents both a (central) supply chain optimum as well as a Nash
equilibrium.
Omitting the game theoretic perspective the overall SC perspective remains: now,
a (central) SC optimum solution suffices for coordination. A third and even “softer”
definition of coordination results if the implemented actions lead to an improved plan
for the SC as a whole compared to a default (or initial) solution. Such a definition is
implicitly supported by Corbett and de Groote who compare their (suboptimal) coordi-
nation mechanism with the default solution (no coordination) (Corbett and de Groote
2000, p.449). Finally, a fourth alternative even calls the default situation coordinated,
which seems to be favoured by Schneeweiss (“worst-case” coordination, Schneeweiss
2003, p.278).
The first two definitions of SC coordination imply that a large number of CP schemes
will be left in a state of non-coordination although solutions generated may be near-
optimal. On the other hand, the fourth proposal will call all solutions coordinated
without looking at feasibility or solution quality. Hence, we regard the third proposal
a good compromise.
A framework for collaborative planning and state-of-the-art 5
The alignment of plans will take place at a certain planning level (e.g. master plan-
ning) with a given degree of detail (level of aggregation) and planning horizon. Hence,
CP is further specified, like collaborative master planning.
CP not only applies to a SC partnership—as described by five criteria by Landeros
and Monczka (1989) based on a survey in the automotive industry—but also to a more
competitive environment. Hence, the term SC member seems more neutral. Still, we
regard an arms-length type of interaction (see Dyer et al. 1998) as well as actions
of moral hazard to be counterproductive for an effective CP. Each SC member is in
charge of a specific planning domain. It comprises a part of the SC and the related
planning processes that are under the control and in the responsibility of a distinct SC
member (Kilger and Reuter 2005).
While intended to be applied to inter-organizational SCs, CP schemes might well
be applied in an intra-organizational setting where SC members belong to the same
company. The main requirement is that CP takes place in the absence of a central
planning instance which may ultimately enforce coordination. If the coordination of
activities (like transport or production activities) is achieved across different SCs this is
called horizontal collaboration (for an example in the distribution of consumer goods
see Fleischmann 1999). However, if the activities considered belong to one single SC
this requires vertical coordination—which will be the focus of this paper.
CP software modules are already offered by some software vendors which basi-
cally support the exchange of (demand and procurement) data plus some additional
feasibility checks. According to Schneeweiss (2003, p. 5) this is only the starting point
of CP.
In the following we will present an overview of research areas where CP schemes
have originated from (Sect. 2). Sections 3, 4 and 5 describe our framework for CP.
Section 6 discusses the issue of fairness which seems to be an important condition for
engaging in and accepting a CP scheme. Finally, Sect. 7 summarizes our findings and
provides some ideas for future research.
2 Related research areas
There are a number of research areas which are closely linked to CP in that these
provide mechanisms to coordinate decentralized decision units. These areas will be
mentioned briefly below with their main focus and limitations. These areas will be dis-
cussed sequentially starting with the area with the largest number of papers considered
here (see Table 1): mathematical decomposition. Within this broad area we discrim-
inate exact mathematical decomposition, heuristic mathematical decomposition and
meta-heuristics.
In principle (exact) mathematical decomposition techniques are applicable for CP
(like Dantzig–Wolfe decomposition (Dantzig and Wolfe 1960) or Benders decompo-
sition (Benders 1962) to name only a few). These techniques usually require a specific
structure of the underlying decision model (coefficient matrix) and aim at finding
an optimal solution to the overall (central) model with less computational efforts.
Mathematical decomposition can be interpreted as a model of a divisionalized orga-
nization with individual and hidden constraints at the divisions as well as common
6
Table 1 Papers applied to the framework for CP and associated research areas
Related research Lu Corbett Barbar- Ertogral Fransoo Gjerdrum Karabuk Fink Schneeweiss Sucky Dudek and Jung Shirod-
areas (1995) and de osoĝlu and Wu et al. et al. (2002) and Wu (2003, and Zimmer (2004b) Stadtler et al. kar and
Groote (2000) (2000) (2001) (2002) 2004) (2004) (2005, 2007) (2005) Kempf
(2000) (2006)
Exact mathematical X
decomposition
Heuristic X X X
mathematical
decomposition
Meta-heuristics X
EOQ models/ X X X
simple contracts
Inventory systems X
Hierarchical X
Planning
Paper fulfils No No No
CP definition (central (central (central
model) model) model)
H. Stadtler
constraints at the organizational level. The organizational level serves as a central deci-
sion unit, which knows the interdependencies among divisions and guides the search
for the organization’s optimal solution by a price-budget directive planning procedure
(Meijboom 1986). Limitations of this area have been lifted in recent years (e.g.
Vanderbeck 2000). A paper originating from this area is Karabuk and Wu (2002)
making use of Augmented Lagrangean techniques (Table 1).
Similar to the Dantzig–Wolfe decomposition heuristic mathematical decomposi-
tion techniques are based on individual mathematical programming models for each
SC member comprising the constraints and objectives of their planning domain. These
models may be used to generate and evaluate purchase and supply proposals (for an
example see Dudek 2004, Dudek and Stadtler 2005, 2007). The models’ decisions are
aligned by heuristic search techniques with the aim of finding a feasible, near optimal
solution for the overall SC (e.g. Jung et al. 2005). CP schemes differ in the type of
information to be exchanged. Here, resource- or price-related information can be made
use of (e.g. Ertogral and Wu 2000). Sometimes mathematical programming models
are solved by a standard MIP solver or by a meta-heuristic (like in Fink 2003, 2004)
(all included in Table 1).
Coordination of static lot-sizing models [also known as economic order quantity
(EOQ) models] in a two party setting (usually termed buyer and vendor) have a long
history (starting with Goyal 1976 and Monahan 1984) and provide a number of valu-
able insights (e.g. how to devise a quantity discount pricing schedule that results in
the SC optimal solution). Recently, research has been extended to three-level supply
chains. Reviews can be found in Thomas and Griffin (1996), Sarmah et al. (2006)
and Li and Wang (2007). They state that research has been limited to “. . .one product
and one machine only and thereby fails to capture the essence of real supply chain”
(Sarmah et al. 2006, p. 13). Papers that originate from this research area and assume
asymmetric information and thus fit into this paper (see Table 1) are from Lu (1995),
Corbett and de Groote (2000) as well as Sucky (2004a,b).
The origin of research into the coordination of decentralized decision units dates
back to the paper of Clark and Scarf (1960). They study a serial multi-echelon inven-
tory system and devise a recursive decomposition approach for determining the optimal
parameters of an echelon stock policy. In the meantime both convergent and divergent
inventory systems have been studied with remarkable results. For a review of this
research area we refer to Thomas et al. (1996) and Minner (2000). A paper which
explicitly deals with decentralized inventory decisions and asymmetric information is
Fransoo et al. (2001) (see Table 1).
Decomposition and coordination is also the topic of Hierarchical Planning. First
introduced by Anthony (1965) and specialized for applications in the area of pro-
duction planning by Hax and Meal (1975) this area has gained great acceptance for
intra-organizational SCs. The planning tasks facing a company are split into several
levels. The closer a planning level is to the actual object of planning (e.g. the shop
floor) the greater the degree of detail and the shorter the planning horizon will be.
Coordination will be achieved by anticipation of lower level decisions, directives and
feedback. In any case there is a central planning unit at the top coordinating the overall
SC (for more details see Schneeweiss 2003). Considering elements from hierarchical
8 H. Stadtler
planning Schneeweiss and Zimmer (2004), Zimmer (2001) devised a CP scheme for
an inter-organizational SC with two SC members (see Table 1).
Papers mentioned above and included in Table 1 fulfill our definition of CP and
advocate model-based CP. However, Table 1 contains three further papers, namely by
Barbarosoĝlu (2000), Gjerdrum et al. (2002) as well as Shirodkar and Kempf (2006)
which ultimately apply a central model and impose the resultant decisions on the SC
(see Table 1). Since the cases described in these papers contain a number of interest-
ing ideas to overcome information asymmetry, we hesitated to totally exclude these
papers from our presentation. Hence, we show the underlying structure of the SC, the
relationships among SC members and the decision situation facing each SC member
in the Appendix (Table 5).
Finally, a few related research areas will be mentioned which are not included here
due to the following reasons:
A field that has attracted many researchers in recent years is SC coordination with
simple contracts. A contract “. . . makes the terms of the relationship explicit” (Tsay
et al. 1998, p. 306). Often idealized decision problems are considered to gain structural
insights, like the news vendor model. Tsay et al. (1998) distinguish eight different con-
tract clauses (like the specification of decision rights, pricing or buyback and return
policies).
Only a few papers deal with several of these clauses simultaneously (Tsay et al.
1998, p. 306). Although there is evidence of a successful application of a clause
in practice (Cachon and Lariviere (2001) mention a revenue-sharing contract in the
video-rental industry) most results are still theoretic and based on stylized models.
Hence, “they fail to address a variety of issues that become relevant to actual imple-
mentation” (Tsay et al. 1998, p. 330). Although models have been extended in the
meantime (see the paper by Wang and Gerchak 2003 and the overview by Cachon
2003) this conjecture still holds. Hence, contract theory will not be considered here
any further. This does not preclude that CP schemes analyzed here incorporate results
from this research area ( Lu 1995; Corbett and de Groote 2000; Sucky 2004a, b).
Also, we will not deal with pure auctions (for an introduction see Milgrom 2004).
In our view auctions are most applicable in pure market interactions at the boundaries
of a SC but not within a SC. Note that auction mechanisms may be utilized to coordi-
nate decentralized (often detailed) schedules (like Fox et al. 2000). This then can be
regarded as a substitute for a central optimization approach.
Likewise, we will not consider Collaborative Planning, Forecasting and Replen-
ishment (CPFR) here since its emphasis (so far) is on the exchange of information and
less on model-based decision support. A short outline of some of the developments
of CPFR should suffice: CPFR is a formalized process which has been worked out by
the standardization committee VICS (Voluntary Interindustry Commerce Standards)
and implemented over 300 companies (VICS 2004, p.5). The CPFR process model
consists of eight planning tasks, which can be subsumed under four main activities:
strategy and planning, demand and supply management, execution and analysis (VICS
2004, S.9). Here “planning” does not refer to the alignment of operational plans but
to the identification and communication of events which may affect demand, such as
promotional activities or product introductions.
While in the original model collaboration is restricted to mere information

exchange, some authors extend the scope of CPFR to joint decision-making of the SC
members involved. Danese (2005, p.458) mentions a concept called “limited CPFR
collaboration”, where plans are synchronized jointly by the SC members (e.g. replen-
ishment plans between a central company and a distribution centre) and exceptions
are managed.
Multi-agent systems may be one way to deal with CP. Here “an agent is an auton-
omous, goal oriented software process that operates asynchronously, communicating
and coordinating with other agents as needed” (Fox et al. 2000, p. 16). While agents
may be in charge of only a single element of a (production) system [like a production
agent representing a single machine which is able to produce a certain number of
products (Grolik et al. 2001)] there may also be very versatile (software) agents repre-
senting a whole decision unit (like Master Planning) of a specific SC member (Stadtler
2004). While we are not interested in the software architecture of these multi-agent
systems the basic logic and models guiding the search for a SC solution may well be
of interest here (e.g. see the proposal of Jung and Jeong 2005, Jung et al. 2005). Since
any CP scheme considered here could be implemented along the guidelines of agent
technology, we will not make use of it as a discriminating criterion.
A first taxonomy of coordination has been advocated by Whang (1995). He first dis-
criminates organizational units according to coordination “within operations”, “cross-
functional” and “inter-organizational”. The second criterion relates to the behaviour of
the people in the organization, namely “single-person perspective”, “team perspective”
and “nexus-of-contract perspective”. A single-person perspective incurs the advantage
that there is a single (central) decision-making unit who has access to all relevant infor-
mation. In the team perspective each party has limited information and action sets.
This requires to communicate and coordinate activities to achieve the global (team)
objective, i.e. in the team perspective the organizational units act separately but share
the same objective. This is in contrast to the “nexus-of-contract perspective” where
there are separate decision making units with private information and individual goals
(opportunism). For each of the resulting nine subcategories Whang presents examples
from literature. One of his conclusions is that in 1995 the research in Operations and
Information Management has heavily leaned towards the single-person perspective of
organizations.
In this paper we will further elaborate the two subcategories defined by “inter-
organizational coordination” and “team-perspective” as well as “nexus of-contract
perspective”. Since the latter involves multiple organizations and has no relevance to an
organizational perspective Whang (1995, p. 420) renamed it into “inter-organizational
interaction”.
Although the taxonomy of Whang (1995) is a first attempt to categorize different
types of coordination it does not span the large variety of characteristics necessary to
describe CP problems. Subsequently, we will present a framework (synonym “typol-
ogy”) of CP problems and schemes intended to describe their main characteristics as
observed in the literature in greater detail. Still, we do not claim to have extracted
all possible characteristics in its totality as would be required for a classification (see
Dyckhoff and Finke 1991, pp. 1 for a discussion of these terms).
10 H. Stadtler
Ideally, a CP scheme should contain a set of activities and rules applicable to a wide
range of decision problems. However, CP schemes presented in the literature often are
closely linked to a specific decision problem. Hence, our framework of CP consists
of three broad categories:
(1) the structure of the SC and the relationships among SC members (Sect. 3),
(2) the decision situation facing each SC member (Sect. 4) and
(3) the characteristics of the CP scheme itself (Sect. 5).
The first two categories specify a CP problem facing a SC. This allows us to cluster
those application areas where CP schemes are already available. Ideally, the third cat-
egory should not depend on the two former categories. Not surprisingly it turned out
that CP schemes actually are designed for specific CP problems.
3 SC structure and relationships among SC members
The structure of the SC constitutes a major factor for CP and the complexity of align-
ing plans. Furthermore, the relationships among SC members will have an influence
on the type and reliability of the information exchanged. Thirdly, requirements for a
SC solution may comprise an objective to be followed by the supply chain as a whole
including the notion of fairness. These issues will be described in the following (see
Table 2). Please note that the information presented in tables often can only provide a
first idea and not a complete description of a criterion specification:
3.1 Criteria (1.1): Structural elements of the SC
For defining the structure of a SC we consider

(1.1.1) the number of SC tiers
(1.1.2) the number of SC members on each tier and
(1.1.3) the business functions SC members fulfil.
The easiest situation, where collaboration can be applied is a two party situation, usu-
ally termed a supplier and a buyer (depicted by “1–1”). In general one has to mention
the number of tiers considered and the number of SC members n i (individual planning
domains) on each tier i (where i = 0 depicts the SC member most downstream the
SC). Additionally one can discriminate the structure of flows between SC members as
serial, divergent, convergent, mixed and cyclic similar to definitions well known from
the bill of materials. Note, that we do not consider the structure of flows within the
planning domain of a SC member. The difficulty in the context of CP is not only the
complexity resulting from the number of links to handle but also the additional deci-
sions to make, e.g. consider the case of a two-level divergent SC (1 − n 0 ) with a scarce
resource on the side of the supplier. Now we may have to decide on the allocation of
the scarce material in light of the (unfilled) demand from the buyers.
According to the business functions to coordinate Bhatnagar and Chandra (1993)
distinguish “multi-plant coordination” and “general coordination” within the context
of intra-organizational collaboration. If decisions have to be linked within the same
Table 2 Criteria for discriminating SC structures and relationships
Criteria for Lu (1995) Corbett and de Ertogral Fransoo Karabuk Fink Schneeweiss and Sucky Dudek and Stad- Jung
SC structure and Groote (2000) and Wu et al. and Wu (2003,2004) Zimmer (2004) (2004b) tler (2005, 2007) et al.
relationships (2000) (2001) (2002) (2005)
1. Structural elements
1.1.1 No. of tiers 2 2 3 2 2 2 2 2 2 2
1.1.2 No. of members on tiers 1–1, 1–n 0 1–1 n 2 –n 1 –n 0 1–n 0 1–1 n 1 –1 1–1 1–1 1–n 0 1–1
1.1.3 Business functions prod. prod. prod. prod. prod. prod. prod. prod. prod. prod.
and and mar- and
retailer keting distr.
1.2. Relationships
A framework for collaborative planning and state-of-the-art
1.2.1 Power S ? – ? ? – B ? ? ?
1.2.2 Behaviour team opp. team team team team team opp. team opp.
1.2.3 Learning eff. – +∞ – – – – – +∞ – yes
1.2.4 Roll. sched. – – – – – – – – yes –
1.3. Required opt. S fair SC SC SC SC S SC SC
SC solution for S, 1–1 near opt. near opt. (near) near opt. near opt. near opt. near opt. near opt.
opt.
B buyer, S supplier, distr. Distribution, – no or not applicable, inv. inventory, +∞ solution lasts forever, n i number of SC members on tier i, ? open question since no SC
member possesses all the mechanisms to exercise power, opt. optimal, opp. opportunism, prod. production, roll. sched. rolling schedules
11
12 H. Stadtler
business function at different echelons in an organization this is termed multi-plant

coordination. Although not limited, the term focuses on the production function. If
several functions, such as production and marketing have to be aligned then the term
general coordination applies.
This idea can be transferred to inter-organizational collaboration, also. Hence, we
distinguish two specifications of the business function within a SC: the planning
domains of all SC members are
– of the same type or
– of a different type.
Consequently, the name of the business function(s) involved have to be indicated.
For instance Barbarosoĝlu (2000) and Fink (2003, 2004) consider the same business
function—production—for all SC members. Fransoo et al. (2001) deal with coordi-
nating a manufacturer and several retailers while Jung and Jeong (2005) show how to
coordinate production with the distribution of final products.
3.2 Criteria (1.2): Relationships among SC members
The relationships among SC members comprise the following issues:

(1.2.1) power of each SC member,
(1.2.2) extent of self-interest governing a SC member’s behaviour,
(1.2.3) learning effects and
(1.2.4) rolling schedules.
The power of a SC member and its relative position in a SC may result from different
sources, like
– product and (production) process know-how,
– number of competitors,
– portion of the value creation with respect to the value of the final product,
– access to the customer base (market) and
– financial resources.
Listing and describing these attributes is rather simple, however, measuring power is
much more difficult.
In the literature on CP the notion of power (or leadership as it is called sometimes)
is usually not discussed explicitly. However, the premises and CP scheme advocated
may reveal in which way power is used by SC members. Still, generalizations often
are difficult: For instance a SC member making the first offer may on the one hand
fix his minimum profit while on the other hand reveal new information which may
be exploited by another SC member when making a counter proposal. (If informa-
tion about the most powerful SC member is provided by the authors explicitly this is
indicated in Table 2, otherwise a “?” is placed.)
Furthermore, the distribution of power may and usually will change over time—it
may even change in the course of a single instance of collaboration. Also, power may
be exercised in different phases of the collaboration (see criterion (2.2)), i.e.
• when designing a SC and fixing the conditions of a collaboration strategy (note,

this may also concern the distribution of the efficiency gain),
• in the planning phase the most powerful member may choose to make the first
offer and thereby fixing his minimum profit (like in Dudek and Stadtler 2005) or
enforce a certain stopping rule,
• in the execution phase the way exceptions are handled and resultant costs are split
may be to the disadvantage of the weaker SC members and finally
• in the evaluation phase the most powerful SC member may impose penalties for
those SC members not conforming to expected performance targets.
The extent of self-interest governing a SC member’s behaviour has already been men-
tioned in Sect. 2 when referring to the taxonomy of Whang (1995). Here, a team
behaviour exists if all actions (or decisions) being favourable for the SC as a whole are
accepted, irrespective of the SC members’ individual interests. Opportunism prevails,
if SC members only implement actions which are in their own interest and may even be
cheating when passing information to others. Note that providing information (either
truthfully or not) can be regarded an action, too (see Hirshleifer and Riley 1992).
According to a survey by Landeros and Monczka (1989) a supplier–buyer part-
nership rests (among other things) on a trustworthy commitment of future conduct—
hence, a team perspective may not be unrealistic.
Looking at the papers considered here, it is not easy to distinguish between team
perspective and opportunism. Dudek and Stadtler (2005, 2007) assume that all parties
provide requested information truthfully while searching for an overall supply chain
optimum (hence, a team perspective applies). The paper of Fransoo et al. separates
the retailers in cooperative (team) and non-cooperative (opportunistic) retailers. The
latter do not share (demand) information and impose a β-service level constraint the
manufacturer has to fulfil (thereby exercising their power over the manufacturer).
Considering this β-service level as a constraint, the remaining SC members exercise
team behaviour. According to Table 2 CP schemes for opportunism are the minority
(Corbett and de Groote 2000; Sucky 2004b; Jung et al. 2005).
Rolling schedules play a role in the planning phase. It is most popular in industry in
order to cope with uncertainty (e.g. of demand). In a SC setting this not only involves
updating and extending an existing plan by one SC member but also renegotiating all
changes with all other SC affected members. One question here is, who will bear the
costs resulting from these changes if there is already a previously approved plan? The
majority of papers assume that a plan once agreed upon will be executed unaltered up
to the planning horizon (an exception is Dudek 2004, pp. 115).
Closely related to rolling schedules is the notion of learning effects. If the nego-
tiation procedure is repeated, then a party may make use of information gained in
previous negotiations. This is especially true if there is an overlap of decisions in
two successive plans (like in rolling schedules). In the extreme this may lead to a
totally different situation, like in Corbett and de Groote (2000), p. 447: once the
buyer has decided to choose a specific purchasing contract all data are revealed to the
supplier. Thus there is no renegotiation and the conditions of the contracts must last
“forever”.
14 H. Stadtler
3.3 Criteria (1.3): Requirements for a SC solution
The objective(s) governing the generation of plans in the CP scheme should not be
mixed with the members individual objectives. Here, we are interested in the way
the—often conflicting—objectives are handled. We discriminate three broad types of
objectives:
– the alignment of flows,
– the search for the SC optimum and
– the search for a fair solution.
If “only” the alignment of (material) flows is looked for, a feasible supply and pro-
curement plan has to be established such that customer orders are fulfilled in due
time. No monetary objectives are considered here (e.g. Friedrich et al. 2002; Stäblein
and Baumgärtel 2006). However, aiming at the SC optimum often means maximizing
profits or minimizing costs for the SC as a whole. Such a solution may incur a loss
for some members and high profits for some others. In these cases side-payments or
discounts may become an issue to yield a win–win situation for each member.
Searching for a fair solution requires a definition of the term fair. Here, a number
of proposals exist which will be discussed in detail in Sect. 6.
Having described the SC structure and the relationships between SC members we
can continue in characterizing the decision situation each member is facing.
4 Characteristics of the decision situation of each SC member
Characteristics for discriminating decision situations fall into four broad categories
which stem from answering the following question: which decisions take place, when,
with which objectives and which information? These four Ws of CP will be described
in greater detail in the following (also see Table 3).
4.1 Criteria (2.1): Decision models—which decisions take place?
Here the real world decision situation (planning tasks) of each SC member should be
described. However, if standard decision models are deployed, their names provide a
clue for the underlying decision situation.
Often it is assumed that each member in the SC faces the same type of deci-
sion model, e.g. both the supplier and the buyer deploy an EOQ model for a single
product. The buyer calculates the order quantity while the supplier calculates opti-
mal production orders. But there are also examples where SC members face different
(basic) decision situations: a resource-constrained-project scheduling problem
(RCPSP) for the buyer and a capacitated lot-sizing problem (CLSP) for the supplier
(e.g. Schneeweiss and Zimmer 2004).
We would like to add that in the course of collaboration both the goal(s) as well as
the action set of a SC member may change. This may require additional constraints
which often destroy the “typical” structure of a standard decision model. As a result
a solution technique applicable for the standard decision model may no longer be
Table 3 Characteristics for discriminating the decision situation
Criteria for Lu Corbett and Ertogral Fransoo Fink (2003, Karabuk Schneeweiss and Sucky Dudek and Jung et al.
decision situation (1995) de Groote and Wu et al. 2004) and Wu Zimmer (2004) (2004b) Stadtler (2005)
(2000) (2000) (2001) (2002) (2005, 2007)
2.1. Decision models EOQ EOQ ML (R,S)- MIP 2-stage CLSP, RCPSP EOQ ML CLSP LP
CLSP inven- sto-
tory chastic
program
2.2 Phases of collaboration cond. cond. oper. oper. oper. oper. oper. cond. oper. oper.
2.3. Information status
2.3.1 Information hidden (from B or S) s&h h cost B all costs all all cap. S s&h cost all all
cost B B
2.3.2 Information exchanged orders order menu orders service rej./ac. plans prices& orders order orders orders
level supply menu

targets quanti-
ties
2.3.3 Degree of uncertainty det. det. det. stoch- det. stoch- det. stoch- det. det.
astic astic astic
(market (yield,
demand) demand)
2.4. Objective(s) min. min. costs min. min. arbitrary max. min. costs min. min. costs min.
costs costs costs profit costs costs
B buyer, oper. operational planning phase, cap. capacities, rej./ac. plans generated by a mediator are transferred to SC members and either rejected or accepted, cond.
collaboration conditions, det. deterministic, S supplier, inv. Inventory, s& h setup and holding, – no
15
16 H. Stadtler
skeleton negotiations adaptations renegotiate or

contract about opera- during renew contract
tional details execution
time
collaboration operational execution phase evaluation phase
conditions planning phase
Fig. 1 Phases of collaboration
applicable. As an example consider the EOQ model where the non-linear objective
function can be minimized by taking the first derivative. In a situation where a supplier
would like to generate a menu of supply proposals to be presented to the buyer addi-
tional constraints result. Now the constrained non-linear optimization model requires
the application of the Karush–Kuhn–Tucker conditions (Sucky 2004b).
The complexity of describing CP schemes becomes clear when dealing with the
decision problem at hand. For each decision problem there may exist a distinct clas-
sification or typology [e.g. for the RCPSP see Brucker et al. (1999) and for lot-sizing
see Drexl and Kimms (1997)]. Further typologies for decision problems in the area of
production have been put forward which so far have not been addressed in conjunction
with CP schemes [e.g. cutting and packing (Dyckhoff and Finke 1991) and assembly
line balancing (Boysen et al. 2007)].
4.2 Criteria (2.2): Phases of collaboration—when does collaboration take place?
Decisions are made at different points in time in the course of collaboration (see Fig. 1).
We discriminate four phases, starting with specifying the
(2.2.1) conditions of collaboration followed by
(2.2.2) planning,
(2.2.3) execution and
(2.2.4) evaluation phase.
Determining the conditions of collaboration often incurs negotiating the terms of a
skeleton or detailed contract valid over a certain period of time. The wholesale price
(Barbarosoĝlu 2000) or a service level (Fransoo et al. 2001) as well as the extent of
information exchange can be set. Furthermore, the auditing and evaluation procedure
may be agreed upon by the parties. In the (operational) planning phase the procure-
ment, production and distribution plans at different levels of the planning hierarchy
have to be aligned.
During the execution phase (of the contract) it might be advantageous to reconsider
certain obligations (like the fulfillment of an order) due to unexpected incoming orders.
Finally, once the duration of a contract comes to an end, an ex post evaluation may
take place. This may give rise to renegotiating the terms of the contract and possibly
its prolongation. An example for the evaluation phase is presented in the paper by
Jammernegg and Kischka (2005) where the contract’s attributes may be renegotiated
in the course of time in order to improve the performance of a SC.
CP schemes proposed in the literature often address decisions relating to just one
planning phase. CP schemes presented in Table 3 are restricted to the first two phases.
Obviously, it would be ideal to take a holistic view and to devise a concept covering
all four phases of collaboration.
4.3 Criteria (2.3): Information status—what is the information status of each SC

member?
The information status in an inter-organizational collaboration is assumed to be asym-

metric. The reasons for asymmetric information can be manifold. On the one hand,
there are practical reasons such as administering a decentralized database may be more
economical and faster than a central database. Also, information gathered decentrally
(like the productivity of specific worker/job assignments) may remain the expertise
of local decision makers. On the other hand, some information might be hidden from
the other SC member(s) in order not to weaken the (future) bargaining power (e.g. a
supplier disclosing large slack capacities may run the risk that the buyer will ask for
price reductions). In an asymmetric information situation the question remains
(2.3.1) which information is hidden to the other party?
Another aspect regards the
(2.3.2) type of information to be exchanged in the course of collaboration.
Here, one can discriminate three subcategories
– quantities (like purchase orders or supplies of a product),
– monetary values,
– additional key performance indicators (KPIs).
As monetary data we may have data from cost accounting as well as data directly
applicable for decision making (like a product’s holding cost coefficient or penalty
costs for late delivery). These data are usually regarded sensitive data, i.e. data that
can do harm to the owner of the data if exploited by a third party. Another type of cost
is the total cost of a “plan” (which may also be regarded a KPI) or a side-payment or
compensation required for accepting a SC member’s proposal.
A special case has been reported by Shirodkar and Kempf (2006). Here, suppliers
have collaborated with the buyer in specifying a model of their planning domain. The
model has been transferred to the buyer in order to combine all these submodels into a
central model. For the operational planning phase the data have to be maintained and
updated by each supplier. In this way information asymmetry is lifted.
The third sub-category, the additional KPIs, is often agreed upon at the start of a
SC relationship. KPIs, like service levels, are calculated continuously or in certain
intervals of time and serve to measure whether the SC is operating as expected.
A final discrimination of the data is the
(2.3.3) degree of uncertainty
A decision model used by a SC member may contain uncertain data due to the
environment (like market demand) or due to the behaviour of SC members. The latter
18 H. Stadtler
has been addressed in the SC literature very often and various ways to overcome this
source of uncertainty have been proposed (e.g. the uncertainty of demand of a sup-
plier can be reduced by transferring the buyer’s production plan in a vendor-managed
inventory (VMI) situation (see Holweg et al. 2005)). As a result either
– deterministic or
– stochastic models
will be constructed.
4.4 Criteria (2.4): Objectives—what are the objectives of the decision problem?
The last category concerns the objective(s) of the decision problem a SC member
strives for. Basically, either profit maximization or cost minimization is considered.
The majority of papers analysed here advocate minimizing costs. If a standard deci-
sion problem is considered then the objective is often also standard (e.g. minimizing
of the sum of setup and stock holding costs per unit time for the EOQ). Models used
in industrial practice often make use of soft constraints. Violating these constraints is
penalized in the objective function (Shirodkar and Kempf 2006).
A time-oriented objective function is used in Fink (2003, 2004) where the objective
of one member is to minimize throughput times.
5 Characteristics of collaborative planning schemes
CP schemes will be described here by only a few structural elements that can be
“observed” and agreed upon by SC members and thus can also be regarded as char-
acteristics of the CP problem (Table 4). Also, we will not go into algorithmic details
since there may be several internal options to generate solutions (like for an LP model).
The interaction between the parties involved in a CP scheme can be documented by a
protocol (see Fink 2004 or Stadtler 2004 for examples). The structural elements to be
presented here define the parties involved, the starting point as well as their “interface”.
Hence, the following four characteristics will be analysed:
(3.1) the incorporation of a mediator,
(3.2) the initial solution,
(3.3) the number of rounds and the number of offers to be exchanged (stopping
criteria) and
(3.4) final results SC members can expect.
A mediator is a third party controlling the rules of the game, e.g. by controlling
the (timing of) interactions among members. A mediator may have the capability of
generating plans and presenting these to all SC members for evaluation and even may
be entitled to propose the distribution of efficiency gains among SC members. An
important issue is the proliferation of data to a mediator required for generating plans
for the SC as a whole, i.e. a mediator must be a trusted entity. A mediator differs from
a central planning function (executed by one SC member) in that a mediator should
not be biased and has to take into account preference or objectives of each SC member
Table 4 Characteristics for discriminating CP schemes
Criteria for Lu Corbett and Ertogral Fransoo Karabuk Fink (2003, Schneeweiss Sucky Dudek and Jung
CP schemes (1995) de Groote and Wu et al. and Wu 2004) and Zimmer (2004b) Stadtler et al.
(2000) (2000) (2001) (2002) (2004) (2005, 2007) (2005)
3.1. Mediator – – yes yes – yes – – – –

3.2 Initial solution – – up – – random – – up down
3.3. No. of plans exchanged
3.3.1 No. of rounds 1 1 n big – n big n big 1 1 n small n small
3.3.2 Parallel offers 1 1 – 1 1 1 1 1
n? n?
→ total no. offers 1 n? n big – n big n big 1 n? n small n small
3.4. Final results
3.4.1 Quality of sol. anal. – comp. – proof comp. comp. – comp. comp.
3.4.2 Side–payments – yes – – yes – yes yes yes –
anal. analytical proof, comp. computational tests, down downstream planning, n big big number (e.g >> 10), sn small small number (e.g. ≤ 10), n ? number unknown, proof
proof of convergence to the local optimum of each SC member, up. upstream planning, – no, ? not mentioned in the paper
19
20 H. Stadtler
adequately. In industrial practice such a mediator in the area of planning has become
known as an application service providing company (Knolmayer et al. 2002, p. 10). So
far, a mediator has rarely been considered in the literature on CP [with the exception
of Ertogral and Wu (2000), Fransoo et al. (2001) and Fink (2003, 2004)].
When starting a negotiation process the solution to start from may be very impor-
tant. In rolling schedules one can expect that the plan agreed to previously (e.g. the
week before) will form the basis for updating, renegotiation and extending a plan in
light of newly available information (e.g. customer orders).
If no such plan exists a plan has to be generated from scratch. A general and easy
way is to generate an initial plan for the SC members randomly (Fink 2003, 2004).
The drawback is that these initial plans probably are infeasible and the CP scheme
must take care to finally obtain feasibility.
A procedure often used in practice is upstream planning (Simpson and Erenguc
2001). Here, the SC member closest to the final customer starts to generate his (opti-
mal) plan with demand forecasts as an input. The supply of input materials is assumed
to be unconstrained. From this (individually optimal) plan the required material sup-
plies are derived and transferred to the suppliers on the next upstream tier. The mate-
rial supplies requested now become the demand of the first tier suppliers. These first
tier suppliers then apply the same logic for generating their (optimal) plans. In this
way supply plans are propagated upstream the whole SC. This procedure assumes that
suppliers will always be able to supply the materials required (Dudek and Stadtler
2005).
In case the sequence of generating plans is reversed, i.e. starting with the most
upstream supplier, downstream planning results. The problem here is to indicate the
demand a supplier has to fulfill: given the supplier gets access to the demand fore-
casts of final products a fixed lead time offset can be applied to estimate when the
corresponding parts demands will take place. This logic is applied by Jung and Jeong
(2005) to a producer and a distributor. We would like to add that downstream planning
is only favourable if the most expensive bottlenecks exist upstream and fixed lead
times can be estimated with sufficient accuracy.
The expected number of rounds to run and the number of offers to be exchanged
in the course of negotiations largely differ among CP schemes analysed. A round is
completed if one party generates an offer or several parallel offers to choose from and
the subsequent evaluation by the other SC member(s). We say there is “no” round if
there is only one instruction (and no “offer”) by one SC member the other members
have to follow, which is the case in classical central planning. Limited coordination
exists (one round) if there is only a single offer or a menu of offers where the other
party has the right to choose from or to “take-it-or-leave-it” (see Corbett and de Groote
2000; Sucky 2004b). The “burden” of collaboration largely depends on the number
of offers to generate and evaluate. Hence, an indicator is the expected total number of
offers resulting from the expected number of rounds times the number of parallel offers
(see Table 4). The number of offers is regarded “small”, if each offer (or plan) can be
evaluated by a human decision maker before it is presented to the other members in
the SC. If this is the case then an interactive CP scheme can be designed, otherwise the
CP scheme must be automated (like in the case of Fink 2003, 2004). An interactive
CP scheme with a small number of offers (e.g. at most ten) increases the chances of
acceptance by the decision maker(s).
The final results of a CP scheme are described by the
(3.4.1) quality of solutions and
(3.4.2) the use of side-payments.
Ideally a CP scheme will result in the SC optimum and a win–win situation for each
SC member while the gains of collaboration are distributed among the members in
such a way that no SC member has a desire to deviate from the SC solution generated
(e.g. a Nash equilibrium). A Nash equilibrium is often looked for in game-theoretic
approaches (like in Corbett and de Groote 2000; Sucky 2004b). If the decision situa-
tion becomes complex an optimality proof rarely exists [Lu (1995) proved optimality
of the solution for the one supplier – one buyer case analytically]. Karabuk and Wu
(2002) present a proof of convergence to the individual optimum of each SC member
based on Augmented Lagrangean theory. However, in the majority of cases computa-
tional tests come into play showing the performance within a certain decision context
(e.g. multi-level capacitated lot-sizing). Given the tests are based on a suitable set of
parameter combinations, results of these tests can provide a good indication for the
suitability of a CP scheme in this context. However, the transfer of results to other
types of decision problems may be misleading.
In case neither optimality nor an equilibrium can be proved certain properties of the
final solution might be stated, like performance bounds. However, no such structural
results have been reported in the literature analysed here.
Obviously the distribution of gains resulting from collaboration plays a major role
for the active involvement and the acceptance of the results of a CP scheme. Side-
payments (sometimes called compensations) may be one instrument to achieve a win–
win situation for each SC member. This leads us to the notion of fairness which will
be discussed in the next section.
6 The notion of fairness
In experimental economics the ultimatum game has gained much attention (Güth
1995). It provides insights into human behaviour and the evaluation of a fair solu-
tion. The ultimatum game assumes that there are two players A and B which play the
game only once. Player A receives an amount of money X which he has to share with
player B. The share s offered to B remains the decision of A, it may range from 0%
< s ≤ 100%. In case player B accepts the share the amount X is shared accordingly
among the two players. However, if player B rejects, the amount X is withdrawn and
neither A nor B get any money.
In the ultimatum game information asymmetry exists, since player A does not know
the threshold value of player B separating the acceptance of share s and its refusal.
From the point of rationality player B should accept any positive share, since the alter-
native is to get nothing. However, experiments have shown that if the share offered
to B is “not substantial” player B prefers “no payment” because B does not grant the
larger share to A – or in our terminology here regards the distribution of payments
unfair. Assuming that A knows the attitude of B, he should offer a fair share in order
22 H. Stadtler
not to lose all the money. In Western culture this results in player A offering a share
of 66.7% on average (Güth 1995, p. 331). Interestingly, experiments have shown that
there are some cultures which regard a significantly smaller share for B to be fair
(because A is regarded the leader of the game).
Our conjecture from these experiments for CP is that fairness is depending on the
culture, role and decision situation a SC member is in.
A proposal considering fair shares explicitly in dyadic channel coordination by a
wholesale price contract (news vendor model) is presented in Cui et al. (2007). Here,
the retailer requires an a priori given fraction (share) of the manufacturer’s payoff in
order to reduce his disutility from inequity. The authors provide formulas—depending
on the fraction required by the retailer—for setting a wholesale price that coordinates
the SC.
Ertogral and Wu (2000) define several objective functions for different measures
of fairness. They analyse a vertical SC where each SC member aims at minimizing
his costs. Costs comprise setup and holding costs up to a given planning horizon. As
a benchmark the minimum cost solution is calculated for each member provided the
primary and secondary customer demands are fulfilled in time (assuming that nei-
ther a lead-time offset nor lot-sizing takes place at downstream production facilities).
Now, any solution generated for the SC as a whole can be evaluated for each member
with respect to the absolute deviation of costs (an increase or decrease) compared to
the benchmark. This also allows us to calculate the average cost deviation for all the
members in the SC. Based on these measures, unfairness results if one member faces
an absolute cost deviation larger than the average cost deviation.
The authors consider three objectives:
– minimizing (the sum of) unfairness across members,

– minimizing the relative cost increase above the member’s-best solution (bench-
mark) and
– minimizing the distance between the member with the least unfair and the most
unfair solution.
Ertogral and Wu (2000) favour the first objective. However, even for the first objec-
tive their computational tests showed that fair solutions sacrifice 37.15% on average
in solution quality [compared with the minimum cost solution of the SC as a whole
(central optimization)].
Our conjecture from this study is that a CP scheme which aims at achieving “fair-
ness” solely by considering the members’ individual cost functions may be counter-
productive with respect to SC competitiveness. It seems much more favourable to
look for a CP scheme which aims at finding the SC optimal solution first and then to
allow side-payments (or discounts) such that a fair solution is reached (which should
be a win–win situation). In order to calculate fair side-payments the above-mentioned
objectives (criteria) may be applied a posteriori.
Gjerdum et al. (2002) propose to introduce fairness into their central model for coor-
dinating two inter-organizational SC members e = 1, 2. At the start each SC member
indicates the minimum profit πemin to the central planner that must be reached. Then
a non-linear objective function is constructed
improved
(πe − πemin ) (1)
e
improved
which allocates profits πe in such a way that, ideally, each SC members will
get the same absolute additional amount.
Fleischmann (1999) analyses the collaboration of two logistic service providers
e = 1, 2 which may collaborate in the distribution of goods to the same customers
by, e.g. sending just one truck (horizontal collaboration). Fleischmann favours the
following rule to allocate the resultant cost savings ex post: a solution is neutral to
competition if the savings S are split such that costs κe are cut by the same percentage α
S
α= (2)
κ 1 + κ2
Note that this allocation rule requires calculating (initial) costs κe of each member
assuming no collaboration.
Transferring the proposal of Gjerdum et al. (2002) to a cost minimization objective
and contrasting it with the one of Fleischmann (1999) reveals that granting the same
absolute cost savings to each SC member will not be neutral to competition provided
initial local costs are different.
While formal rules or axioms for the allocation of gains of a collaboration may be
shown to be fair one should keep in mind that the perception of fairness by the parties
involved will often incur some subjective elements and will be situation dependent
(e.g. on the distribution of power among parties).
7 Summary and outlook
We have presented a framework for collaborative planning (CP) which allows to con-
trast and cluster various contributions in this relatively new research area. A selected
number of papers dealing with operational, model-based CP has been analysed accord-
ingly.
These ten papers primarily address the following characteristics of collaborative
planning (see the mind-map type Fig. 2. Here, the underlined characteristics indicate
that at least five of the ten CP approaches analysed possess this characteristic): a two
tier SC with one SC member on each tier is addressed. Production decisions have
to be aligned with the aim of obtaining a near optimal SC solution. Mainly lot siz-
ing decisions are addressed—both static and dynamic—for the operational planning
phase. Orders—including order menus and sequences of orders—are exchanged in a
deterministic environment. The CP schemes do without a mediator and do not need a
(specific) initial solution. The number of rounds ranges from 1 to a very large number
with only one plan exchanged per party per round. Computational tests provide some
insights into the quality of solutions.
Unfortunately, a widely applicable or even generic CP scheme for more complex
SCs is still missing (with Fink 2004 coming very close to this challenge). Also, a holis-
tic view or concept covering several phases of collaboration has still to be elaborated.
24 H. Stadtler
(1-1)
( n1 - 1 ) 1
1.1.1/2 No. tiers &
l ot s iz e no. SC m on ea ch ti er (1-n0)
LP 1 2.1 Decision models 1.1 Structural (n2-n1-n0) 1
elements
MIP 1
1.1.3 Business functions p r o d uc t i o n
stoch. inv.1
1. marketing 1
SC structure distribu tion 1
conditions & relationships
1.2.1 Power of each SC m
o p er ati o nal 2.2 Phases of tea m
execution 0 collaboration 2. 1.2.2 SC m behaviour
1.2 Relationships opportun.
Decision
evalu at i on 0 situation yes 1
1.2..3 Learning effects
al l 1.3 Required align flows 0 no
The W´s 1 . 2 . 4 R o l l i n g s c h e d u le s ye s 1
s&h SC solution SC (near) optimum
2.3.1 Hidden no
costs 1 information fair 1
cap 1
3.
orders 2.3.2 Inf. exchanged 2.3 Information CP schemes 3.1 Me di a to r yes
upstrea m
prices 1 status no
3.2 Initial downstream 1
service levels 1
solution random 1
deterministic costs + 2.4 Object ives
2.3.3 Uncertainty 3.3 No. of plans none
profit 1 1
stoch astic exchanged
3.4 Final 3.3.1 No. of nsmall
other 1 proof
re s ul ts rounds
Expl ana tions: computat ional tests nb i g
SC m Supply Chain member 3.4.1 Quality
3.3.2 No. of 1
no guarantee
x for at l east five CP approaches this result a pplies parall e l offers nsmall
yes
1
f o r o n e C P a p p r o a c h e s t h i s r e s u lt a p p l i e s 3.4.2 Side-
payments no nb i g
0
there is no CP approach with such a feature
Fig. 2 Summary of characteristics considered in CP approaches analysed
The performance of a CP scheme can be evaluated in two different ways. One way
is a comparison with an optimal central solution. Some might argue that decentral
planning in practice may even result in better solutions than a central one because a
central unit will not dispose of the level of detail and timeliness of information of a
decentral decision unit. Also, the experience how to handle soft data (like preferences
for job assignments) may be larger in decentral decision units. We will not contradict
these arguments. However, in a laboratory experiment or in theory we can assume that
there is an ideal central decision unit which possesses all the decentralized knowledge.
If there is evidence that a CP scheme performs well in these laboratory experiments
then we expect favourable results also in real-world applications. In the absence of a
central solution a CP scheme will be evaluated according to the improvements over
the initial solution versus the additional efforts incurred.
Research contributions in CP are limited to assumptions typical in the operations
management literature, like people are predictable in their actions, emotionless and
observable (Boudreau et al. 2003). In other words, organizational behaviour, communi-
cation or cultural aspects are mostly omitted in designing CP schemes. Incorporating
results from game theory (like in Corbett and de Groote 2000) is one step in this
direction but should be extended in future [the reader is referred to Akkermans et al.
(2004), Bendeloy et al. (2006) and Marble and Lu (2007) for valuable insights and
recommendations in this area].
Once adopted by industry CP schemes could be the object of empirical research
[like empirical research on Advanced Planning Systems, see Buxmann et al. (2004)
and Roussel et al. (2002)]. But so far CP is more a research area than a ready-to-use
product. Future research in CP should come up with CP schemes that
• work well for more than two tiers,
• work well in a great number of decision situations (problems) even in a mix across
a SC,
• address, support or even secure a fair distribution of the gains among SC members
and
• allow renegotiations of already accepted plans in rolling schedules.
Still, there is a further important obstacle to overcome: while companies have realized
that a SC perspective is necessary for improving competitiveness and that isolated
planning domains lead to local optima they are reluctant to share information and to
consider compensations to SC members required to reach a win–win situation as a
result of CP.
Appendix
Table 5
Table 5 Criteria for discriminating SC structures, relationships and decision situations
Criteria for SC structures and Barbarosoĝlu (2000) Gjerdrum et al. (2002) Shirodkar and Kempf (2006)
relationships and
decision situation of each
SC member
1.1 Structural elements

1.1.1 No. of tiers 2 2 2
1.1.2 No. of members on tiers 1 − n0 1–1 n1 − 1
1.1.3 Business functions production production production
1.2. Relationships
1.2.1 Power S ? B
1.2.2 Behaviour S Opportunistic team B Opportunistic
1.2.3 Learning effects, – – –
1.2.4 Rolling schedules – – –
1.3. Required solution S opt. (fair) SC fair B opt.
2.1. Decision models CLSP/ LP MIP MIP
2.2 Phases of collaboration conditions oper. oper.
2.3. Information status
2.3.1 Information hidden all, except: none none
2.3.2 Information exchanged orders expected from B prices model of S
2.3.3 Degree of uncertainty stochastic deterministic deterministic
2.4. Objective(s) min. costs min. costs min. costs+ viol.
B buyer, ? neither B nor S dominate, MIP individual mixed integer programming model, – no, n i number
of SC members on tier i, oper. operational planning phase, opt. optimal solution, S supplier, viol. violation
of soft constraints
26 H. Stadtler
Acknowledgements The author is indebted to Martin Albrecht and Carolin Püttmann for their valuable
contributions to this paper.
References
Akkermans H, Bogerd P, van Doremalen J (2004) Travail, transparency and trust: a case study of computer-
supported collaborative supply chain planning in high-tech electronics. Eur J Oper Res 153:445–456
Anthony RN (1965) Planning and control systems, a framework for analysis. Boston
Barbarosoĝlu G (2000) An integrated supplier–buyer model for improving supply chain coordination. Prod-
uct Planning Control 11:732–741
Bhatnagar R, Chandra P (1993) Models for multi-plant coordination. Eur J Oper Res 67:141–160
Bendeloy E, Donohue K, Schultz KL (2006) Behavior in operations management: Assessing recent findings
and revisiting old assumptions. J Oper Manage 26:737–752
Benders JF (1962) Partitioning procedures for solving mixed-variables programming problems.
Numerische Mathematik 4:238–252
Boudreau J, Hopp W, McClain JO, Thomas LJ (2003) On the interface between operations and human
resource management. Manufact Service Oper Manage 5(3):179–202
Boysen N, Fliedner M, Scholl A (2007) A classification for assembly line balancing problems. Eur J Oper
Res 183:674–693
Brucker P, Drexl A., Möhring R, Neuman K, Pesch E (1999) Resource-constrained project scheduling:
notation, classification, models, and methods. Eur J Oper Res 112:3–41
Buxmann P, von Ahsen A, Diaz LM, Wolf K (2004) Usage and evaluation of supply chain management
software—results of an empirical study in the European automotive industry. Informat Syst J 14:295–
309
Cachon GP (2003) Supply chain coordination with contracts. In: Graves SC, de Kok AG (eds) Supply chain
management: design, coordination and contracts. Handbook in operations research and management
science, vol 11. Elsevier, Amsterdam pp 229–239
Cachon GP, Lariviere MA (2001) Turning the supply chain into a revenue chain. Harvard Business Rev
79:20–21
Cachon GP, Lariviere MA (2005) Supply chain coordination with revenue-sharing contracts: strength and
limitations. Manage Sci 51:30–44
Christopher M (1998) Logistics and supply chain management, strategies for reducing cost and improving
service, 2nd edn. Financial Times Pitman Publishing, London
Clark AJ, Scarf H (1960) Optimal policies for a multi-echelon inventory problem. Manage Sci
6:475–490
Corbett CJ, de Groote X (2000) A supplier’s optimal quantity discount policy under asymmetric informa-
tion. Manage Sci 46:444–450
Cui TH, Jagmohan SR, Zhang ZJ (2007) Fairness and channel coordination. Manage Sci 53(8):1303–1314
Danese P (2005) A taxomomy of supply chain collaborations: an investigation of CPFR initiatives, paper
presented at the EurOMA conference on operational and global competitiveness, Budapest, Hungary,
19–22, June 2005
Dantzig GB, Wolfe P (1960) Decomposition principle for linear programs. Oper Res 8:101–111
Drexl A, Kimms A (1997) Lot sizing and scheduling—survey and extensions. Eur J Oper Res
99:221–235
Dudek G (2004) Collaborative planning in supply chains. A negotiation-based approach, Lecture notes in
economics and mathematical systems 533. Springer, Berlin
Dudek G, Stadtler H (2005) Negotiations-based collaborative planning between supply chain partners. Eur
J Oper Res 163:668–687
Dudek G, Stadtler H (2007) Negotiation-based collaborative planning in divergent two-tier supply.
Int J Product Res 45:465–484
Dyckhoff H, Finke U (1992) Cutting and packing in production and distribution. Physica, Heidelberg
Dyer JH, Cho DS, Chu W (1998) Strategic supplier segmentation: The next best practice in supply chain
management. Calif Manage Rev 40(2):57–77
Ertogral K, Wu SD (2000) Auction-theoretic coordination of production planning in the supply chain. IIE
Trans 32:931–940
Feldmann M, Müller S (2003) An incentive scheme for true information providing in supply chains.
OMEGA 31(2):63–73
Fleischmann B (1999) Kooperation von Herstellern in der Konsumgüterdistribution. In: Engelhard J, Sinz
EJ (eds) Kooperation im Wettbewerb. Wiesbaden, 168–186
Fink A (2003) Automatisierte Koordinationsmechanismen zur Ausgestaltung komplexer Verträge im Rah-
men mediatisierter Verhandlungsprozesse. In: W Uhr, W Esswein, E Schoop (Hrsg.), Wirtschaftsin-
formatik 2003, Band I, Medien, Märkte, Mobilität, Springer, pp 281–301
Fink A (2004) Supply chain coordination by means of automated negotiations. Proceedings of the 37th
Hawaii International Conference on System Sciences, IEEE
Fransoo JC, Wouters MJF, de Kok T (2001) Multi-echelon multi-company inventory planning with limited
information exchange. J Oper Res Soc 52:830–838
Friedrich JM, Mertens P, Eversheim W, Kampker R. (2002) Der CW-SCM-Ansatz, Eine komponentenbas-
ierte Supply-Chain-Management-Software für kleine und mittlere Unternehmen. Wirtschaftsinforma-
tik 44:117–130
Fox MS, Barbuceanu M, Teigen R (2000) Agent oriented supply chain management. Int J Flexible Manufact
Syst 12:165–188
Gerchak Y, Wang Y (2004) Revenue-sharing vs. wholesale-price contracts in assembly systems with ran-
dom demand. Product Oper Manage 13:23–33
Gjerdrum J, Shah N, Papageorgiou LG (2002) Fair transfer price and inventory holding policies in two-
enterprise supply chains. Eur J Oper Res 143:582–599
Grolik S, Stockheim T, Wendt O, Albayrak S, Fricke S (2001) Dispositive Supply-Web-Koordination durch
Multiagentensysteme. Wirtschaftsinformatik 43:143–155
Güth W (1995) On ultimatum bargaining experiments: a personal view. J Econ Behav Org 27:330–344
Hax AC, Meal HC (1975) Hierarchical integration of production planning and scheduling. In: Geisler
MA (ed) TIMS-studies in management science, vol. 1, logistics. North Holland/TIMS, Studies in
Management Sciences, Amsterdam, pp 127–172,
Hirshleifer J, Riley JG (1992) The analytics of uncertainty and information. Cambridge University Press,
Cambridge
Holweg M, Disney S, Holmström J, Smaros J (2005) Supply chain collaboration: making sense of the
strategic continuum. Eur Manage J 23:170–181
Jammernegg W, Kischka P (2005) Dynamic, customer-oriented improvement of supply networks. Eur
J Oper Res 167:413–426
Jung H, Jeong B (2005) Decentralised production–distribution planning system using collaborative agents
in supply chain network. Int J Adv Manufact Technol 25:167–173
Jung H, Chen FF, Jeong B (2005) A production–distribution coordination model for third party logistics
partnership, 2005 IEEE international conference on automation science and engineering (CASE)
Karabuk S, Wu SD (2002) Decentralizing semiconductor capacity planning via internal market coordination.
IIE Trans 34:743–759
Kilger C, Reuter B (2005) Collaborative planning. In: Stadtler H, Kilger C (eds) Supply chain management
and advanced planning. Concepts, models, software and case studies. Springer, Berlin, pp 260–278
Knolmayer G, Mertens P, Zeier A (2002) Supply chain management based on SAP Systems, order man-
agement in manufacturing companies. Springer, Berlin
Landeros R, Monczka R (1989) Cooperative buyer/seller relationships and a firm’s competitive posture.
J Purchas Mater Manage 25(3):9–18
Li X, Wang Q (2007) Coordination mechanisms of supply chain systems. Eur J Oper Res 179:1–16
Lu L (1995) A one-vendor multi-buyer integrated inventory model. Eur J Oper Res 81:312–323
Marble RP, Lu Y (2007) Culturalizing enterprise software for the Chinese context: An argument for accom-
modating guanxi-based business practices. Int J Product Econ 107:364–379
Meijboom BR (1986) Horizontal mixed decomposition. Eur J Oper Res 27:25–33
Myerson RB (1991) Game theory: analysis of conflict. Harvard University Press, Cambridge
Monahan JP (1984) A quantity discount pricing model to increase vendor profits. Manage Sci
30:720–726
Milgrom P (2004) Putting auction theory to work. Cambridge University Press, Cambridge
Minner S (2000) Strategic safety stocks in supply chains. Lecture notes in economics and mathematical
systems, pp 490, Springer, Berlin
Roussel J, Kuehn H, Schoepke P (2002) Unlocking the value from supply chain planning solutions. Supply
Chain Forum Int J 3(1):30–33
28 H. Stadtler
Sarmah SP, Acharya D, Goyal SK (2006) Buyer–vendor coordination models in supply chain
management. Eur J Oper Res 175:1–15
Schneeweiss C (2003) Distributed decision making. 2nd edn. Springer, Berlin
Schneeweiss C, Zimmer K (2004) Hierarchical coordination mechanisms within the supply chain. Eur
J Oper Res 153:687–703
Shirodkar S, Kempf K (2006) Supply chain collaboration through shared capacity models. Interfaces
36:420–432
Simpson NC, Erengüc SS (2001) Modeling the order picking function in supply chain systems: formulation,
experimentation, and insights. IIE Trans 33:119–130
Stadtler H (2004) Advanced planning systems—basics and beyond. In: Fandel G, Backes-Gellner U,
Schlüter M, Staufenbiel JE (eds) Modern concepts of the theory of the firm, managing enterprises
of the new economy. Springer, Berlin, pp 285–300
Sucky E (2004a) Koordination in supply chains. Spieltheoretische Ansätze zur Ermittlung integrierter Bes-
tell- und Produktionspolitiken. Gabler, Wiesbaden
Sucky E (2004b) Coordinating order and production policies in supply chains. OR Spectrum 26:493–520
Thomas DJ, Griffin PM (1996) Coordinated supply chain management. Eur J Oper Res 94:1–15
Tsay A, Nahmias S, Agrawal N (1998) Modeling supply chain contracts: a review. In: Tayur S,
Ganeshan R, Magazine M (eds) Quantitative models for supply chain management, Boston
Vanderbeck F (2000) On Dantzig-Wolfe decomposition in integer programming and ways to perform
branching in a branch-and-price algorithm. Oper Res 48:111–128
Voluntary Interindustry Commerce Standards (2004): Collaborative planning, forecasting and replenish-
ment (CPFR). 14/07/06: http://www.vics.org/standards/CPFR_Overview_US-A4.pdf, Accessed on
14 July 2006
Whang S (1995) Coordination in operations: a taxonomy. J Oper Manage 12:413–422
Whang Y, Gerchak Y (2003) Capacity games in assembly systems with uncertain demand. Manufact Ser-
vice Oper Manage 5:252–267
Zimmer K (2001) Koordination im Supply Chain Management. Ein hierarchischer Ansatz zur Steuerung
der unternehmensübergreifenden Planung. Gabler, Wiesbaden
Simulation and optimization of supply chains:
alternative or complementary approaches?
Christian Almeder · Margaretha Preusser ·

Richard F. Hartl

OR Spectrum (2009) 31:95–119
DOI 10.1007/s00291-007-0118-z
Abstract Discrete-event simulation and (mixed-integer) linear programming are

widely used for supply chain planning. We present a general framework to support
the operational decisions for supply chain networks using a combination of an optimi-
zation model and discrete-event simulation. The simulation model includes nonlinear
and stochastic elements, whereas the optimization model represents a simplified ver-
sion. Based on initial simulation runs cost parameters, production, and transportation
times are estimated for the optimization model. The solution of the optimization model
is translated into decision rules for the discrete-event simulation. This procedure is
applied iteratively until the difference between subsequent solutions is small enough.
This method is applied successfully to several test examples and is shown to deliver
competitive results much faster compared to conventional mixed-integer models in
a stochastic environment. It provides the possibility to model and solve more real-
istic problems (incorporating dynamism and uncertainty) in an acceptable way. The
limitations of this approach are given as well.
Keywords Supply chain management · Optimization · Discrete-event simulation ·

Hybrid method
C. Almeder (B) · M. Preusser · R. F. Hartl

University of Vienna, Brünnerstr 72, 1210 Vienna, Austria
M. Preusser
R. F. Hartl

30 C. Almeder et al.
1 Introduction
In recent years intra-company supply chains have been growing significantly spanning
production and distribution sites all over the world. At the same time global competi-
tion has increased, such that there is a strong demand for new decision support tools
on strategic, tactical and operational levels. Biswas and Narahari (2004) classified the
relevant research on such decision support systems into three categories:
(a) Optimization models mainly for multi-echelon inventory control. In most cases
these models are deterministic and used for strategic or tactical decisions.
(b) Analytical performance models, which consider a dynamic and stochastic envi-
ronment. They are used to investigate design or principal management decisions.
Such systems are represented as Markov chains, Petri nets or queuing models.
(c) Simulation and information models, which are used to analyze complex dynamic
and stochastic situations and to understand issues of supply chain decision mak-
ing.
For the first and the second categories, it is often necessary to make several simpli-
fications from the real-world case in order to develop solvable models. Nevertheless
the problem size is usually very limited. Although there are promising developments
of combinations of these two categories (cf. Sect. 2), many of them remain on a stra-
tegic level and the stochastic property is considered by a small number of different
scenarios.
In this paper we develop a new solution approach by applying a LP/MIP formu-
lation in the context of a discrete-event simulation. So we are able to combine the
advantages of models from all categories mentioned above by considering a detailed
representation of a dynamic and stochastic environment and allow the application of
optimization methods in this context. Our investigations are based on a general supply
chain network model with different facilities (suppliers, manufacturers, distributors)
and different transportation modes connecting these facilities. We assume that there is
a central planner with perfect information such as for intra-company supply chains or
supply chains with a dominant member. This problem setting is motivated by a case
study about a global supply chain network in the paper industry (Gronalt et al. 2007).
The goal is to reduce costs by simultaneously optimizing the production/transportation
schedule and reducing inventory levels. We are aiming for a robust solution, in the
sense that a stochastic environment is considered. Comparing our problem to the tasks
in the supply chain matrix (cf. Stadtler 2005), the problem is a combination of several
operational tasks: production planning, distribution planning and transport planning.
In addition to other approaches, we assume a stochastic simulation model for these
tasks and combine it with classical optimization approaches.
Our goal is to achieve an optimal operation plan for supply chain networks by com-
bining optimization models and simulation models. We do not use the optimization on
top of the simulation, where an optimization algorithm uses the simulation model as a
black-box evaluation function (cf. Glover et al. 1999). Instead we include simulation
and optimization in an iterative process in order to gain the advantages of optimiza-
tion (exact solution) and simulation (nonlinearities, complex structure, stochasticity).
In the previous research (Almeder and Preusser 2004; Preusser et al. 2005a,b) we
Simulation and optimization of supply chains 31
Solutions of simulation
simulate
experiments aggregate
Decision rules in
Optimization
D-E model
model
derive optimize
Linear solution
Fig. 1 Interaction between simulation and optimization
developed a rough idea of this concept. In the current paper we extended this con-
cept, such that it is possible to apply it to a wide range of supply chain problems.
Furthermore we analyze in detail the advantages and disadvantages of this approach
and present results for different test cases.
The supply chain is represented as a discrete-event model (D-E model) and a sim-
plified version is modeled as an optimization model. We start by performing several
simulation runs in order to get average values of the parameters (e.g., unit transportation
costs) which are then fed into the optimization model. After solving the optimization
model the result is transformed into decision rules that are used in the discrete-event
model. Then we start again with further simulation experiments (see Fig. 1), and so
on.
Our contribution is twofold:
• Development and analysis of a general framework (Fig. 1) and a toolbox for the
combination of discrete-event simulation and optimization of supply chains.
• For stochastic supply chains, an iterative combination of simulation and linear
programming is empirically shown to be competitive compared to deterministic
MIP-models.
The paper is organized as follows: We start with a literature review in Sect. 2, fol-
lowed by a description of the general model framework in Sect. 3. In Sect. 4 we explain
the linkage between the simulation model and its linear version. Finally, we report on
different test results in Sect. 5 and give conclusions and an outlook for possible future
research in Sect. 6.
2 Literature review
2.1 Supply chains
Aspects of the integration of transport and production planning within supply chains
have been investigated in several papers (cf. Erengüc et al. 1999). Combined plan-
ning approaches for different decision levels (e.g., tactical and operational decisions)
can be found in Meyr (2002) and Schneeweiss (2003). There are numerous papers
dealing with linear or mixed-integer programs for supply chain networks and network
flows. Yaged (1971) discussed in his paper a static network model which includes
nonlinearities. He tries to optimize the product flow by solving a linearized version of

the network and to improve the flow in the network. Paraschis (1989) discussed several
different possibilities to linearize such networks and Fleischmann (1993) presented
several applications of network flow models, which are solved through linearization.
Pankaj and Fisher (1994) showed that based on an MIP model the coordination of
production and distribution can reduce the operating cost substantially. Dogan and
Goetschalckx (1999) showed that larger supply chain design problems can be solved
using decomposition. A recent case study about a supply chain of the pulp indus-
try modeled as a MIP is given by Gunnarsson et al. (2007). In general the problems
solved with LPs and MIPs usually include several simplifications in order to keep
them solvable.
Recent publications also included stochastic elements in the optimization models.
Santoso et al. (2005) considered a stochastic programming approach for the sup-
ply chain network design. They used a sample average approximation and Benders
decomposition to solve design problems for a supply chain while considering future
operational costs. For that purpose they developed a linear model with uncertain cost
factors and demand. Although they used a fast algorithm, realistic problems with
sample sizes of up to 60 scenarios need several hours to be solved. Alonso-Ayuso et
al. (2003) considered a similar combined design and operation problem. Their sto-
chastic programming approach was able to solve medium sized problems with about
100 binary decisions within 1 h. Leung et al. (2007) presented a robust optimization
model for a simultaneous production planning for several sites in a supply chain under
uncertainty. But still they are restricted to rather small models and consider only four
different scenarios.
In the field of supply chain simulation Kleijnen (2005) gave a short overview of
simulation tools and techniques used for supply chains. He distinguished between
four different approaches: spreadsheet simulation, system dynamics, discrete-event
dynamic systems simulation, and business games. Clearly, discrete-event simulation
is the most powerful tool to consider complex stochastic systems. Numerous software
packages for discrete-event simulation are available, both very specialized ones for a
specific part of the supply chain and general ones with a high functionality in modeling
and visualization of supply chains (cf. Kelton et al. 2002; Kuhn and Rabe 1998). One
example is the Supply Net Simulator presented by Stäblein et al. (2007). It allows
simulating the behavior of individual members in a supply chain network. They used
an agent-based approach, where each member optimizes its own operations in the
sense of an advanced planning system. But there is no interaction between simulation
and optimization.
2.2 Optimization and simulation
Most of today’s simulators include possibilities to do a black-box parameter optimiza-

tion of a simulation model. Glover et al. (1999) presented the successful development
of OptQuest (© OptTek Systems, Inc.1 ), an optimization toolbox containing different
1 http://www.opttek.com.
algorithms (mainly metaheuristics) designed to optimize configuration decisions in

simulation models. The simulation model is used only for the evaluation of the objec-
tive value, no further structural information is used. Swisher et al. (2000) and Fu (2002)
stated in their papers that there is still a big gap between optimization methods for
simulation-based optimization used in commercial software and methods available in
research literature.
Truong and Azadivar (2003) developed an environment for solving supply chain
design problems, where they combine simulation with genetic algorithms and mixed-
integer programs. Strategic decisions regarding facility location and partner selection
are considered.
The work by Lee and Kim (2002), possibly the most related work in this context,
shows a combination of simulation and optimization for the case of a production-
distribution system. They use simulation to check the capacity assumptions used for a
simpler linear model in a more realistic environment with stochastic machine break-
downs and to update these capacity parameters for the optimization. After several
iterations they end up with a solution of the optimization model which is also within
the constraints of the stochastic simulation model. Their method is quite similar to
our approach, but they aim for more realistic capacity estimation for the optimization
model. In contrast, we try to find a robust plan for production, stocking, and transpor-
tation considering stochastic and nonlinear operations and costs by estimating delays
and cost factors based on simulation experiments.
3 The supply chain network model
The general description of the supply chain originates from a case study about a supply
chain in the paper industry (cf. Gronalt et al. 2007). Several production sites are used
to manufacture different paper products, which are delivered either directly or via
hubs to customers all over the world. The main task in this case study was to develop
a 1-year plan for production quantities and transportation links. In this case study a
static model was developed, which was used to get rough estimates quickly.
Inspired by this case study we formulate the following problem setting. The basis
for our supply chain model is a predefined network, i.e., the locations of all actors and
the connections between them are given. Within the network we differentiate between
three types of participants connected by transportation links:
• suppliers providing raw materials;

• customers who demand certain products at a specific time;
• production/warehouse sites where production, stocking, and transshipment takes
place.
The whole supply chain is order-driven, that means products are manufactured or
transported only if a subsequent member of the supply chain requests it. So the origin
for all activities is the predefined deterministic demand of the customers. All activities
are based on time periods, which might be days or shorter time periods.
The suppliers are used as source for raw materials, which are sent to production sites
if requested. Production/warehouse sites can store incoming products. These products
can be used to manufacture new products, or they are simply transferred to the output
inventory. From there they are sent to subsequent members of the supply chain.
The simulation model is implemented using AnyLogic (© XJ Technologies), a
Java-based simulation tool. The model is constructed as a library including several
different modules. These modules represent the four types of participants in the supply
chain network plus a general control module necessary for controlling the simulation
experiments as well as the communication with the optimization model which was
developed using Xpress-MP (© Dash Optimization). This model is a simplified deter-
ministic version of the library modules of the simulation model. In this section we will
explain the different modules of our supply chain network.
3.1 Module supplier
Simulation model. This module is used to generate certain products, store them, and
deliver them if demanded. It has one input port to receive orders for products and
one output port to deliver products. If this module receives an order through the input
port, then it sends the requested amount of products via the output port. If the amount
exceeds the current inventory level, only the available amount is sent. As soon as new
products arrive in the inventory they are delivered until the whole order has been ful-
filled. The costs arising in this module are only inventory costs for storing products
prior to delivery. These costs may have any user-defined functional form. According
to the given parameters in each period, new products are generated and added to the
stock.
Optimization model. We also developed a simplified representation as an optimi-

zation model. We denote by JS the set of suppliers within a network, by P the set
of products, and by T the number of periods. The representation of the supplier’s
behavior in the optimization model can be formulated as follows (If p and t are free
indices, i.e. not used as a summation index, then the set of equations is meant to be
valid ∀t = 1, . . . , T, p ∈ P.):

T
p out p
TCiS = out
Hi li (t) ∀i ∈ JS , (1)
p∈P t=1
out p out p p p
li (t) = li (t − 1) − out
f i (t) + Si (t) ∀i ∈ JS , (2)
out p
li (t) ≥ 0 ∀i ∈ JS . (3)
For a complete list of parameters and variables, see Appendix C. The overall cost
p
of supplier i is denoted by TCiS , consisting only of the holding cost out Hi (·) of the
output inventory, expressed by the right-hand side of (1). Equations (2) are the inven-
p
tory balance equations for the output inventories out li (t). The stock is diminished by
p p
the outflow of materials, out f i (t), and increased by the given supply Si (t). The last
set of constraints (3) guarantees that the inventory level cannot be negative.
The simulation and the optimization model are connected via the holding costs in
(1) which represent the user-defined cost function in the simulation.
3.2 Module production
Simulation model. This module is the core of the whole model. It represents a
production site as well as a transshipment point. It consists of an input and an out-
put storage. Items are either transformed into new items or simply transferred from
the input to the output storage. This module has one input port and one output port
for orders, as well as one input and one output port for products. The input storage
is replenished by ordering products via the output port for orders from a supplier
or another production module. The ordering policy may be either autonomous (e.g.,
an (s,S)-policy or any user-defined policy) or it is determined by the result of the
optimization model. Products are received through the product input port and stored
in the input inventory. The production of new products or the transfer of products
is initiated by an internal order placed by the output inventory (either autonomous
or based on the solution of the optimization model). The delay for production and
transfer is a user-defined function. It may contain stochastic elements and depend
on other parameters (e.g., the current load). Production and transfer have limited
capacities and furthermore production is restricted to the availability of raw materials
(other products). If these capacities do not allow producing (or transferring) a lot as a
whole, it is split into several batches. Through the input order port the module
receives orders from other production or customer modules. Products are sent through
the output product port according to these orders and based on availability. Costs arise
in this model for inventory holding (input and output), for production, and for transfer.
Optimization model. The optimization model for the production node is as follows
(we denote by J I the set of production nodes in the supply chain network):

T
p p
T
p p
TCiI = Wi m i (t) + Z i u i (t)
p∈P t=1 p∈P t=1

T
T

p in p out p out p
+ in
Hi li (t) + Hi li (t) ∀i ∈ J I , (4)
p∈P t=1 p∈P t=1
p p
p p
m i (t) ≤ prod
Capi (t), ai · m i (t) ≤ prod
Ci (t) ∀i ∈ J I , (5)
p∈P
p p
p p
u i (t) ≤ ta Capi (t), di · u i (t) ≤ ta
Ci (t) ∀i ∈ J I , (6)
p∈P
in p p p
li (t) = in li (t − 1) + in f i (t)
p p
αi ( p ) · m i (t) − u i (t) + ri (t)
p p
− ∀i ∈ J I , (7)
p ∈P
out p out p p p p
li (t) = li (t − 1) − out
f i (t) + χt≥δ p · m i (t − δi )
i
p p p
+χt≥σ p · u i (t − σi ) + si (t) ∀i ∈ J I , (8)
p
i
in p p p
li (t) ≤ invin
Capi (t), qi · in li (t) ≤ in
L i (t) ∀i ∈ J I , (9)
p∈P

out p p p p
li (t) ≤ invout
Capi (t), qi · out li (t) ≤ out
L i (t) ∀i ∈ J I , (10)
p∈P
p p in p out p
m i (t) ≥ 0, u i (t) ≥ 0, li (t) ≥ 0, li (t) ≥0 ∀i ∈ J I . (11)
The overall cost of a production node i is represented by TCiI . These costs

p
consist of production costs (the production amounts are denoted by m i (t)) trans-
p
fer costs (the transfer amounts are denoted by u i (t)) and the holding costs of the
input and the output inventory. Constraints (5) and (6) restrict the production and the
transfer for each product individually, as well as for the total production and transfer.
In the latter case the amounts are multiplied by the resource requirements. The dis-
tinction between individual capacity constraints for each product and global capacity
constraints are necessary to cover general situations where different resources as well
as common resources are necessary for production. Equations (7) are the inventory
balance equations for the input inventories. The current inventory level is determined
by the inventory level of the previous period, the inflow from other nodes, the required
raw materials for production, the transfer amount and some external inflow (from out-
side of the system); αi ( p ) represents the units of raw material p which is necessary
p
to produce one unit of product p . The inventory balance equations for the output
p p
inventories (8) are similar, but the production and transfer delays (δi , σi ) have to be
considered before a new product arrives in the output inventory. Function χt≥ε is an
indicator function, used in order to avoid the use of production and transfer amounts
for negative periods. Equations (9) and (10) are used to restrict the stock of the input
and the output inventory (for each product separately and accumulated using the space
p
requirements qi ). These two types of restrictions allow modeling a dedicated-storage
as well as a random-storage policy.
The simulation model and the optimization model are connected through the cost
p p
factors in Eqs. (4) and production and transfer delays (δi , σi ), which are user-defined
functions in the simulation model possibly containing stochastic and nonlinear
p p
elements. Furthermore the production and transfer amounts (m i (t), u i (t)) of the
optimization model are used to determine production plans in the simulation model.
p
For example, if m i (t) > 0 for a specific product and period, then in the simulation
p
model the amount given by m i (t) is ordered in period t.
3.3 Module customer
Simulation model. According to a given demand table, the customer places orders at
the production sites. Due to stochastic features within the simulation, it is not possible
to time deliveries exactly. Therefore the customer has an input inventory, which is
used to satisfy the demand. The inventory level can be negative (shortages), as well as
positive (oversupply). In both cases penalty costs occur, which are higher for shortages.
The module has one output port for sending requests and one input port for receiv-
ing products. The orders are sent either according to the demand table (including a
standard delay time for transportation) or according to the solution of the optimization
model.
Optimization model. The optimization model for the customers’ behavior can be
written as follows (we denote by JC the set of customer nodes in the supply chain
network):

T
p p
TCiC = Ri (in bi (t)) ∀i ∈ JC , (12)
p∈P t=1
in p p in p p p
li (t) − in bi (t) = li (t − 1) − in bi (t − 1) + in f i (t)
p p
−Di (t) + ri (t) ∀i ∈ JC , (13)
in p
li (t) =0 ∀i ∈ JC . (14)
In (12) we calculate the cost at the supplier which consists only of penalty cost
p
for back orders in bi (t). Equations (13) are the inventory balance equations where the
p
customers’ demands Di (t) are considered. It is assumed that all customers are just-
in-time customers. Therefore, constraints (14) ensure that no oversupply (positive
stock level) is possible, i.e. it is not allowed to send more products than demanded
by the customers. This JIT assumption may be dropped and holding costs for positive
stock may be included. In the simulation model the JIT assumption is weakened,
because stochastic transportation times may cause an unwanted early delivery. These
early deliveries are penalized.
The differences between the simulation model and its representation as an optimi-
zation model are the penalty cost factors in (12) and the JIT assumption expressed in
(14).
3.4 Module transport
Simulation model. This module is used to transport products between different mod-
ules. It receives products through its single input port and sends it (according to some
time delay) through the output port to the next module (Production or Customer). It
has a limited capacity and organizes the transports according to a FIFO rule. It is also
possible to split shipments if the available capacity does not allow single shipment.
The time delay may be stochastic and may depend on other parameters. User-defined
costs arise for transportation and may include transportation time, amounts, and fixed
charge parts.
Optimization model. The representation of the transport modules as an optimization

model can be formulated as follows. Each transport module is identified by the indi-
ces of the nodes, which it connects. Furthermore, we need an additional index v ∈ V
denoting the different transport modes (if v is not used as a summation index, the
equations are valid for all v ∈ V ):

T
v p v p
TCiTj = Ci j xi j (t) ∀i ∈ JS ∪ J I , j ∈ J I ∪ JC , (15)
t=1 v∈V p∈P

v p v v p
g · v xi j (t) ≤ v
p p
xi j (t) ≤ Capi j (t), Ci j (t) ∀i ∈ JS ∪ J I , j ∈ J I ∪ JC ,
p∈P
(16)

p v p v
in
f j (t) = xi j (t − τi j ) ∀ j ∈ J I ∪ JC , (17)
i∈Js ∪J I v∈V
v τ <t
ij

p v p
out
f i (t) = xi j (t) ∀i ∈ JS ∪ J I , (18)
j∈J I ∪JC v∈V
v p
xi j (t) ≥0 ∀i ∈ JS ∪ J I, j ∈ J I ∪ JC . (19)
The total transportation cost for transports from member i to member j of the
supply chain network is denoted by TCiTj and v xi j (t) gives the transportation amount
p
for each period, each product, and each transportation mode. Constraints (16) limit
the transportation to a product-specific and an overall capacity limit. Equations (17)
and (18) represent the inflow of products to member j and the outflow of products
from member i.
The connection between the simulation and the optimization model is established
by the transportation cost functions in (15) and the transportation delays v τi j in (17) on
the one hand, and through the transportation amounts v xi j (t) on the other hand. These
p
transportation amounts are used to define ordering schemes for the Production and
Customer modules in the simulation model. That means if v xi j (t) > 0, then member
p
j of the supply chain sends a request for v xi j (t) units of product p to member i of the
p
supply chain at time t.
3.5 Supply chain optimization model
The optimization model of the whole supply chain network is defined by minimizing
the total cost

min TCiS + TCiI + TCiC + TCiTj (20)
i∈JS i∈J I i∈JC i∈JS ∪J I j∈J I ∪JC
subject to the constraints (1)–(19). If we assume that all cost functions are linear, i.e.
that the objective (20) is a linear function, we can write it as follows:

v p
· v xi j (t) +
p p p
min ci j wi · m i (t)
i j∈J p∈P t=1,..T v∈V i∈J I p∈P t=1,..T
p p

in p p
+ zi · u i (t) + hi · in li (t)
i∈J I p∈P t=1,..T i∈J I p∈P t=1,..T

out p p p p
+ hi · out li (t) + ρi · in bi (t). (21)
i∈JS ∪J I p∈P t=1,..T i∈JC p∈P t=1,..T
supplier production customer
transport_1 transport_2
param (LP) param (LP) param (LP)

p p
out
hi
p wi , z i , in p
ri
param (LP) in
hi
p
,
out p
hi , param (LP)
v p v v p v
cij , t ij p p c ij , t ij
di , s i
DV (Sim) DV (Sim)
v p
DV (Sim) v p
x ij (t) p p x ij (t)
m i (t) , u i (t)
Fig. 2 An example of the module configuration for a simple supply chain consisting of one supplier, one
production site, and one customer (dashed lines indicate information flow and solid lines indicate material
flow). For the case of a linear program the information below the modules represents the parameters which
are calculated during the simulation runs and transferred to the LP, param (LP), and the decision variables
of the LP used as decision rules in the simulation model, DV (Sim.)
Hence, we get a linear programming model which we will use in connection with
the simulation model as depicted in Figure 1. A detailed description of the linear model
can be found in Preusser et al. (2005a,b).
This model is a pure linear program which can be solved easily with any standard
LP-solver within very short time. If necessary, it is possible to extend the model formu-
lation to consider more features, e.g., fixed-cost transportation, binary decisions, step
functions, etc. These extensions lead to a mixed-integer formulation, thus increasing
computational time (cf. Appendix A).
3.6 Supply chain simulation model
Implementing a simulation model in AnyLogic means to arrange the according mod-

ules and connect them. In Fig. 2 an illustrative example shows, how these modules
can be connected in order to maintain information flow (direct connections bet-
ween Supplier, Production, and Customer) and material flow (via Transport modules).
Furthermore the cost and delay functions for each module must be specified.
4 Connecting the optimization with the simulation
In order to couple the optimization model and the simulation model, we first have to
define the required data and the way they should be exchanged. We decided to use an
MS Access database to store all necessary information which is:
• General network structure: This includes the number of actors in the supply chain
and the according links between them.
• General parameters used in the simulation and optimization models: These sets of
parameters include all capacity limitations, resource parameters, bill-of-materials,
predefined supply at the suppliers, and predefined demand at the customers.
aggregated results (transportation delays, production delays,…)
general general
simulation parameters MS Access parameters optimization
model database model
(AnyLogic) (ODBC) (Xpress)
decision rules (ordering plans, production schedules,…)
Fig. 3 This scheme shows the data exchange between the simulation and the optimization model via the
MS Access database in the middle
• Results of the optimization model (= parameters for the simulation): The results
of the optimization used as decision rules in the simulation are production and
transfer quantities, as well as transportation amounts (m i (t), u i (t), v xi j (t)).
p p p
• Results of the simulation model (= parameters for the optimization model): The
main results of the simulation experiments used in the optimization model are the
cost parameters and the delays for production, transfer and transport.
The simulation model is designed as the master process, which controls the data
communication and the LP/MIP-solver. The simulation model and the optimization
model retrieve and store values from and to the database using the Open Database
Connectivity (ODBC) interface (see Fig. 3).
To initiate the optimization process in our system, a few simulation runs are per-
formed using the data from the database. Missing decision rules which, in later iter-
ations, are generated using the results of the optimization model, are substituted by
autonomous decision rules (like an (s,S)-policy for the replenishment). These first
simulation runs are only necessary to generate initial parameter values for the opti-
mization model, but their results will be ignored in further iterations in order to avoid
biasing effects caused by the autonomous decision rules. The results of the initial runs
(delays, per unit costs, etc.) are aggregated and according means and variances are
stored in the database (see Sect. 4.1). Afterwards Xpress-MP is executed. It loads the
general data and the simulation results from the database, computes the solution of
the optimization model and stores the results (ordering and delivery plans, production
and transfer schedules, etc.) in the database. Then we start again with five simulation
experiments using now the newly computed decision rules (see Sect. 4.2), based on
the solution of the optimization model. Further on we will denote this algorithm by
SimOpt (or SimLP, if a pure LP model is used for the optimization part, and SimMIP,
if a mixed-integer formulation is used). In Table 1 a pseudo code of this SimOpt
algorithm is given.
4.1 Aggregating simulation results
Since the simulation model may contain stochastic and nonlinear elements, it is
necessary to perform several simulation runs and combine the results. For the cost
parameters, necessary for a linear optimization model, we calculate average per unit
cost. That means, e.g., for the production costs we accumulate the total cost for the
Table 1 Pseudo code for the combined simulation optimization approach SimOpt
SimOpt:
Load necessary simulation parameters from the database
Perform a few simulation runs using autonomous decision rules
Aggregate results and store them in the database
while stopping criteria are not met
Load aggregated parameters into LP/MIP-Solver
Solve the optimization model
Write new decision rules to the database
Load new decision rules into simulation model
Perform simulation runs using these decision rules
Aggregate results and store them in the database
end-while
whole planning horizon for a certain product and divide these costs by the number of
products produced. Other parameters, called critical parameters, have a direct influ-
ence on the material flow (e.g., transportation delay). The use of average values for
those parameters would most probably lead to bad results. In about half of the cases the
delay would be longer than assumed and would cause additional delays in subsequent
operations. Therefore, it seems reasonable to use, e.g., a 90%-quantile (based on a
normal distribution with estimated mean and variance calculated from the different
simulation runs) for such delay parameters. This results in an overestimation of the
delays for the optimization model, because the time is determined such that 90% of
the observed delays will be shorter, but it ensures that a smooth material flow through
the network is possible.
For the critical parameters it is useful to combine results from the previous itera-
tions with current ones, in order to enlarge the sample size and to get better estimates
of the mean and the variance.
4.2 Decision rules based on the solution of the optimization model
There are several possible ways to use the solution of the LP-model within the simula-
tion model. One method, which we apply here, is to use the transportation, production
and transfer results (v xi j (t), m i (t), u i (t)) as a given ordering plan. According to the
p p p
transportation values v xi j (t) module j sends a request to module i at time t for the
p
given amount v xi j (t) of products of type p using the transportation mode v. Simi-
p
larly, production and transfer results can be used. In some cases, due to the stochastic
features of the simulation model, it may happen that some of the modules are out-
of-stock for a specific product. Since unfulfilled orders are backlogged, these requests
are fulfilled as soon as the products are available.
More complex procedures would be, e.g., to use results of the sensitivity analysis
(dual variables, reduced costs) to determine the critical parameters, to observe these
parameters during the simulation runs, and to adapt decision rules (use a different
310000
simulation
260000 optimization
total cost
210000
160000
110000
60000
10000
1 -O p t
2 -O p t
3 -O p t
4 -O p t
5 -O p t
6 -O p t
1 -S i m
2 -S i m
3 -S i m
4 -S i m
5 -S i m
6 -S i m
iteration
Fig. 4 Objective values of the optimization model and the simulation model for each iteration for a deter-
ministic model considering fixed costs for production, transfer and transport
solution of the optimization model) if the observed parameters reach a certain thresh-
old.
For our examples we use the first approach for translating the solution of the opti-
mization model into decision rules for the simulation model (cf. Sect. 3). In this paper
we wish to investigate the direct interactions between the solution of the optimization
model and the simulation results. The analysis of more complex decision rules goes
beyond the scope of this paper and might by a subject for further research.
5 Tests and results
We wish to investigate the following research questions with empirical tests using a
set of test instances:
• Does this method converge in practice for realistic test cases?
• If we can observe convergence, is the result optimal or at least a good approxima-
tion?
• Is this method advantageous compared with traditional planning methods?
Although it is not possible to prove general convergence for all our test instances,
we observe fast convergence of the objective values of the simulation and optimization
model. Figure 4 shows a typical situation using the results of the deterministic test
instance D1-L described in Sect. 5.1.
We start with the simulation model using an autonomous rule for replenishing the
inventories. Since we start with all inventories empty, it takes a long time, until the
orders are fulfilled. This causes high costs and an overestimation of transportation
and production delays. Therefore, the first solution of the linear model leads also to a
high objective value mainly consisting of penalty costs for late (or even no) deliveries.
Consequently, the simulation model leads to a similar objective function in iteration
2, because it uses the delivery plans of the solution of the linear model. Due to the fact
that the solution of the linear model causes a somehow synchronized material flow,
the measured delays are much lower now. Therefore, the cost of the solution of the
linear model in the second iteration decreases. After three iterations the simulation
and the linear model have converged to the same solution.
5.1 Deterministic problems with fixed costs
In order to verify the quality of the solutions, we create a set of 12 examples. For these
test instances we consider a simple supply chain consisting of three actors (a supplier,
a producer and a customer) and a time horizon of 30 periods. For transportation of
products two transport modules are used, which connect the supplier and the producer
as well as the producer and the customer. Two types of products are demanded by the
customer: product 1 which is provided by the supplier and sent via the producer to the
customer and product 2 which is manufactured by the producer using product 1 as a
raw material. The cost structure is as follows:2
• The transportation costs consist of fixed costs per delivery, which are subject to a
step function. A transport costs 100, 200 or 300 monetary units, depending on the
amount delivered.
• The costs of production and transfer are separated into variable costs and fixed
costs. The variable production costs are set to 30 and the fixed part is 50 monetary
units. Transferring products costs 15 per product unit plus a fixed part of 10.
• Delayed deliveries are penalized by 100 monetary units per product unit and period.
Concerning the demand at the customer we distinguish between instances with high
demand and others with low demand. The difference lies in the frequency of orders
sent off by the customer. In high demand cases the occurring orders in each period
are around the maximum possible quantities which could be delivered considering the
capacities of the supplier and the producer. In low demand models the ordered amounts
cover approximately 70% of the possible deliveries in each period. Instances D1-L
to D5-L (see Table 2) represent five different realizations for low demand models.
Accordingly, D6-H to D10-H correspond to five different realizations of high demand
models. The last two instances, D1a-L and D6a-H, are modifications of instances D1-L
and D6-H, respectively. The former ones consider exactly the same ordering amounts
as D1-L and D6-H but the fixed costs for production and transfer at the intermediate
node are increased to 1000 and 500, respectively.
For examples of this size it is possible to formulate an exact mixed-integer model
and determine the optimal solution. The corresponding MIP formulation consists of
1,342 constraints, 1,080 continuous and 300 binary decision variables. For the simu-
lation approach the nonlinear parts are only considered in the simulation model itself,
the connected linear (non-integer) model does not include any of them. See Table 2
for the resulting total costs of the simulation and the optimal solution (MIP).
The gap between our SimLP approach the optimal solutions gained by solving the
exact MIP formulation varies between 0.44 and 3.46% and averages in 1.87%. As
we would expect, for the two test instances with high fixed cost the gap increases.
2 All datasets are available at http://www.univie.ac.at/bwl/prod/download/SCM-Data.

Table 2 Comparison of total

Instance SimLP Exact MIP Difference (%)
costs between our
simulation-based optimization
D1-L 53640 52947 1.31
approach SimLP and the exact
MIP-model for deterministic D2-L 55032 53860 2.18
test cases classified by the D3-L 52626 52394 0.44
occurrence of customer demand D4-L 54442 53600 1.57
(H—high demand, L—low
demand) D5-L 55198 54057 2.11
D6-H 59885 58830 1.79
D7-H 61257 60129 1.88
D8-H 59028 58347 1.17
D9-H 60403 59501 1.52
D10-H 61436 60365 1.77
D1a-L 63720 61587 3.46
MIP solutions marked with (*) D6a-H 76165 73761* 3.26
are best solutions found after
one hour calculation time Average 77165 73760 1.87
For test instances with low demand the variation of the gap seems higher than for the
test instances with high demand. But on average there seems no significant difference
between the results of those two groups of instances.
Based on the above results we may conclude that the error caused by neglecting
fixed cost is low as long as the fixed costs are low compared with other costs. If the
fixed costs increase (relative to the other costs), the nonlinear properties should be
considered in the optimization model used for the SimOpt approach, i.e. a SimMIP
should be used instead of the SimLP (see also the following subsection).
5.2 Test instances with stochastic transport delays
In order to measure the quality of our solutions in a stochastic environment we prepare

a set of test examples including stochastic transportation times. We compare our Sim-
LP approach using a simplified linear model without binary variables with an exact
MIP-model. This MIP-model does not cover stochastic features and we have to provide
estimated values of the transportation times. Within the simulation we consider uni-
formly distributed transportation delays between 1 and 9 for transportations from the
supplier to the producer and between 1 and 5 for transportations from the producer to
the customer. For estimating the delay parameters we perform runs using 90%-, 70%-,
and 50%-quantiles. The corresponding transportation delays for the MIP-model are
set according to the used quantile. For the small test cases there would probably be
no noticeable difference between the results of a 99%- and a 90%- quantile. There-
fore, we test a high quantile (90%, risk averse), the average value (50%), and some
intermediate value (70%). The maximum runtime is set to 30 minutes for the MIP-
model, i.e. we report the best solution found after this time limit, while the simulation
approach converges after a few seconds. For the simulation approach we again process
8 iterations, each consisting of 5 simulation runs and one LP computation. Finally the
solution of the MIP-model and the solution of the SimLP are evaluated by performing
Table 3 Difference of the mean total costs of 20 runs between the SimLP and SimMIP method and the
solution found by a deterministic MIP model classified by the occurrence of customer demand (H—high
demand, L—low demand)
Instance MIP SimLP SimMIP
Cost Quant. (%) Cost Quant. (%) Diff. (%) Cost Quant. (%) Diff. (%)
S1-L 66400 90 62601 90 −5.72 61637 90 −7.17

S2-L 61338 90 60635 90 −1.15 60282 90 −1.72
S3-L 63323 90 63566 70 0.38 63618 70 0.47
S4-L 63122 90 64067 90 1.50 64060 90 1.49
S5-L 60954 90 62399 90 2.37 62229 90 2.09
S6-H 72485 90 72342 90 −0.20 70871 90 −2.23
S7-H 70928 90 70751 90 −0.25 71040 90 0.16
S8-H 73257 90 77537 70 5.84 74999 70 2.38
S9-H 73501 90 74637 70 1.55 72845 90 −0.89
S10-H 71606 90 70230 90 −1.92 73934 90 3.25
S1a-L 71511 90 71686 90 0.25 70350 90 −1.62
S6a-H 88442 70 90582 90 2.42 88442 70 0.00
Av.-L 64441 64159 −0.39 63696 −1.08
Av.-H 75037 76013 1.24 75355 0.44
Average 69739 70086 0.42 69526 −0.32
The total costs are reported in the columns Cost. The quantile which lead to the best results is reported
in the column Quant. The difference with respect to the solution of the deterministic MIP is denoted in
column Diff.
20 independent simulation runs. Furthermore, we replace the simplified linear model

with the MIP model (SimMIP approach) and perform the same tests. The results for
all three methods are displayed in Table 3, where negative percentage values imply
that the simulation achieves a better result than the exact MIP-model.
The results of the SimLP approach, where we combine the pure LP model with
the simulation model, are on average slightly worse compared with the deterministic
MIP approach. For the low demand cases alone we can observe a small improvement.
Considering the variation of the results these differences are not significant.
Furthermore, we test a second approach, the SimMIP approach, where all nonlinear
features of the model are also considered in the optimization part which is represented
by a MIP model. So the difference between the SimMIP and the exact approach is only,
that in the first case the parameters are estimated based on simulation experiments and
in the latter case the parameters are determined using the known distribution functions.
Here we see that the solution quality can be raised for 9 out of 12 test instances and
on average the result is slightly better than the deterministic MIP approach.
Even if for some instances the 70%-quantile yields the best results, the 90%-
quantile leads to only slightly higher costs. Using the 50%-quantile, i.e. the expected
value, always caused much higher costs.
Additionally we analyze the variation of the 20 final simulation runs for each
method. There is no significant difference of the variation for all methods. The coef-
ficient of variation for the total costs is always around 0.07.
For larger test instances it would not be possible to solve the MIP or to apply the Sim-
MIP method. In a preliminary study (cf. Mitrovic 2006) we focused on that issue and
tried to find the approximate limits of solving MIP formulations of supply chain prob-
lems by the means of three state-of-the-art LP/MIP-solvers. The tests were conducted
on a PC (Intel P4 2.4 GHz, 1 GB RAM) using Windows 2000. We used a set of different
sized supply chain network problems considering fixed costs in transportation, pro-
duction and transfer. The best performing solver succeeded in solving problems with
20 supply chain actors, 8 products, 5 periods and 2 transportation modes, considering
3,360 binary variables before and 250 binary variables after presolving. The next in
size, which included 10 products instead of 8, could not be solved within a time limit
of one hour. Thus, there is a trade-off between a good approximation resulting from
MIP-models or fast computational times. Definitely, important decisions involving
high fixed costs should be considered within the optimization model of our SimOpt
method.
5.3 Quantile tests on larger instances
In addition to the small instances used in the previous subsections we generate a set of
12 instances representing larger supply chain networks. Using these test instances we
analyze the influence of the quantile on the results if the uncertainty is concentrated in
a specific part of the supply chain. The size and structure of these test cases is shown
in Fig. 5.
This fictitious supply chain network consists of 10 actors: 3 suppliers, 4 production
nodes, and 3 customers. The intermediate nodes are separated into two layers and all
of them are authorized to produce and also transfer products. All actors are connected
sup 1 cust 1
prod 1 prod 3
sup 2 cust 2
prod 2 prod 4
sup 3
cust 3
Fig. 5 Exemplary supply chain network. For simplicity the transport modules have been omitted
Table 4 Total costs of SimLP

Instances Quantile
for test example with ten supply
chain actors and stochastic 90% 70% 50%
transportation delays at the
beginning of the supply chain L1-L-S 274239 −0.19% 11.20%
L2-L-S 274241 9.98% 6.00%
L3-L-S 274995 5.76% 10.85%
L4-L-S 275366 5.83% 11.51%
L5-L-S 273214 1.33% 9.80%
L6-H-S 270286 2.72% 10.10%
L7-H-S 270491 2.23% 4.51%
L8-H-S 270438 −0.59% 10.92%
The results for the 90%-quantile
L9-H-S 267766 2.71% 8.65%
are taken as basic values. For the
remaining quantiles the L10-H-S 270155 −0.57% 10.12%
difference to the corresponding L1a-L-S 333772 2.77% 4.04%
basic value is given. The S
L6a-H-S 346953 1.42% 4.52%
indicates that there is more
stochasticity close to the supplier Total Avg. 283493 2.78% 8.52%
by one transportation mode. The customers request 4 different products. Products 1

and 2 are on the one hand final products, which have to be delivered to the customers
and on the other hand raw materials used to produce products 3 and 4.
First we consider the case with the stochasticity concentrated at the beginning of
the supply chain. Hence, for the connection between the suppliers and the first layer
of production sites we assume stochastic transportation times, which are uniformly
distributed between 1 and 5. The transportation times between the two layers of pro-
duction nodes are uniformly distributed between 1 and 3. For the remaining links we
assume deterministic transportation times of 1. The costs functions for production and
transport consist of fixed costs and variable costs. If all binary decisions would be con-
sidered in a MIP-model, this would lead to more than 3,800 binary variables, which
is beyond the size of problems we could solve with the best MIP/LP-solvers within
several hours. In comparison, our SimLP algorithm takes about 12 min for one test
instance to converge to a solution. We evaluate three different values for the quantile
used for the estimation of the delay parameters: 90, 70, and 50%. See Table 4 for the
results.
In this case it seems that the 90%-quantile is most robust choice, although the 70%-
quantile delivers only slightly worse results on average and in some cases even better
ones, whereas the 50%-quantile leads to the worst results for all instances. Due to the
fact that there is less stochastic near to the customer, it is possible to reduce the safety
factors for the delays to some extent without increasing the costs too much, because
lost time at the beginning of the supply chain can be made up at the end.
We also conduct experiments where the transportation delays at the end of the sup-
ply chain are stochastic, i.e. the connection between production nodes and customers
are uniformly distributed between 1 and 5. Transportation times between the two lay-
ers of production nodes are again uniformly distributed between 1 and 3. Remaining
transportation times are set to 1. The corresponding results are summarized in Table 5.
Table 5 Total costs of SimLP

Instance Quantile
for test example with ten supply
chain actors and stochasticity 90% 70% 50%
concentrated near the end of the
supply chain L1-L-C 263957 11.70% 43.96%
L2-L-C 263252 16.81% 45.02%
L3-L-C 263948 13.79% 48.81%
L4-L-C 263114 12.65% 50.60%
L5-L-C 263707 13.98% 51.41%
L6-H-C 258440 18.11% 52.09%
L7-H-C 257963 14.60% 59.29%
The results for the 90%-quantile L8-H-C 260606 12.67% 41.68%
are taken as basic values. For the
L9-H-C 258762 24.74% 52.44%
remaining quantiles the
percentage difference to the L10-H-C 258936 17.52% 43.16%
corresponding basic value is L1a-L-C 335837 7.15% 32.56%
given. The C indicates that there
L6a-H-C 335973 9.61% 32.95%
is more stochastic at the
customer Total Avg. 273708 14.45% 46.16%
For these instances the best choice would be to use the highest safety factor (90%),
because if there are delays at the end of the supply chain, there is no chance to catch
up.
For the test instances L1a-H-C and L6a-H-C with high fixed cost, we also apply a
SimMIP method where we include only the high fixed-charge production costs. The
low fixed-charge transportation costs are still neglected. For all quantiles the Sim-
MIP method delivers slightly lower costs (L1a-L-C: −4.48%/−10.86%/−1.19% for
90%/70%/50%-quantiles; L6a-H-C: −3.47%/−2.96%/−0.37% for 90%/70%/50%-
quantiles) but the calculation times are more than five times longer.
If we assume stochastic transportation times for the whole supply chain (all trans-
portation delays are uniformly distributed between 1 and 5), the results are similar to
those in Table 5, i.e. the 90%-quantile is always the best choice.
6 Conclusions
In this paper we have presented a new approach that combines the advantages of com-
plex simulation models and abstract optimization models. We have shown that our
method is able to generate competitive solutions quickly, even compared with tradi-
tional planning approaches that are much more time consuming. Our investigations
can be summarized as follows:
• In many cases the SimLP method seems to be a good trade-off between solu-
tion quality and computational time. If the nonlinear elements in the model are
dominating it is better to apply the SimMIP approach and consider these nonlin-
earities in the optimization model as along as the computational time for solving
the optimization model is acceptable.
• Furthermore, we investigated the impact of safety times for delays on the solution
quality. If we use the 90%-quantile, we can generate robust plans, but for specific
situations we might get better results with less safety time. Only for the case if
stochastic is near the customer, then the 90%-quantile is clearly the best. Never-
theless, the choice of the quantile depends on the structure of the supply chain and
has to be fine-tuned in each case.
• Using the 50%-quantile, i.e. the expected values for the delays, always leads to
pure results. If the uncertainty is concentrated far away from the customer, the
cost increase by using the expectation value is about 10% whereas the increase is
almost 50% if the uncertainty occurs close to the customer.
Further research for different aspects of this method is still possible and neces-
sary. The aggregation step and the generation of new decision rules is an open field.
One possibility is to interpret the solution of the optimization model only as a tar-
get strategy and use adaptive decision rules to approximate this target strategy in an
uncertain environment. The use of sensitivity results of the optimization model might
lead to improved decision rules. Further investigations are possible for the bound-
aries between the simulation and the optimization model. The question, which aspects
should be included in the optimization model, is not completely answered yet. If more
complex models are used, other fast solution methods (e.g., heuristics, metaheuristics,
etc.) should be taken into consideration.
We conclude by answering the question posed in the title of this paper: simulation
and optimization are complementary approaches and it is worthwhile combining them.
Acknowledgment We wish to thank Martin Grunow and two anonymous referees for their valuable
comments on this manuscript.
Appendix A: MIP formulation for fixed-charge transportation cost
The objective (20) of the optimization model can be transformed into a mixed-integer
program considering for example
fixed transportation costs. In case, the transportation
v p v p
cost functions Ci j xi j (t) can be written as follows:
v
ci j if v xi j (t) > 0
p p
v p v p
Ci j xi j (t) = ∀i, p, t, v. (22)
0 otherwise
In order to capture this situation, it is necessary to introduce binary decision vari-

ables v i j (t) which indicate if there is positive transportation. So by adding the fol-
p
lowing constraints:
v p
≤ G · v i j (t)
p
xi j (t) ∀i, p, t, v (23)
it is possible to formulate the transportation costs as the linear functions

v
Ci j (t) = v ci j v i j (t)
p p p
∀i, p, t, v. (24)
The resulting mixed-integer linear program includes now |J I | × |P| × T × |V |

binary decision variables.
A similar approach can be used for modeling step functions like
⎧v p
if v xi j (t) ∈ (v X i j , v Yi j ]
p p p
⎨ ci j
v v p
xi j (t) = v di j if v xi j (t) ∈ (0, v X i j ]
p p p p
Ci j ∀i, p, t, v. (25)
⎩
0 otherwise
Here we need 2 different binary decision variables v i j (t) and v i j (t) to represent
p p
this situation. If we add two additional constraints
v p
≤ G · v i j (t)
p
xi j (t) ∀i, p, t, v, (26)
v p
· v i j (t) + v X i j (t)
p p
xi j (t) ≤G ∀i, p, t, v, (27)
the cost functions can be written as

v
Ci j (t) = v di j v i j (t) + v p
− v di j
p p p p v p
ci j i j (t) ∀i, p, t, v. (28)
The resulting mixed-integer linear program includes now 2 × |J I | × |P| × T × |V |

binary decision variables.
Appendix B: Notation
Notation used for the optimization model

J set of locations J = JS ∪ J I ∪ JC
j ∈ JS raw-material supplier (starting nodes)
j ∈ JC customer (end nodes)
j ∈ JI nodes between supplier and customer
P set of products
V set of transportation modes
T number of periods
Decision variables
p
m i (t) amount of product p (product p is the end product of the produc-
tion process at location i) that starts to be produced at location i
in period t
p
u i (t) amount of product p that starts to be transacted in location i in
time period t
v x p (t) flow of product p from location i to location j with transportation
ij
mode v (sent away in period t)
Costs, delays, and general parameters

p
ai factor indicating the amount of capacity units required to produce
one unit of product p at location i
αi ( p ) amount of product p required to produce one unit of product p at
p
location i
in b p (t) amount of backorders of product p at customer i in period t
i
v C p (·) transportation cost function of product p transported from location
ij
i to location j with transportation mode v
v C (t) maximum transportation capacity of transportation mode v on the
ij
way from location i to location j
prod C (t) maximum production capacity at location i in period t
i
ta C (t) maximum transaction capacity at location i in period t
i
invin Cap p (t) maximum amount of product p that can be held in the inbound
i
inventory of intermediate i in period t
invout Cap p (t) maximum amount of product p that can be held in the outbound
i
inventory of intermediate i in period t
v Cap p (t) amount of product p that transportation mode v can transport from
ij
location i to location j in period t
prod Cap p (t) amount of product p that can be produced at location i in period t
i
ta Cap p (t) amount of product p that can be transacted at location i in period t
i
vc p cost factor used in case of linear transportation costs for deliveries
ij
of product p between location i and location j with transportation
mode v
p
Di (t) demand for product p at location i in period t
p
di factor indicating the amount of capacity units required to transact
one unit of product p at location i
p
δi amount of periods required to produce product p at location i
in f p (t) amount of product p arriving at location j in period t
j
out f p (t) amount of product p sent away at location j in period t
j
vgp factor indicating the amount of capacity units required to transport
one unit of product p with transportation mode v
in H p (·) inbound inventory cost function for product p at location i
i
out H p (·) outbound inventory cost function for product p at location i
i
in h p cost factor used in case of linear inventory costs for the inbound
i
inventory of actor i and for product p
out h p cost factor used in case of linear inventory costs for the outbound
i
inventory of actor i and for product p
in L (t) maximum capacity of inbound inventory at location i in period t
i
out L (t) maximum capacity of outbound inventory at location i in period t
i
in l p (t) inbound inventory level of product p at location i in period t
i
out l p (t) outbound inventory level of product p at location i in period t
i
p
qi factor indicating the amount of capacity units required to hold one
unit of product p at the inventory of location i
p
Ri (·) penalty cost function at location i for product p
p
ri (t) amount of product p which is already transported at period 0 and will arrive
at location i in period t (or external increase of inventory)
p
ρi cost factor used in case of linear penalty costs at customer c and for product
p
p
S j (t) supply of product p at location j in period t
p
si (t) amount of product p which is already in production process in period 0 and
will be finished in period t (or external increase of inventory)
p
σi amount of periods required to transact product p at location i
vτ amount of periods transportation mode v requires to go from location i to
ij
location j
p
Wi (·) production cost function of product p at location i
p
wi cost factor used in case of linear production costs at site i and for product
p
p
Z i (·) transaction cost function of product p at location i
p
zi cost factor used in case of linear transaction costs at site i and for product
p
References
Almeder C, Preusser M (2004) Mixed analytical / DEVS approach to ARGESIM comparison C14 “Supply
Chain Management” using Xpress-MP and Anylogic. SNE Simulat News Eur 41/42:50
Alonso-Ayuso A, Escudero LF, Garín A, Ortuño MT, Pérez G (2003) An approach for strategic supply
chain planning under uncertainty based on stochastic 0–1 programming. J Global Optim 26:97–124
Arbib C, Marinelli F (2005) Integrating process optimization and inventory planning in cutting stock with
skiving option: an optimization model and its application. Eur J Oper Res 163:617–630
Biswas S, Narahari Y (2004) Object oriented modeling and decision support for supply chains. Eur J Oper
Res 153:704–726
Dogan K, Goetschalckx M (1999) A primal decomposition method for the integrated design of multi-period
production-distribution systems. IIE Trans 31:1027–1036
Erengüc SS, Simpson NC, Vakharia AJ (1999) Integrated production/distribution planning in supply chains:
an invited review. Eur J Oper Res 115:219–236
Fleischmann B (1993) Designing distribution systems with transport economies of scale. Eur J Oper Res
70:31–42
Fu MC (2002) Optimization for simulation: theory vs practice. INFORMS J Comput 14:192–215
Glover F, Kelly JP, Laguna M (1999) New advances for wedding optimization and simulation. In:
Farrington PA, Nembhrad HB, Sturrock DT, Evans GW (eds) Proceedings of the 1999 Winter Simu-
lation Conference, pp. 255–260
Gronalt M, Hartl RF, Preusser M (2007) Design eines globalen Liefernetzes für Papier – Eine Fallstudie.
In: Corsten H, Missbauer H (eds) Productions- und Logistikmanagement. Vahlen Verlag, pp 597–616
Gunnarsson H, Rönnqvist M, Carlsson D (2007) Integrated production and distribution planning for Södra
Cell AB. J Math Modell Algorith 6:25–45
Kelton WD, Sadowski RP, Sadowski DA (2002) Simulation with Arena, 2nd edn. McGraw-Hill, New York
Kleijnen JPC (2005) Supply chain simulation tools and techniques: a survey. Int J Simulat Process Modell
1:82–89
Kuhn A, Rabe M (1998) Simulation in Production und Logistik (Fallbeispielsammlung). Springer, Heidel-
berg
Lee YH, Kim SH (2002) Production-distribution in supply chain considering capacity constraints. Comput
Indust Eng 43:169–190
Leung SCH, Tsang SOS, Ng WL, Wu Y (2007) A robust optimization model for multi-site production
planning problem in an uncertain environment. Eur J Oper Res 181:224–238
Meyr H (2002) Simultaneous lotsizing and scheduling on parallel machines. Eur J Oper Res 139:277–292
Mitrovic D (2006) CPLEX, XPRESS-MP, GLPK comparison by means of supply chain network models.
Diploma thesis, Vienna University of Technology
Pankaj C, Fisher ML (1994) Coordination of production and distribution planning. Eur J Oper Res 72:503–
517
Paraschis IN (1989) Optimale Gestaltung von Mehrproduktsystemen. Physica-Schriften zur Betriebswirts-
chaft, vol 26. Physica-Verlag, Heidelberg
Preusser M, Almeder C, Hartl RF, Klug M (2005a) LP Modelling and simulation of supply chain networks.
In: Günther HO, Mattfeld DC, Suhl L (eds) Supply Chain Managament und Logistik: Optimierung
Simulation, Decision Support. Physica-Verlag, Heidelberg, pp 95–114
Preusser M, Almeder C, Hartl RF, Klug M (2005b) Hybrid Supply Chain Modelling—Combining
LP-models and discrete-event simulation. In: Maroto Álvarez MC, Alcaraz Soria J, Ruiz GR, Crespo
Abril F, Vallada R (eds) ORP3—Operational research peripatetic postgraduate programme proceed-
ings, ESMAP, S.L., Universidad Politécnica de Valencia, pp 163–170
Santoso T, Ahmed S, Goetschalckx M, Shapiro A (2005) A stochastic programming approach for supply
chain network design under uncertainty. Eur J Oper Res 167:96–115
Schneeweiss C (2003) Distributed decision making, 2nd edn. Springer, Heidelberg
Shapiro JF (1993) Mathematical programming models and methods for production planning and sched-
uling. In: Graves SC, Rinnooy Kan AHG, Zipkin PH (eds) Handbooks in operations research and
management science: logistics of production and inventory. North–Holland, Amsterdam
Stadtler H (2005) Supply chain management and advanced planning—basics, overview and challenges. Eur
J Oper Res 163:575–588
Stäblein T, Baumgärtel H, Wilke J (2007) The Supply Net Simulator SNS: An artificial intelligence
apporach for highly efficient supply network simulation. In: Günther HO, Mattfeld DC, Suhl L (eds)
Management logistischer Netzwerke. Physica-Verlag, Heidelberg, pp 85–110
Swisher JR, Jacobson SH, Hyden PD, Schruben LW (2000) A survey of simulation and optimization tech-
niques and procedures. In: Joines JA, Barton RR, Kang K, Fishwick PA, (eds) Proceedings of the 2000
Winter Simulation Conference, pp 119–128
Truong TH, Azadivar F (2003) Simulation based optimization for supply chain configuration design. In:
Chick S, Sánchez PJ, Ferrin D, Morrice DJ (eds) Proceedings of the 2003 Winter Simulation Confer-
ence, pp 1268–1275
Yaged B (1971) Minimum cost routing for static network models. Networks 1:139–172
Revenue management and demand fulfillment:
matching applications, models, and software
Rainer Quante · Herbert Meyr ·

Moritz Fleischmann

OR Spectrum (2009) 31:31–62
DOI 10.1007/s00291-008-0125-8
Abstract Recent years have seen great revenue management successes, notably
in the airline, hotel, and car rental businesses. Currently, an increasing number of
industries, including manufacturers and retailers, are exploring ways to adopt simi-
lar concepts. Software companies are taking an active role in promoting the broa-
dening range of applications. Additionally technological advances, including smart
shelves and radio frequency identification (RFID), are removing many of the barriers
to extended revenue management. The rapid developments in supply chain planning
and revenue management software solutions, scientific models, and industry appli-
cations have created a complex picture, which is not yet well understood. It is not
evident which scientific models fit which industry applications and which aspects
are still missing. The relation between available software solutions and applications
as well as scientific models appears equally unclear. The goal of this paper is to
help overcome this confusion. To this end, we structure and review three dimensions,
namely applications, models, and software. Subsequently, we relate these dimensions
to each other and highlight commonalities and discrepancies. This comparison also
provides a basis for identifying future research needs.
R. Quante (B)
Institute for Production Management, Vienna University of Economics and Business Administration,
Nordbergstraße 15, 1090 Vienna, Austria
H. Meyr
Chair of Production and Supply Chain Management, Technical University of Darmstadt,
Hochschulstr. 1, 64289 Darmstadt, Germany
M. Fleischmann
RSM Erasmus University, P.O. Box 1738, 3000 DR Rotterdam, The Netherlands

58 R. Quante et al.
Keywords Revenue management · Demand fulfillment · Manufacturing · Software ·

Advanced planning systems
1 Introduction
Recent years have seen great revenue management successes, notably in the airline,
hotel, and car rental businesses. These successes essentially rely on identifying and
exploiting differences in the customers’ willingness to pay. Some approaches exploit
these differences by offering multiple product variants, tailored to different customer
segments, such as different fare classes in the airline industry. Other approaches stick to
a single product variant but adjust its price dynamically over time. This is practiced, for
example, by many budget airlines and fashion retailers during end-of-season clearance
sales. Currently, an increasing number of industries, including manufacturers and
retailers, are exploring ways to adopt similar concepts. Software companies are taking
an active role in promoting the transfer of revenue management concepts to a broader
range of applications. Technological advances, such as smart retail shelves and RFID
tags for real-time inventory visibility further support this development by decreasing
many potential barriers. Finally, increasing customer service and revenues through
intelligent demand fulfillment provides a way for companies to respond to the ever
increasing pressure of global competition.
The growing interest in revenue management applications is also reflected in intensi-
fied scientific research, as documented by a rapidly increasing number of publications
(see e.g. Fleischmann et al. 2004). Supply chain planning and advanced planning
software is also gradually incorporating ideas of revenue management. For example,
advanced planning systems (APS) extend the traditional available-to-promise and
capable-to-promise-logic of demand fulfillment modules to a profitable-to-promise
logic (SAP 2003). This development coincides with an ongoing consolidation in the
APS market, from a multitude of small vendors, like Red Pepper or Numetrix, to a few
big business application and business intelligence software companies, like Oracle or
SAP. Nevertheless, many niche players remain successful, due to their “greater ability
to manage the complexities of the supply chain, superior calculation power, greater
agility, and improved integration capabilities achieved through open standards and
service-oriented architectures” (ARC Advisory Group 2006). Due to acquisitions, big
software companies offer a wide range of supply chain management modules (several
dozens to more than a hundred), instead of just a few supply chain planning software
packages. Consequently, these companies need to re-arrange and re-structure their
supply chain management (SCM) software portfolios, as illustrated by the example of
Oracle, that needs to position its own advanced planning solutions together with the
supply chain planning, pricing, and demand management modules obtained through
the acquisitions of Peoplesoft (including the former Red Pepper and Numetrix soft-
ware), Retek, ProfitLogic, and Demantra (Oracle 2007).
Given these rapid developments, it is not surprising that not only software com-
panies, their customers, but also the scientific community is struggling to maintain a
clear picture of the resulting situation: which software modules serve which planning
purposes in which business applications? Which scientific models that have proven
RM and DF: matching applications, models, and software 59
successful in a certain type of industry are transferable or adaptable to which other

types of business with similar characteristics? The goal of this paper is to help over-
come these confusions. To this end, we structure and review three dimensions, namely
applications, models, and software. Subsequently, we relate these dimensions to each
other highlighting commonalities and discrepancies. This comparison also provides a
basis for identifying future research needs.
We do not pretend to cover every potential application, nor every modeling detail.
Instead, we focus on a few prototypical cases (“types”) on each of the three aforemen-
tioned dimensions. Moreover, within each dimension we build on existing reviews and
classifications whenever possible.
The scope of our analysis includes short- and mid-term demand fulfillment
decisions in a supply chain. For this delineation we follow Fleischmann and Meyr
(2004) who define the planning tasks of demand fulfillment relative to the position of
the decoupling point, which divides the supply chain into forecast-driven and order-
driven processes (Sharman 1984; Hoekstra and Romme 1992). Demand fulfillment, as
understood in this article, comprises the decisions at and downstream of the decou-
pling point. These decisions are based on customer orders and primarily deal with
managing the due dates of these orders.
We complement demand fulfillment with the concept of revenue management.
According to Talluri and van Ryzin (2004), revenue management concerns demand
management decisions aimed at increasing a firm’s revenues. The authors distinguish
quantity-based and price-based revenue management approaches. The first approach
relies on exploiting customer heterogeneity. It segments customers into multiple classes
and prioritizes them when allocating scarce capacity. The key idea is that giving prio-
rity to high-margin segments yields higher revenues than selling scarce capacity on a
first-come-first-served basis. The second revenue-management approach uses pricing
decisions as a lever for demand management. This includes adjusting prices dynami-
cally over time in response to non-stationary demand or a finite selling season or via
auctions as a price-discovery mechanism.
Traditionally, the demand fulfillment and APS perspective is common in manu-
facturing, whereas revenue management applications are mainly found in the service
industries. In this paper we argue that the planning tasks of both concepts are actually
very similar, and we systematically compare them. By highlighting analogies and dis-
tinctions, we aim to provide a basis for expanding the traditional domains of application
for both concepts.
To summarize, our contribution is threefold. First, we unite the currently distinct
concepts of demand fulfillment and revenue management and compare them to each
other. Second, we link the three dimensions of applications, models, and software
discussing similarities and differences between them. We note that Elmaghraby and
Keskinocak (2003) address similar aspects for dynamic pricing in a retail environment.
Our paper differs from their’s by a broader scope in terms of planning tasks and appli-
cations. At the same time, we consider aggregated model types in the literature, rather
than reviewing individual modeling contributions. Third, we provide a supply chain
framework for revenue management and demand fulfillment. We believe that many
other applications beyond the examples illustrated in this paper are worth exploring.
Our presented framework provides a means for doing so in a systematic way.
60 R. Quante et al.
The remainder of the paper is structured as follows. Section 2 introduces our

framework, which we use for structuring the three dimensions of applications,
models, and software. Sections 3–5 analyze each of these dimensions separately.
Section 6 then links the three dimensions and identifies alignments and discrepancies
between them. Section 7 summarizes our main insights and discusses opportunities
for future research.
2 A supply chain framework for revenue management and demand fulfillment
In this section, we present a framework, which we use in the remainder of the paper for a
structured analysis of revenue management and demand fulfillment (RM&DF) aspects
in each of the three dimensions of applications, models, and software. The framework
is based on the elements of the supply chain depicted in Fig. 1. It is motivated by the
aforementioned definition of demand fulfillment by Fleischmann and Meyr (2004).
Recall that demand fulfillment encompasses supply chain decisions downstream of the
decoupling point (DP). At the same time, we can also represent revenue management
decisions in this framework. We link pricing decisions to the item “final product” and
capacity allocation decisions to the item “demand”. We explain these elements in more
detail below.
Our framework includes supply chain elements that are directly related to revenue
management and demand fulfillment decisions. Specifically, we consider the following
elements, from upstream to downstream: Replenishment represents either an external
supplier or internal production. According to the definition of demand fulfillment, the
replenishment policy at the DP of the receiving party is based on demand forecasts.
This is a key point in our subsequent analysis. The decoupling point itself is the
next element of our supply chain framework. It holds the inventory that is needed to
hedge against forecast errors and replenishment uncertainty. The DP plays a pivotal
role in our analysis since many RM&DF decisions are dependent on the available
inventory and on future replenishment orders. For a more detailed discussion of the
DP concept and its impact in different production environments [including make-
to-order (MTO), assemble-to-order (ATO) and make-to-stock (MTS)] we refer to
Fleischmann and Meyr (2004). The supply chain may contain additional production
processes downstream of the DP. For MTS production this is not the case, since
the DP holds the final product. In contrast, limited downstream capacity, possibly of
multiple production stages, plays a critical role in an MTO supply chain. By definition,
production downstream of the DP is order driven. Therefore, it has to be considered
g n)
l in tio ity t
oup t uc c l uc
c in od pa na od
De Po ( Pr Ca Fi Pr
__
Replenishment Demand
Forecast-driven ( Demand Planning) Order-driven ( Demand Fulfillment)
Fig. 1 Supply chain framework

in the analysis to capture the effects of production lead times or order fulfillment.
The elements furthest downstream in our framework pertain to the final product and
corresponding customer demand.
RM&DF are closely tied to decisions in the depicted supply chain and therefore
depend on its specific characteristics. For example, the current inventory at the DP or
the remaining production capacity may influence pricing decisions or promised due
dates. By capturing these characteristics, the framework provides a systematic basis for
identifying RM&DF requirements in different applications. Moreover, by structuring
models and software tools in a similar way, we can compare the three dimensions to
each other.
In order to characterize different types of supply chains we describe each element
of the framework by a number of attributes, which are relevant to RM&DF. Each
supply chain then corresponds with a specific value for each attribute. For example,
the product life cycle is an attribute of the final product, which can take the values short
or long. We do not seek to describe every potential attribute of any given supply chain
element but rather to focus on those attributes that are the most relevant to RM&DF.
Using these attributes allows us to reduce the complexity of the intended compa-
rison. Specifically, we group instances together by types that have the same or simi-
lar values in many attributes. In this way, we identify applications that have similar
RM&DF requirements. In addition, we can compare application, model, and software
types to each other to see which tools are available for supporting RM&DF decisions
in a given context.
We proceed as follows. Sections 3–5 address applications, models, and software,
respectively. In each case, we first introduce corresponding supply chain attributes
and briefly discuss their potential values and their relevance to RM&DF. Then we
characterize a set of instances in terms of these attributes and cluster them into types.
Section 6 finally compares the types across the three dimensions.
3 Industry applications
In this section we analyze and compare RM&DF decisions in different industries.

As explained in Sect. 1, we do not seek completeness, but rather explore a set of
examples covering a broad range of different applications. To this end, we include
examples from the service industries, retail, and manufacturing. Specifically, within
the service industries, we look at the airline industry as a classical user of revenue
management. Given the diverse airline demand management strategies, we make a
distinction between premium and budget airlines. The retail sector is a driving force
in many recent pricing strategies. We include examples of fashion retail and consumer
packaged goods in our analysis. Finally, we include three manufacturing examples
that reflect different DPs, namely MTS, ATO, and MTO.
3.1 Application-oriented supply chain attributes
In order to analyze and compare RM&DF decisions in these different environments

we characterize them in terms of our supply chain framework. To this end, we consider
62 R. Quante et al.
in
g n)
pl ctio ity
u t u c ct
co in od pa al odu
De Po ( Pr Ca n
Fi Pr
__
Refilling of Decoupling Flexibility of Perishability Profit

inventory point capacity Perishable heterogeneity
Buyer-driven MTO High Durable πtk =
refilling ATO Low Life cycle rtk – ctk + sk
Vendor-driven MTS Short
refilling Long π: Profit
Pricing r: Revenues
flexibility c: Costs
s: Strategic
At order time component
Short term
Mid term t: Time index
k: Customer class
None index
Fig. 2 Supply chain framework: application dimension
the attributes depicted in Fig. 2, which we explain from upstream to downstream in

what follows. Again, we do not claim completeness of our selection but rather focus
on arguing why the selected attributes are relevant to RM&DF. Table 1 summarizes
the evaluation of the attributes for the aforementioned industry applications.
3.1.1 Replenishment-related attributes
At the upstream end of the framework, we distinguish different roles of the supply
chain members with respect to the forecast-based replenishment of inventory located at
the DP. We capture these differences in the attribute refilling of inventory. The famous
“rationing game” describes a powerful supplier who distributes scarce supply among
his customers proportional to their ordered quantities. This mechanism encourages
customers to inflate orders in order to obtain the quantity they truly desired. We denote
situations in which suppliers have the power to decide on production and deliveries
by the attribute value vendor-driven refilling. Rationing only occurs in the case of
scarce supply capacity. If supply exceeds demand, orders will, in general, be fulfilled
as requested. We describe this situation as buyer-driven refilling.
If the supplier has the power to decide on deliveries this results in unreliable deli-
very times and limited replenishment capacity from the buyer’s perspective. A more
powerful buyer implies more reliable deliveries and a perception of unlimited supply
capacity. In conclusion, the attribute affects the buyer’s supply flexibility and thereby
his fulfillment decisions.
To illustrate the application of this attribute, consider the automotive industry,
fashion retail, and airlines. In automotive production, suppliers are usually in a weak
position due to the extreme competition in the market—hence refilling is buyer dri-
ven. A similar situation occurs in the case of fashion retail, although this may differ
between cheap and luxury fashion. For airlines, the concept of DP refilling does not
Table 1 Application examples
Attribute Application
Manufacturing: Manufacturing: Manufacturing: Service: Retail:

consumer goods configurable machines (budget/premium) (fashion/consumer
computers airlines packaged goods)
Refilling of inventory Vendor-driven Partly vendor-driven Buyer-driven

Decoupling point MTS ATO MTO MTO MTS
RM and DF: matching applications, models, and software
Flexibility of capacity High High Low

Perishability Durable Durable Durable Perishable Durable
Life cycle Long Short Long Long Short/Long
Pricing flexibility Mid-term At order time At order time Short-term/Mid-term Short-term/Mid-term
Profit heterogeneity πtk = rt − ctk + sk πtk = rtk − ck + sk πtk = rtk − ck + sk πt = rt − c/πk = rk − c πt = rt − c
63
64 R. Quante et al.
really apply since seats are “sold” only temporarily (i.e., for a given flight) and no
orders are placed to get new seats.
3.1.2 Decoupling-point-related attributes
As discussed in Sect. 2, the decoupling point represents a bundle of characteristics that

separate forecast-driven from order-driven processes. Another common term is order
penetration point. See Sharman (1984); Hoekstra and Romme (1992); Fleischmann
and Meyr (2004) and Meyr (2003) for a detailed analysis of the DP concept. We
distinguish three attribute values.
In MTO production systems the DP lies upstream in the supply chain. Production
is triggered by incoming customer orders. The inventory at the DP consists of raw
materials, which are used in downstream production processes. The bottleneck in
MTO is usually the production capacity downstream of the DP. MTO production is
appropriate primarily for high-value and customer-specific products.
In ATO systems, production downstream of the DP includes the final product assem-
bly. Inventory at the DP essentially consists of components that are usually delivered
by external suppliers. Bottlenecks may occur at either the DP inventory or the downs-
tream assembly process.
MTS production is based on forecasts. Final products are produced on stock. This
is a common strategy for standard products without customer-specific requirements.
The location of the DP has a strong impact on RM&DF decisions by influencing
customer service times, order fulfillment, and control of replenishment orders. For
example, deciding on a (promised) delivery date becomes more important as the DP
moves upstream in the supply chain. At the same time, customer service times increase.
Furthermore, the different production environments require different types of forecasts
with different levels of aggregation.
The literature provides several real-life examples of RM&DF in supply chains with
different DPs. For example, Spengler et al. (2007) show an MTO application in the iron
and steel industry. Harris and Pinder (1995) mention custom-made textile and custom
equipment manufacturing as ATO examples. Meyr (2008) shows an application of
MTS in the lighting industry. Given the scope of our paper, it is worth pointing out
how the DP concept can be applied not only to manufacturing industries, but also
to retail and service industries. In retail, inputs are the final products, while outputs
are the sold products. The “production” process involves bundling and offering the
right products at the right time, corresponding with MTS “production”. In the case
of a service business, “production” inputs include working time, material, etc., and
the output coincides with the transformation process. Since the “production” step is
dependent on the customer’s presence it can be characterized as MTO.
3.1.3 Capacity-related attributes
Similar to DP inventory replenishment, the degree of flexibility of downstream

capacity also impacts a firm’s fulfillment decisions by providing an alternative lever
for matching supply and demand. Possibilities to adapt capacity levels to short-term
demand fluctuations include, e.g., shutting down unnecessary machines or production
lines in the case of excess capacity and hiring temporary workers or extending regular
working hours in the case of shortages. General strategies for balancing capacities and
demand can be found in, e.g., Chase and Aquilano (1995) or Bertrand (2003). These
options may be further restricted by technical or regulatory requirements.
Evaluating this attribute for our application examples, we observe that airline
capacity tends to be very inflexible in the short term, due to long planning lead times
for flight and crew schedules. In manufacturing, flexibility depends strongly on the
technology used and on working-time agreements with labor unions. In the process
industry, machines cannot easily be switched off due to the long and costly running-up
phase. In contrast, computer manufacturing is highly flexible due to a high number of
manual tasks. In retail, capacity flexibility concerns changing opening hours and thus
working times.
3.1.4 Product-related attributes
We observe several characteristics of the final product that impact RM&DF. The first
one is the degree of perishability, i.e., the maximum storage time. The values of this
attribute range from perishable to durable. Perishability describes the flexibility regar-
ding the time horizon for selling the product. An inherent characteristic of perishable
products is their low salvage value after expiration.
Product perishability has important consequences for RM&DF. For perishable pro-
ducts, sales in the current period usually do not influence future sales. This is intuitive
for necessity items that are frequently consumed. Moreover, since necessity items are
consumed repeatedly, there is hardly a possibility for the customer to wait until the
price decreases. Perishability also increases the importance of optimal replenishment.
Since overstocking is expensive for perishable products, proper demand forecasts and
corresponding optimal order quantities are crucial. In the case of durable goods, to-
day’s purchases may affect future sales. For example, a customer typically buys a
computer only once within a few years. Therefore, price discounts early in the product
life cycle may hurt future sales (Elmaghraby and Keskinocak 2003).
Another relevant product-related attribute concerns the length of the product life
cycle, i.e., the duration of the selling season. It can range from a few months, as in the
case of fashion goods, to multiple years, as for basic food items.
This attribute also has a strong impact on RM&DF. Short life cycles limit the
availability of historical demand data and thereby complicate reliable forecasting.
Moreover, due to the high frequency of new product releases, the customer will learn
to anticipate the shape of the future price path. For example, video games usually have
a short selling season and new ones are released frequently. Therefore, prices of video
games are rapidly decreasing since they out-date quickly. Anticipating future prices,
customers behave strategically and buy at their individual best price. In the case of
long life cycles, anticipating future prices is more difficult for the consumer, due to a
lack of experience. We refer to Elmaghraby and Keskinocak (2003) for a more detailed
discussion of this attribute.
A third product-related attribute worth considering is the degree of pricing flexi-
bility, i.e., a company’s possibilities to change prices. A customized product usually
also has a customized price, as opposed to a standard product that sells for the same
66 R. Quante et al.
price during a longer time horizon. The attribute values range from changing at order
time, as in the case of highly customized products, over short-term, up to mid-term
price changes.
An online store may serve as an example of almost costless short-term price changes.
Mid-term price changes are found in industries where the physical or organizational
infrastructure for frequent price changes do not exist; for example, retail prices publi-
shed in catalogs. Additionally, not only the frequency of price changes but also their
potential magnitude characterizes the degree of pricing flexibility.
The impact of this attribute on RM&DF decisions is large since, for example, price-
based demand management relies crucially on pricing flexibility. If this flexibility is
not given, quantity-based demand management by reserving products for specific
customer segments becomes more important.
3.1.5 Demand-related attributes
On the downstream end of our supply chain framework, we distinguish multiple ways
in which profitability may differ between orders. We denote this attribute as profit
heterogeneity. The formula in Fig. 2 displays three factors for differentiating customer
order profitability, namely revenues r , costs c, and strategic importance s. Each of these
factors may vary over time (denoted by the index t) and between customers (denoted
by the index k).
For example, airlines often charge different prices according to the remaining boo-
king time and other factors like remaining capacity, which implies different revenues
rt at different points in time. Different fare classes, such as business and economy,
represent an example of different revenues rk from different customer segments k at
the same point in time. In retail, different prices rt are charged at different points in
time t according to the remaining selling season however they typically apply to all
customers.
Similarly, the costs of serving a customer order may also be a differentiator. Note
that only those costs which can still be influenced when accepting the order are relevant
here. This includes, for example, transportation costs, taxes, and any variable costs
of downstream production. Again, these costs can be invariant (c), differ between
customer classes (ck ) and/or differ between different points in time (ctk /ct ).
Finally, customers may differ in their strategic importance, which may go beyond
immediate costs and revenues. We capture this in our framework through the parameter
s. For example, loyal customers may be extremely important and should be treated
better than occasional customers.
As discussed in Sect. 1, heterogeneity lies at the very heart of RM. Price-based
approaches are driven by the time dependence of the above profit elements. Quantity-
based approaches seek to exploit customer segmentation. It is worth noting that
segmentation may be attractive even in the case of constant unit revenues if the
overall profit π is heterogeneous due to varying costs (e.g. πtk = r − ctk ) and/or
different strategic importance (e.g. πk = r + sk ). This contradicts the common
claim that sunk costs of unused capacity are a major prerequisite for applying RM
(Weatherford and Bodily 1992). Meyr (2008) and Fischer (2001) describe applications
of quantity-based approaches with strategically important customers.
Supply Flexibility
Capacity-based (MTO) Inventory-based (MTS + ATO)
Inflexible Flexible Inflexible Flexible
Inflexible
Pricing Flexibility
Service: Manufacturing: Retail:
Premium Consumer Consumer
Airlines Goods Pack. Goods
Manufacturing:
Flexible
Service:
Manufacturing: Configurable
Budget
Machines Computers
Airlines
Retail: Fashion
Fig. 3 Different application types
3.2 Application types
Table 1 summarizes the various attribute values for the selected applications briefly
introduced at the beginning of this section. Missing attribute values indicate that an
attribute is not applicable to the specific case. For example, the capacity flexibility
attribute is not applicable to make-to-stock production since there is no production
process downstream of the DP in this case. Capacity may be crucial upstream of the
DP (e.g., in consumer goods manufacturing where scarce production capacities often
lead to push-based, “vendor-driven” refilling of finished goods inventories) but the
production process can no longer be influenced upon order arrival and therefore does
not make part of RM&DF, as defined in Sect. 1. Similarly, refilling of inventory is not
applicable to the MTO cases. These examples illustrate that the different attributes in
Table 1 are not mutually independent. We can therefore further simplify the charac-
terization of the selected applications. Figure 3 highlights that the applications differ
from each other with respect to the criteria supply flexibility and pricing flexibility.
The first criterion combines the first three attributes of Table 1, whereas the second
one corresponds directly with the original attribute pricing flexibility. The supply is
considered as inflexible if there is either a partly vendor-driven refilling of inventory
or low flexibility of capacity. Note also that these criteria correspond with the main
planning tasks among the list of attributes whereas the remaining attributes further
describe the context.
We exploit the classification displayed in Fig. 3 for matching the application with
available models and software in Sect. 6. In particular, we discuss the implications of
the different degrees of flexibility on RM&DF requirements. Before doing so however,
we first describe the model and the software dimensions in Sects. 4 and 5, respectively.
4 Models and methods
4.1 Model-oriented supply chain attributes
This section addresses the model dimension of RM&DF. Analogous with the previous
section, we first discuss a number of model-oriented attributes of our underlying supply
68 R. Quante et al.
g n)
l in ctio ity
up t u c ct
o
c in od pa al odu
De Po ( Pr Ca n
Fi Pr
__
Replenishment Level of Capacity at Price ptk Customer order

quantity yt inventory It None None acceptance xtk
None None Data Data None
Data Data Decision Decision variable* Decision variable
Decision Decision variable Perishability Demand reliability
variable variable Perishable / Durable Deterministic
Replenishment Stochastic
*Discrete set, continuous
lead time
Deterministic
Stochastic
Fig. 4 Supply chain framework: Model dimension
chain framework and then derive and review a set of distinct model types. Figure 4 lists
a number of attributes that we believe are relevant for characterizing RM&DF models,
as we discuss below. Again, we do not claim the list of attributes to be exhaustive.
We derive our model attributes from the basic elements of mathematical optimiza-
tion models, namely decision variables, objective function, constraints, and input data.
We do not address the objective function separately since the decision variables include
sufficient information for our analysis. In addition, we jointly consider constraints and
data as given model inputs. Figure 4 displays potential decisions in straight bold cha-
racters and pure data attributes in italic bold characters.
A basic set of decision variables is derived from the product flow through the supply
chain. These pertain to the replenishment quantity at the DP yt , the resulting inventory
level It , and the number of customer orders accepted xtk . As in the previous section,
the index t denotes the time dependence and k indicates a distinction by customer
class. These quantities
are interrelated through the basic inventory balance equation
It = It−1 + yt − k xtk , stating that the inventory at the end of period t is equal
to the inventory at the end of the preceding period plus replenishment minus total
accepted demand from all customer classes. Each of the factors in this equation may
be modeled either as decision variables, as exogenous data, or it may not be considered
at all (none).
We make the same distinction for the sales price ptk and also consider whether its
domain is discrete or continuous. With respect to the production capacity downstream
of the DP at we distinguish between capacitated models (data), models with capacity
as a decision (e.g. overtime), and uncapacitated models, which do not consider scarce
capacity downstream of the DP (none).
In addition, we consider a few pure data attributes whose different values appear
to be relevant to RM&DF (see Fig. 4). These include the replenishment lead time,
which may be completely known (deterministic) or uncertain (stochastic), product
perishability, and demand reliability. Many other attributes will be valuable for refining
Replenishment consideration
None Data Decision variable
•Replen. yt: None •Replen. yt: Data •Replen. yt: Decision
variable
Demand / Price consideration

Data
Stochastic
•Price p: Data Order
•Customer order acc. xt: - Inventory
Promising
Decision var. Control
Price-based
•Price pt: Decision var. Markdown /
•Customer order acc. xt: Trade Integrated
Decision var. Pricing /
Promotions Pricing
Auctions
Quantity-based
•Price pk: Data
•Customer order acc. xtk:
Inventory
Traditional RM aATP
Decision var. Rationing
Fig. 5 Different model types
the analysis, such as salvage values, number of different products, backordering or lost
sales, or the number of sales channels. However, in order not to overload our model
we do not include all of these details in our analysis below.
4.2 Model types
Table 2 characterizes different streams of literature in terms of the attributes listed

in Fig 4. As for the different application types in Sect. 2, the different attributes in
Table 2 are also not independent. We therefore again condense the representation by
concentrating on the most distinctive attributes. It turns out that pricing and order
acceptance, on the one hand, and control over the replenishment quantity, on the other
hand, suffice to reasonably characterize most model types. Re-arranging the model
types along these two variables results in the compressed view shown in Fig. 5.
Models in the first row of Fig. 5 take demand and price as entirely exogenous. They
satisfy demand first-come-first-served (FCFS) at a given price. In particular, they do
not involve any customer segmentation. The next two rows of Fig. 5 correspond with a
more active demand management. Models in the second row still consider only a single
customer class, but allow price changes which provide a lever for influencing demand.
The last row shows models that explicitly recognize heterogeneous customers. In
responding to a customer request they have to make a trade-off between accepting a
current, low-priority order now versus reserving the resources for high-priority orders
expected in the future.
The demand side of Fig. 5 corresponds with the classification of RM-models of
Talluri and van Ryzin (2004). We follow their terminology and label models as price-
based or quantity-based. The columns reflect the way the different models handle
inventory replenishment at the decoupling point. Note that both dimensions of the
figure correspond with key decision variables regarding demand and supply. Also note
that these dimensions are closely related to those of the application types discussed in
the previous section. We analyze this analogy in detail in Sect. 6.
In the remainder of this section, we briefly discuss each of the above model types.
Given the scope of our analysis, we primarily build on available review papers in
70
Table 2 Model examples
Attribute Model type
Order Stoch. inventory Pricing Markdown Trade Integrated aATP Traditional RM Inventory Auctions
promising control management promotions pricing rationing
Replenishment quantity Data Decision None None Data Decision Data None Decision None
Replenishment lead time Det/Stoch Det. Det. Det. Det. Det/Stoch
Level of inventory Decisiona Decisiona None Data Data Decisiona Decisiona None Decisiona None
(Production) Capacity Data None None None None None Data Data None None
Price Data: p Data: p Dec: pt Dec: pt Dec: pt Dec: pt Data: pk Data: pk Data: pk Dec: ptk
Perishability Durable Durable Durable Durable Durable Durable Durable Perish Durable Dur/Per
Order acceptance Dec: xt Dec: xt Dec: xt Dec: xt Dec: xt Dec: xt Dec: xtk Dec: xtk Dec: xtk Dec: xtk
Demand reliability Det/Stoch Stoch Stoch Stoch Stoch Stoch Det Stoch Stoch Stoch
a derived decision
R. Quante et al.
the literature, rather than review the individual models. Again, we see a more refined
analysis as a valuable issue for future research.
4.3 Review of model types
4.3.1 Single-class exogenous demand models
We have not found any models fitting in the upper left cell of Fig. 5. This is not
surprising since models with a given price and no consideration of replenishment or
inventory, respectively, have nothing to decide on, neither on the demand nor on the
replenishment side of the supply chain.
In the next cell to the right, the so-called order promising models consider price
(or demand), current inventory, and future replenishment quantities as given. This
results in information about product availability and delivery times. For each incoming
customer order the model decides real-time on the due date. The decision is made in
a greedy fashion, purely based on availability. An introduction and overview of this
so-called “real-time mode” or “single-order-processing” models is given in Ball et al.
(2004), Chen et al. (2001), and Fleischmann and Meyr (2004). Additionally, a broad
overview of due date management models with an emphasis on stochastic models is
included in the work of Keskinocak and Tayur (2004).
The upper right cell of Fig. 5 holds the large class of stochastic inventory control
(SIC) models, which focus on optimal inventory replenishment. Some of these models
primarily address the structure of optimal replenishment policies, as for example the
famous (s, S)-policy proven by Scarf (1960) (if inventory position is below s, order up
to S). Other models seek to determine optimal control parameters of such policies, such
as the optimal ordering time, order quantity, and inventory review intervals. Many SIC
models build on the classical newsvendor model, which seeks to determine the optimal
order quantity for a perishable product under stochastic demand. An overview of
single-period newsvendor problems is given by Khouja (1999). Silver (1981) provides
an overview and typology of many standard inventory problems, such as the ones
mentioned above. General up-to-date overviews of inventory models can be found in
the textbooks by Silver et al. (1998), Porteus (2002) and Tempelmeier (2006).
4.3.2 Price-based models
The model types in the middle row of Fig. 5 treat price as a decision, which influences
the demand. Pure pricing models aim to determine an optimal selling price, without
considering replenishments. For example, given a price–demand relation, the goal is
to find the price which maximizes total revenues. Mild et al. (2006) review factors
influencing demand and show how to find optimal prices.
Markdown models determine the right price path for inventory clearance for a
given amount of inventory, which cannot be replenished during the planning horizon.
Elmaghraby and Keskinocak (2003) classify several dynamic pricing models with and
without replenishment decisions, the latter ones including markdown models.
72 R. Quante et al.
Auctions, as discussed for example by Talluri and van Ryzin (2004, Sect. 6), take a
fundamentally different approach to pricing. They provide a price-discovery mecha-
nism and thereby an alternative to posting fixed prices. This approach is particularly
valuable if little demand information is available. The aforementioned authors discuss
the close connection between auctions and dynamic pricing.
Trade promotion models represent a type of pricing model that considers reple-
nishments as an exogenous input. These models therefore fit in the second column of
Fig. 5. Neslin (2002) provides an overview and discusses the reasons for promotions.
Research in integrated pricing (IP) models dates back to Whitin (1955) who extends
the EOQ-formula (economic order quantity) as well as the classical newsvendor model
with price decisions. This field has seen extensive research in the last decades, which
is summarized, for example, by Petruzzi and Dada (1999). Recent research focuses on
multiple period models, which are discussed in the well-known literature reviews of
Chan et al. (2004), Elmaghraby and Keskinocak (2003) and Yano and Gilbert (2003).
Few models exist for environments in which replenishment, prices, and due dates are
set simultaneously. Some models of this type and other models dealing with setting
due dates can be found, for example, in the previously mentioned review paper by
Keskinocak and Tayur (2004).
From an application-oriented perspective it is worthwhile comparing IP and a
successive application of pricing and SIC models. While IP models recognize the
interdependence between pricing and replenishment and therefore determine decisions
simultaneously, they do so at the cost of a more simplified demand and supply repre-
sentation. Pure pricing models may include sophisticated demand functions, including
reference price effects, promotion effects, and competition (Mild et al. 2006). Simi-
larly, SIC models consider factors such as multiple suppliers and quantity-discounts.
IP models typically cannot deal with these factors due to tractability (Elmaghraby and
Keskinocak 2003, Sect. 4).
4.3.3 Quantity-based models
Models in the bottom row of Fig. 5 take prices as exogenous but manage demand by
means of rationing strategies. In contrast with the models of the top row, the models
distinguish multiple customer classes and prioritize them rather than fulfilling orders
in an FCFS manner.
The type traditional revenue management (TRM) in the first cell of the third row
refers to models that are common in airline applications. In these models, a given
number of units of a perishable product (e.g., seats on a flight on a specific day) are
allocated to customers with different priorities or different willingness to pay. The
basic question is whether to accept a given order or to reserve capacity in anticipation
of more profitable future orders. McGill and van Ryzin (1999) and Pak and Piersma
(2002) provide an overview and a short history of research in traditional revenue
management with a focus on airline applications. Boyd and Bilegan (2003) discuss
models focusing on e-commerce applications. The recent review by Chiang et al.
(2007) includes an overview of RM practices in different industries.
Allocated available-to-promise models (aATP) are similar to the order promising
type of the top row except for differentiating between multiple customer classes. Scarce
resources (inventory on hand, planned stock at the DP or capacity downstream of the

DP) are allocated to these classes according to customer profitability or other priority
measures. Within each class, customer requests are usually handled FCFS, just as in
traditional order promising. Guerrero and Kern (1988) introduce the general problem
of accepting and refusing orders and discuss the requirements and implications of
order promising mechanisms. For reviews of the mostly deterministic models of this
type we refer to Kilger and Meyr (2008), Pibernik (2005) and Meyr (2008).
If customer requests do not have to be answered instantaneously several customer
orders can be collected and jointly promised in a batch, thereby creating higher degrees
of freedom for selecting the most important or profitable orders within a simultaneous
optimization process. Overviews of these so-called “batch order promising models”
can again be found in the work of Ball et al. (2004), Chen et al. (2001) or Fleischmann
and Meyr (2004).
A review of integrated due-date management and job-scheduling models with
deterministic orders is provided by Gordon et al. (2002). Their article considers batch
models in which due dates are determined according to current capacity and the desi-
red delivery date. Keskinocak and Tayur (2004) give a general overview of due-date
setting models.
aATP and TRM models are similar in that they decide about demand fulfillment
with respect to different customer classes. The most significant difference between the
two models is the perishability of resources. TRM considers “perishable” products,
e.g., empty seats on a specific flight, which are lost after the departure date, whereas the
ATP quantities managed in aATP models are storable, in general. Another difference
concerns the time horizon. TRM models typically consider a fixed day of capacity
availability, e.g., the departure date of a flight. In contrast, aATP models consider
multiple periods linked through the storability of excess inventory. Furthermore, aATP
models usually assume deterministic demand whereas demand in TRM models is
stochastic.
The last model type within our framework concerns inventory rationing (IR) models.
Similar to the relationship between aATP and traditional order promising, IR models
extend traditional SIC models by distinguishing and prioritizing multiple customer
classes. For an early review we refer to Kleijn and Dekker (1998). Like traditional
SIC models, IR models may consider deterministic or stochastic replenishment lead
times. A further distinction within this class of models is the number of demand classes
considered, which may be general or limited to two classes.
For lack of recent reviews, we refer to a few individual articles that reflect two broad
research streams within the type of IR models. Ha (1997) and De Véricourt et al. (2002)
propose models with multiple demand classes and stochastic replenishment times,
thus assuming limited production (= replenishment) capacities. In contrast, Melchiors
et al. (2000) and Arslan et al. (2005) model deterministic replenishment lead times
and unbounded replenishment quantities. All these models take decisions on ordering
and rationing levels, which are typically expressed in policies like (s, S, R) where s is
the reorder point, S the order-up-to level, and R the protection level between customer
classes.
IR and aATP models differ in terms of exogenous versus endogenous replenishment.
Specifically, IR models consider replenishment decisions with stationary deterministic
74 R. Quante et al.
or stochastic lead times. In contrast, aATP models typically focus on capacitated,

dynamic and deterministic arrivals of push-based production (=replenishment) quan-
tities. To this end, aATP usually assumes deterministic and dynamic demand forecasts
whereas IR models assume stochastic demand.
In addition to the model types captured in Fig. 5, a few recent research streams
combine several types by simultaneously considering multiple attributes. For example,
Pibernik and Yadav (2008) present a model uniting the distinct features of the types
aATP (deterministic production quantities) and IR (stochastic demand). Kocabıyıkoğlu
and Popescu (2005) jointly analyze price and allocation decisions with two customer
classes. Since most quantity-based models assume exogenous prices, this seems to
be a promising direction for future research. Bitran and Caldentey (2003) formulate
a general model of this problem and review the current state of research. Another
approach is pursued in Ding et al. (2006) in which trade promotion models are com-
bined with inventory rationing models. The authors denote the resulting new pro-
blem type by ADP, referring to the allocation of available stock, discounting and
prioritization of customers.
5 Software applications
5.1 Software types
The software market for demand and supply chain solutions has changed in the recent
years. For many years the focus was on the supply side. The interest is now, however,
turning to end-to-end solutions including the demand side. Big supply chain solution
providers like Oracle and SAP are investing large amounts into the acquisition of
demand-related know-how. For example, in 2005 SAP took over Khimetrics, a leading
vendor of markdown, price, and promotion-optimization solutions. As noted in Sect. 1,
Oracle—after taking over one of its largest competitors in supply chain solutions,
Peoplesoft, in 2005—simultaneously invested in the demand solutions of Demantra
(2006), ProfitLogic (2005) and Retek (2005), all of them leading vendors of retail
revenue management software. Another big consolidation occurred in 2006 when
JDA Software—a provider of specialized retail solutions—took over Manugistics, a
supply chain solution provider focusing on profit optimization in the consumer goods
industry.
The scope of our current analysis is restricted to software supporting short-term
decision making in RM&DF. These solutions draw data from other software systems,
such as customer relationship management systems on the demand side (Buttle 2004)
and enterprise resource planning systems (Stadtler and Kilger 2008) on the supply
side. Since these systems themselves do not focus on decision making we do not
include them in our analysis.
As discussed in the previous section, scientific optimization models are fairly well
described in the literature. One can easily identify data, decision variables, restrictions,
and solution strategies. Moreover, the solution quality is often analyzed in detailed
numerical studies. This is different for commercial software solutions. Usually, avai-
lable information is scarce and reveals little of the underlying technology. Software
Replenishment consideration
Demand / Price consideration

Purchasing &
Traditional Materials
Data -
Order Promising Requirements
Planning
Markdown- /
Pricing- / Promotion Enterprise Profit
Price-based
Auction- Optimization Optimization
Management
Demand
Revenue Master
Quantity-based Fulfillment &
Management Planning
ATP
Fig. 6 Different software types
users can only assess the supported input data, available options, and the resulting
output that is automatically calculated. The solution quality can hardly be evaluated
objectively and is usually judged by user experience.
Our analysis of software modules reflects this limited availability of objective infor-
mation. We build our characterization of software types and functionalities primarily
on available software reviews and whitepapers. As a starting point we use essentially
the same attributes and values as for the scientific models (see Fig. 4). Model data and
decisions roughly correspond with software input and output, respectively. Due to this
correspondence, we omit a detailed discussion of attributes in this section.
Figure 6 structures software types along the same axes as the model types in Sect. 4.
We choose names according to the functionality of commercial software modules on
the market. The remainder of this section briefly reviews each of these software types.
5.2 Review of software types
5.2.1 Single-class exogenous demand solutions
The mid-upper cell of Fig. 6 denoted by traditional order promising contains traditional
software modules for short-term order promising under known inventory availability.
When a customer order arrives, the software simply determines whether the order can
be satisfied out of available inventory. If not, the order is backlogged according to a
standard lead time without considering future capacity or additional incoming supply.
It is easy to see that this approach can lead to an order peak after the standard lead time
and thus to severe capacity problems in the future. Kilger and Meyr (2008) illustrate
this situation in a simple example.
Refilling of inventory is usually left to purchasing and materials requirements
planning modules, which are part of enterprise resource planning (ERP) systems.
Essentially, these systems support refilling of non-bottleneck material and compo-
nents from a single vendor. An overview of these classical systems can be found,
for example, in the textbook of Vollmann et al. (2005). Since these classical systems
provide sufficient solution quality only for very simple settings, specialized inventory
76 R. Quante et al.
modules consider extensions such as capacitated replenishment, stochastic demand,

and multiple suppliers (Stadtler 2008). Such modules usually are part of larger
advanced planning and supply chain planning software suites. Additionally, there
are specialized vendors of supply-chain wide inventory optimization tools, such as
Optiant (2007) with its inventory suite Powerchain and Smartops Smartops (2007).
5.2.2 Price-based solutions
Markdown management systems are mainly used in retail, for example for end-
of-season stock clearance. An example of markdown management systems is B_Line,
described by Mantrala and Rao (2001) under the name MARK. The system takes
possible prices and corresponding demand probability distributions for each period as
inputs and can find both markdown and markup price paths. The output consists of
a specific price in each period. Furthermore, MARK is capable of finding a suitable
amount of initial inventory by iterating through a discrete set of possible inventory
levels. Elmaghraby and Keskinocak (2003, Sect. 3.2) describe the capability of mark-
down solutions.
Software systems of the type pricing management are relatively new. This is due
to improvements in computing power and increased availability of past sales data.
The rise of data warehouses and cheap computing power has recently allowed the use
of automated pricing systems for many applications. Pricing management systems
are based on complex price–demand functions for which suitable parameters have
to be estimated, a process requiring vast amounts of past sales data. For example,
to estimate price elasticity, the sales data must include a certain degree of diversity,
corresponding with at least a few past price changes. Capacity or inventory restrictions
are usually not considered in these types of software (see for example Mild et al.
2006).
The quick expansion of e-commerce applications has boosted the use of auction
systems. The large number of different systems merits a review in its own right and
exceeds the scope of our analysis. We refer to Kambil and van Heck (2002) for a
systematic introduction to this field. Vakali et al. (2001) discuss the characteristics of
internet-based auction systems and present a short survey of popular applications.
Similar to the previously described markdown systems, promotion optimization is
also used in retail environments, as described by Elmaghraby and Keskinocak (2003,
Sect. 3.2). Very detailed information about the capability of such systems can be found
online, for example from the vendors mentioned at the beginning of this section.
The term enterprise profit optimization (EPO) was coined by the software company
Manugistics, who claims to be the first vendor offering an integrated pricing and supply
solution (Manugistics 2002). Furthermore, Manugistics software is meant to be able
to allocate scarce resources to the most profitable customers, thus simultaneously
applying ideas of quantity-based RM&DF. Demand and supply planning is realized in
many solutions, but not in an integrated way and not including price decisions. Most
APS forecast demand for different price levels and then successively analyze—within
the context of mid-term planning—several what–if scenarios and their effects on the
total supply chain.
5.2.3 Quantity-based solutions
APS software modules that support mid-term, aggregated supply and demand decisions
are known as master planning modules (Meyr et al. 2008b). They receive determi-
nistic demand forecasts and prices as inputs from the demand planning module of
APS (Kilger and Wagner 2008) and then determine the best combinations of sales,
production and replenishment quantities and the corresponding inventories under
given capacity constraints. Quantities can be allocated to different customer classes.
In terms of our supply chain framework in Fig. 1, master planning modules deal
with forecast-driven demand planning (i.e. push-based, vendor-driven refilling of the
DP) and therefore fall outside the scope of our definition of RM&DF. However,
we feel that they deserve mention since their resulting allocations serve as the pri-
mary input for the short-term, capacity-checked order promising, executed by the
Demand Fulfillment and ATP modules of APS. A detailed list of options conside-
red in master planning modules can be found in the work of Rohde and Wagner
(2008).
By taking capacity and inventory replenishments into account, demand fulfillment
and ATP modules of APS extend the aforementioned traditional order promising. They
determine due dates for incoming customer orders, which promise to be more reliable
than simple standard lead times. In addition, if ATP quantities are allocated to custo-
mer priority classes—in the usually implemented aggregated way—order promising
differentiates with respect to customer importance, based on customer profitability or
strategic impact.
To find a reliable due date for a customer order, the software searches for demand
fulfillment alternatives according to pre-defined “search dimensions”. These include
the time dimension, i.e., checking for ATP back- or forwards in time, the product
dimension, i.e., substitute products, the location dimension, and the customer dimen-
sion, i.e., checking for availability in other priority classes (Kilger and Meyr 2008).
Usually, the software systems do not take the profitability of different fulfillment
alternatives into account during this search. However, recent systems not only consider
available-to-promise quantities (available inventory) or capable-to-promise quantities
(available capacity), but also follow a profitable-to-promise (PTP) logic that enables
them to compare customer orders and fulfillment alternatives according to their prio-
rity. Usually, simple rules are defined as search strategies for the different dimensions
(Meyr et al. 2008a Sect. 18.3.1).
Revenue management software is widely used by airlines, hotel chains, and car
rental agencies. RM software systems basically take the given capacity and offe-
red tariffs as input and decide on acceptance or rejection of customer orders. One
of the main differences with demand fulfillment and ATP is that RM software
focuses on revenues rather than costs. Furthermore, RM systems usually forecast
demand in much more detail than demand-fulfillment modules, e.g., for each flight,
on each day, and for each customer class. These forecasts require a large amount
of historical sales data in order to be reliable. Modern revenue management sys-
tems can handle many additional industry-specific issues, such as overbooking and
connecting flights in the airline context (Talluri and van Ryzin 2004 Sects. 10.1.3,
11.2).
78 R. Quante et al.
6 Matching applications, models, and software
In this section, we match the three previously discussed dimensions applications,

models, and software against each other in order to identify alignments and misali-
gnments. In particular, our goal is to identify the most appropriate models and soft-
ware types for RM&DF decisions in the different applications introduced in Sect. 3,
namely service, retail, and manufacturing. We also seek to highlight remaining
research needs.
We build our discussion around the structure of Figs. 3, 5, and 6. Specifically,
we compare the supply flexibility observed in a given application with the way that
replenishment decisions are supported by different model and software types. Simi-
larly, we tie the observed demand flexibility to the supported demand management. In
this way, we identify for each application the most appropriate cell(s) in Figs. 5 and 6.
Subsequently, we discuss the match/mismatch with models and software within that
cell in more detail by including the additional attributes highlighted in the preceding
sections. This allows us to recognize empty spots and future research needs.
Figure 7 summarizes the match between the different types of applications and the
main model and software attributes. In the remainder of this section we explain this
match by application type.
6.1 Service industries
Service industries can be characterized as MTO since the “production” step requires
the presence of the customer. Hence, replenishment decisions do not play a role and
one primarily needs to consider available capacity downstream of the decoupling point
(currently available seats in an airplane). This suggests that the first column in both
Figs. 5 and 6 is the most relevant for service industries.
Moreover, since capacity is inflexible and cannot be substituted nor stored, the
means for matching supply to demand are limited, which makes demand management
the main lever in the short term. Models and methods in the top row of our figure
hardly support this step and therefore appear insufficient for these applications. A more
intelligent way of demand management is required, which recognizes the customers’
Replenishment
Data - (Manufacturing) (Retail)

Demand Management
Retail / Retail /
Price-based (Online) Retail
Service ATO-Manuf.
Quantity-based Service /
Manufacturing Online Retail
(MTO-Manuf.)
Fig. 7 RM&DF in different applications

willingness to pay. This can be achieved either through price-based (second row)
or quantity-based (third row) RM&DF, which renders models and software of the
type markup/-down, pricing, and (traditional) revenue management the most intuitive
candidates (see Figs. 5, 6 and 7).
Which of these types are best suited essentially depends on the pricing flexibility
and possibility of customer segmentation. This corresponds with the two examples
of service applications sketched in Sect. 3, namely budget and premium airlines.
Because of their online pricing flexibility, budget airlines can use markup models
and software to determine the (increasing) mid-term price path of a certain flight,
as well as short-term temporary discounts. Premium airlines benefit from tailoring
fares to specific customer segments, which leads to the quantity-based approaches of
traditional revenue management.
Of course, these matches are hardly surprising since they confirm well-known
common practice. However, the examples show that our approach is able to identify
these matches correctly. This supports the application of the same approach to less
obvious examples as in the following cases.
6.2 Retail
Inventory replenishment in retail environments is primarily buyer-driven (see Table 1)

and thus endogenous. This suggests that the third column of Figs. 5 and 6 is the most
appropriate in this context. There are, however, exceptions. For example, when reple-
nishment decisions are made on the mid-term and cannot be revised on the short-term
(e.g. in the fashion industry due to long lead times and short life cycles) replenishment
has to be taken for granted at the time of order fulfillment or it may be non-existent
altogether. Similarly, replenishment may be of little concern in the case of ample
capacity and short lead times. Thus, the first two columns may also be relevant to
retail.
Under the general assumption of buyer-driven inventory replenishment and the MTS
decoupling point, supply, and production can be considered as relatively unproblematic
in retail (see Sect. 3). On the other hand, retail is close to the final consumer and
therefore has a traditional focus on price setting, even if prices can only be changed on
the mid-term. Therefore, the second row of our scheme is of primary interest. Because
of the relatively simple supply and production requirements, all of the price setting
models in this row appear relevant under certain circumstances. This includes pricing
models (e.g., if inventory cannot or needs not be considered), markdown models (for
short life cycles), trade promotion models (for long life cycles with given replenishment
contracts), and integrated pricing models (with simultaneous replenishment decisions).
Similarly, the corresponding software types, namely pricing, markdown manage-
ment, promotion optimization, and EPO software can be applicable, with the appro-
priate choice depending on the detailed application attributes in Table 1. All these
models and software can be applied on the mid-term, or on the short term if pricing
flexibility is high enough.
Opportunities for customer segmentation (third row) are generally more limited
in a retail environment. It is hard to collect sufficient data for identifying natural
80 R. Quante et al.
consumer segments or to define and fence off “artificial” consumer segments analogous
to the service industries. Fencing usually requires additional services (e.g., business
customers receiving bonus miles, preferred transfer, etc.) which may not be obvious in
retail—essentially requiring a shift from a purely product-oriented retailer to a more
general service provider. Other examples of segmentation in retail include the selling
of a product as part of a bundled package, or the use of multiple distribution channels
(Agatz et al. 2007). Inventory rationing models can then be applied for refilling and
allocating stock to different channels.
Online retailers face particularly promising opportunities for gathering data for cus-
tomer segmentation as well as for changing prices almost instantaneously. Therefore,
they can integrate price- and quantity-based RM with flexible replenishment strate-
gies. Models (and software) for this type of application are rare and offer promising
research opportunities (see also Sect. 4).
6.3 Manufacturing
Our third main application type is the manufacturing industry. Unlike in the preceding
application types, production processes are the most important, and usually most
costly, process steps. The decoupling point as the interface between forecast-driven
demand planning and customer-oriented demand fulfillment (see Fig. 1 and Sect. 3)
describes whether a certain production process is operated under demand (un)certainty,
what type of stocks (raw material, components, final products) must be held, where
the main bottlenecks (stocks, production capacity) can be expected, and the length of
customer service times. Because of these significant differences we consider MTS,
ATO, and MTO manufacturing applications separately.
6.3.1 Make-to-stock
In MTS environments, all production processes are executed based on forecasts. Due
to upstream capacity limitations, production planning decides on short-term reple-
nishment of DP inventories in a push-based, “vendor-driven” manner. Thus, models
including replenishment decisions (third column of Fig. 5) can only support mid-term,
forecast-based demand planning, but not short-term demand fulfillment. In order to
make use of the (uncertain) information on future DP inventory replenishments, as
implied by the production plans, demand fulfillment models of the second column
appear the most appropriate.
Because of the MTS market conditions and contracting practice, pricing deci-
sions typically have to be taken on a mid-term basis (see Table 1). For example,
the demand planning module of an APS forecasts several price-demand scenarios
including, e.g., different alternatives for price discounts or promotions. These scena-
rios are passed to the master planning module, which checks each of them with respect
to supply chain constraints, selects the most profitable one, and generates directives for
the (forecast-based) short-term production planning. Thus, short-term pricing flexibi-
lity is rather limited, which rules out the models and software in the second row of
Fig. 5 for demand fulfillment in MTS manufacturing. Price-based approaches appear
mainly applicable on the mid-term planning level, e.g., to determine demand forecasts
in conjunction with optimal prices.
Thus order promising and aATP models remain as the most applicable models
for RM&DF in MTS manufacturing. Both models consider the current level of DP
inventory. Order promising in a MTS environment searches through the ATP quantities
in an FCFS manner to be able to fulfill a customer order. Newer approaches process
several customer orders in a batch and allocate ATP to the most profitable customer
orders. Due date setting is not relevant in this environment. aATP models overcome the
disadvantage of batch order promising, namely not providing a real-time order promise
and forcing the customer to wait. These models are, however, dependent on forecast-
based information on DP inventories as provided by the master and production plans,
and on the possibility of customer segmentation. The ATP search rules that are used to
consume the allocated ATP quantities of the different customer classes follow similar
ideas as traditional RM methods. However, since products in MTS manufacturing are
durable and can be stored (see Table 1), they also have to be able to “search over time”,
i.e., to take future inventory replenishments into account.
6.3.2 Assemble-to-order
In ATO production, the most important supply chain elements are the inventory kept at
the DP (components to be assembled) and the scarce assembly capacity downstream
of it. Since customer order service times are longer than in MTS environments, order
promising becomes more difficult and more important. In addition, new customer-
oriented planning tasks arise, such as supply and demand matching and the short-term
production planning of the assembly process (see Fleischmann and Meyr 2004).
To determine the suitable columns in Figs. 5 and 6, one should consider whether or
not replenishments can be controlled in the short term. In the aforementioned example
of configurable computer manufacturing, long replenishment lead times essentially
rule out this option. If availability of the computer components, e.g., high-end CPUs,
is limited and they cannot be replenished as required (i.e., vendor-driven), an integrated
demand fulfillment and replenishment is impossible, which rules out the third column.
On the other hand, DP inventory is as important in ATO as in MTS and thus information
about future replenishments should not be neglected. Otherwise promised due dates
will not be reliable. Therefore, again, models of the first column (especially traditional
RM methods) are not appropriate or at least have to be adapted to these needs.
In general, the capacity downstream of the DP involves some degree of flexibility.
Therefore, demand management is less important for matching supply and demand
than in entirely inflexible environments (see service applications above). However,
ATO usually concerns customized products, and therefore yields opportunities for
customer segmentation and exploiting differences in willingness to pay. Therefore,
the focus of ATO RM&DF should lie on selling the available resources to the different
customer segments in the most profitable way. aATP models and software (third row
of Figs. 5 and 6) outperform the simpler FCFS order promising (first row) in that
respect.
For real-time order promising, the advantages of aATP over an FCFS logic are
similar to the ones in the MTS case, except that multi-stage bills-of-material have to be
82 R. Quante et al.
taken into account. Batch order (re-)promising models, however, are more important in
the ATO case because they can also be applied for supply-demand-matching (which is
part of the short-term production planning), i.e., to select the most important customer
orders to release next to the shop floor for assembly. Standard APS software can be
used for this purpose (see Kilger 2008), but usually does not provide as sophisticated
methods as the corresponding optimization models (Ball et al. 2004).
For customized products, pricing can also be part of order promising because prices
often are only communicated upon customer request, due to the high number of poten-
tial product configurations. However, traditional trade promotion models and software
as listed in the central cell of Figs. 5 and 6 are not appropriate for this planning task.
On the other hand, traditional pricing or EPO models/software usually do not take into
account future inventory availability at the DP, as determined by the mid-term pro-
duction planning. Therefore, we see room for future research on price-based RM&DF
with given replenishments.
Even though their business model has suffered recent criticism, Dell has shown
that direct selling can build a bridge between traditional manufacturing and traditional
retail and that pricing flexibility—at least for non-standard components such as addi-
tional memory—can be increased through direct access to customers (Kraemer et al.
2000). In this case, both pricing flexibility and customer segmentation can be exploi-
ted simultaneously. Therefore, as noted for the case of online retailing in the previous
subsection, integrated price- and quantity-based approaches offer further research
opportunities, also for ATO manufacturing.
6.3.3 Make-to-order.
In MTO production all—usually multi-stage and very complex—production processes

are executed to order and production capacity is critical. In contrast, DP inventory only
consists of basic materials that can be easily replenished. Therefore, replenishment
decisions are of lesser importance and we can concentrate on the models and software
of the first and second column of Figs. 5 and 6. Although short-term production
planning in MTO is also customer-oriented and thus part of demand fulfillment, we
mainly focus on the order promising aspects of MTO in what follows. The key task is
to quote reliable due dates and to decide which orders to accept in order to maximize
profit.
Pricing flexibility in an MTO setting is high. However, automatic price-based
approaches are not generally applicable, because prices for the usually complex and
expensive MTO products must be negotiated. Standard pricing software (second row)
cannot really be used for determining minimum acceptable prices since sufficient past
sales data are hardly available for the highly customized products in MTO. A more
appropriate way of finding minimum acceptable prices may be using the production
planning software to assess marginal costs, e.g., by simulating production plans with
and without the given order. Due to the complexity of the production system this can
often only be done in an aggregate, approximate manner.
Similar to ATO, the high customization of MTO products offers opportunities for
segmentation strategies, provided that the customer base is sufficiently heterogeneous
to identify and separate different segments. Customers of MTO products are naturally
segmented by the differences in production costs between different orders. Therefore,

fencing may not be needed.
We further narrow the choice of appropriate cells of Figs. 5 and 6 by noting that the
underlying models and software should include available production capacity. This
leaves us with the model types of order promising, aATP, and traditional revenue
management. Traditional RM assumes fixed production costs, non-flexible capacity
and perishability. These assumptions apply to only a few, exceptional MTO industries
(e.g., perishable products in some chemical industries). In the majority of cases, tra-
ditional RM must at a minimum be modified to be able to support due-date setting in
MTO production.
Traditional order promising models for MTO focus on due date setting under limited
production capacities. They either check capable-to-promise (CTP) quantities in an
FCFS manner (Dickersbach 2004; Kilger and Meyr 2008), estimate due dates by means
of stochastic queuing theory (Keskinocak and Tayur 2004), or simulate the short-term
production planning deterministically, as discussed above. Whereas the first and the
last types of models are usually implemented in standard APS, the second type is only
found in specialized software.
Differences in the profitability of customer orders are not exploited by traditional
order promising models. Some APS allow a profitability assessment of different ful-
fillment alternatives by evaluating their revenues and costs during the rule-based (ATP
and) CTP check. The aATP logic can be extended similarly for additional CTP checks.
This again allows prioritizing different customer segments, thereby transferring RM
ideas to MTO manufacturing. However, since CTP quantities represent production
capacities only in a very aggregate manner, the resulting due dates and thus also the
estimated profits are not very reliable for complex production systems. In this case
a more detailed simulation of short-term, customer-oriented production planning, as
described above, appears necessary. This, however, often yields prohibitively long
computation times, which implies long reaction times to customer requests.
Of course, short-term production planning models and software can also be applied
to batch promising of multiple customer orders, which offers additional degrees of
freedom. However, the approach suffers from the same inherent problems as in the
single-order case. We conclude that the biggest challenge in MTO order promising is
not so much the availability of models and software but how to implement them in
practice to yield a good balance between solution quality/reliability and short response
times.
7 Conclusion
In this paper, we have analyzed and structured revenue management (RM) and demand
fulfillment (DF) decisions. We have presented a framework that covers the underlying
supply chain processes and allows for a systematic comparison of different business
environments.
The conceptual integration of revenue management and demand fulfillment is a
major contribution of this paper. Both concepts have emerged in different industries.
Demand fulfillment is a standard component of advanced planning systems that are
84 R. Quante et al.
mainly applied in manufacturing industries. Revenue management, on the other hand,

is tightly linked to service industries, notably to airline ticket sales. These different
backgrounds and associated terminology and connotations complicate a systematic
comparison and thus a mutual exchange of ideas between both concepts. In this paper,
we argue that revenue management and demand fulfillment actually concern essen-
tially the same supply chain processes.
The main conceptual distinction is the differing degree of demand management.
Traditional DF essentially treats demand as exogenous. It assumes given prices and
treats customer orders in a first-come-first-served manner. For each incoming customer
order it searches for the fastest fulfillment possibility and promises a corresponding
due date to the customer. In contrast, RM seeks to more actively manage demand. It
considers short-term price adjustments and prioritizes different customer segments,
thereby exploiting differences in the customers’ willingness to pay.
The framework introduced in Fig. 1 in Sect. 2 allows for a uniform treatment of
RM and DF decisions. In this paper, we have applied it to characterize and compare
a few exemplary application environments. The selected examples are far from com-
prehensive. Yet they serve to demonstrate the applicability of our framework, which
we hope will be useful for analyzing many other examples. Moreover, the framework
also supports a classification of RM and DF models and software systems. We have
identified two key distinctive factors, namely the aforementioned demand manage-
ment and the replenishment strategy at the supply chain decoupling point. Similar to
the former, the latter can be endogenous or exogenous. We have used both of these
factors (see Figs. 5, 6, 7) to structure reviews of optimization models and software
systems, which are relevant for RM&DF, and to compare them to the requirements of
selected applications in the service, retail, and manufacturing industries in Sect. 6.
This led to the following main insights (cp. also Fig. 7):
– Replenishment is not an issue in service industries. Scarce capacities (e.g., of

seats in an airplane, hotel rooms, etc.), which cannot be stored and thus also not
be replenished, are the center of attention. Therefore, fewer means are available
for balancing supply (=capacities) and demand than in other industries, which
motivated the early use of RM techniques. On the other hand, adapting these
traditional RM techniques for manufacturing or retail applications requires the
integration of replenishment and storability.
– Retail is closest to the end consumer. Production and limited production capacities
do not play a noteworthy role, but replenishment is important and can take on
different shapes. Various types of price-based demand management are applicable
here. Possibilities for differentiated stock allocation are still limited. Exceptions
include multi-channel retailing, where some channels may have priority over
others, or online retailing, where a lot of information about customer behavior
is available and the customers do not see the physical inventories.
– Manufacturing industries are heterogeneous. Inventory holding, replenishment,
and limited production capacities can occur separately or in any combination,
and the means for balancing supply, capacity, and demand are manifold. Thus, an
active demand management in the sense defined above was not as important in the
past as in service industries and is still in its infancy today. However, it arouses
more and more interest. The decoupling point concept helps structure the variety
of manufacturing applications. The major potential for a more active demand
management seems to lie in differentiated capacity allocation, i.e., quantity-based
RM&DF. A short-term pricing potential is mainly given for ATO manufacturing,
due to the high degree of customization in combination with a pressure for short
customer order lead times.
Unlike in service environments, order promising in manufacturing industries is a multi-

period problem, i.e., production earlier or deliveries later than the customer’s requested
date are possible. Therefore, traditional RM techniques typically cannot be applied as
is, except possibly in a few MTO manufacturing industries. Otherwise, they have to
be adapted to deal with the holding and (future) replenishment of decoupling point
inventory (MTS, ATO), costs (MTS, ATO, MTO) and throughput-time estimates for
downstream processes (ATO, MTO). The allocation and ATP consumption rules cur-
rently used in APS serve this purpose, but are very basic. We see a need for future
research here. Incorporating further ideas of traditional (quantity-based) RM might
lead to more sophisticated methods and software.
Additionally, we have identified further promising research opportunities: In online
retail the integration of price- and quantity-based demand management ideas with
flexible replenishment strategies seems to offer interesting potentials. The same is
true for direct sellers in ATO manufacturing who additionally have to take care of the
product assembly.
For due-date re-promising and demand-supply-matching, which are further plan-
ning tasks of ATO manufacturing, batch optimization models appear helpful and have
already been proposed in the scientific literature. However, to develop optimization
methods that are scalable to practical needs and can find their way into commercial
planning software, further research seems necessary.
Last but not least, short-term production planning and scheduling modules of APS
already offer the basic functionality to estimate the due dates and costs of a cer-
tain customer request in complex MTO environments. However, for most practical
applications the effort in terms of modeling complexity and computation time is still
too high. Thus we see an interesting trade-off between detailed but complex cost and
throughput time projection models as a basis for demand management and simpler
but possibly less accurate projections. Even if this problem can eventually be handled
the question will remain how to use these projections in price negotiations with the
customer and whether another customer is about to call who is willing to pay an even
higher price.
Acknowledgements The authors are grateful to the Vienna Science and Technology Fund (WWTF) for
funding this research.
References
Agatz N, Fleischmann M, van Nunen J (2007) E-fulfillment and multi-channel distribution—a review. Eur
J Oper Res (in press)
ARC Advisory Group (2006) Supply chain planning worldwide outlook. http://www.arcweb.com/
StudyPDFs/Study_scp.pdf
86 R. Quante et al.
Arslan H, Graves SC, Roemer T (2005) A single-product inventory model for multiple demand classes.
Working paper 4542-05, Massachusetts Institute of Technology (MIT), Sloan School of Management
Ball MO, Chen CY, Zhao ZY (2004) Available-to-promise. In: Simchi-Levi D, Wu SD, Shen ZJ (eds)
Handbook of quantitative supply chain analysis—Modeling in the e-business era. Kluwer, Dordrecht,
pp 447–483
Bertrand JWM (2003) Supply chain design: flexibility considerations. In: de Kok A, Graves S (eds) Supply
chain management: design, coordination and operation vol 11. Elsevier, pp 133–198
Bitran G, Caldentey R (2003) An overview of pricing models for revenue management. Manuf Service
Oper Manage 5(3):203–229
Boyd EA, Bilegan IC (2003) Revenue management and e-commerce. Manage Sci 49(10):1363–1386
Buttle F (2004) Customer relationship management, 1st edn. Butterworth Heinemann, Oxford
Chan LMA, Shen ZJM, Simchi-Levi D, Swann JL (2004) Coordination of pricing and inventory decisions:
a survey and classification. In: Simchi-Levi D, Wu SD, Shen ZJ (eds) Handbook of quantitative supply
chain analysis—modeling in the e-business era. Kluwer, Dordrecht, pp 335–392
Chase RB, Aquilano NJ (eds) (1995) Production and operations management: manufacturing and services,
7th edn. Irwin, Chicago
Chen CY, Zhao ZY, Ball MO (2001) Quantity and due date quoting available to promise. Infor Systems
Front 3(4):477–488
Chiang WC, Chen JCH, Xu X (2007) An overview of research on revenue management: current issues and
future research. Int J Rev Manage 1(1):97–128
De Véricourt F, Karaesmen F, Dallery Y (2002) Optimal stock allocation for a capacitated supply system.
Manage Sci 48(11):1486–1501
Dickersbach JT (ed) (2004) Supply chain management with APO. Springer, Heidelberg
Ding Q, Kouvelis P, Milner JM (2006) Dynamic pricing through discounts for optimizing multi-class ful-
fillment. Oper Res 54(1):169–183
Elmaghraby W, Keskinocak P (2003) Dynamic pricing in the presence of inventory considerations: research
overview, current practices, and future directions. Manage Sci 49(10):1287–1309
Fischer ME (2001) “Available-to-Promise”: Aufgaben und Verfahren im Rahmen des Supply Chain Mana-
gement. Theorie und Forschung Band 704, Wirtschaftswissenschaften Band 63. Roderer, Regensburg
Fleischmann B, Meyr H (2004) Customer orientation in advanced planning systems. In: Dyckhoff H,
Lackes R, Reese J (eds) Supply chain management and reverse logistics.. Springer, Berlin pp 297–321
Fleischmann M, Hall JM, Pyke DF (2004) Smart pricing. MIT Sloan Manage Rev 45(2):9–13
Gordon V, Proth JM, Chu C (2002) A survey of the state-of-the-art of common due date assignment and
scheduling research. Eur J Oper Res 139:1–25
Guerrero HH, Kern GM (1988) How to more effectively accept and refuse orders. Prod Invent Manage J
29(4):59–63
Ha AY (1997) Inventory rationing in a make-to-stock production system with several demand classes and
lost sales. Manage Sci 43(8):1093–1103
Harris FH, Pinder JP (1995) A revenue management approach to demand management and order booking
in assemble-to-order manufacturing. J Oper Manage 13(4):299–309
Hoekstra S, Romme JE (1992) Integral logistic structures:developing customer-oriented goods flow.
McGraw-Hill, London
Kambil A, van Heck E (2002) Making markets: How firms can design and profit from online auctions and
exchanges. Harvard Business School Press, Boston
Keskinocak P, Tayur S (2004) Due-date management policies. In: Simchi-Levi D, Wu SD, Shen ZJ (eds)
Handbook of quantitative supply chain analysis—Modeling in the e-business era.. Kluwer, Dordrecht,
pp 485–547
Khouja M (1999) The single-period (news-vendor) problem: Literature review and suggestions for future
research. Omega 27:537–553
Kilger C (2008) Computer assembly. In: Stadtler H, Kilger C (eds) Supply chain management and advanced
planning, 4th edn. Springer, Heidelberg, pp 381–398
Kilger C, Meyr H (2008) Demand fulfilment and ATP. In: Stadtler H, Kilger C (eds) Supply chain mana-
gement and advanced planning, 4th edn. Springer, Heidelberg, pp 181–198
Kilger C, Wagner M (2008) Demand planning. In: Stadtler H, Kilger C (eds) Supply chain management
and advanced planning, 4th edn. Springer, Heidelberg, pp 133–160
Kleijn MJ, Dekker R (1998) An overview of inventory systems with several demand classes. Econometric
institute report 9838/a, RSM Erasmus University
Kocabıyıkoğlu A, Popescu I (2005) Joint pricing and revenue management with general stochastic demand.
Working paper, INSEAD
Kraemer KL, Dedrick J, Yamashiro S (2000) Refining and extending the business model with information
technology: Dell computer corporation. Infor Soc 16(1):5–21
Mantrala MK, Rao S (2001) A decision-support system that helps retailers decide order quantities and
markdowns for fashion goods. Interfaces 31(3):146–165
Manugistics (2002) Enterprise profit optimization. http://www.manugistics.com/documents/
epo_whitepaper.pdf
McGill JI, van Ryzin G (1999) Revenue management: research overview and prospects. Transport Sci
33(2):233–256
Melchiors P, Dekker R, Kleijn M (2000) Inventory rationing in an (s, Q) inventory model with lost sales
and two demand classes. J Oper Res Soc 51(1):111–122
Meyr H (2003) Die Bedeutung von Entkopplungspunkten für die operative Planung von Supply Chains.
Zeitschrift für Betriebswirtschaft 9:1–22
Meyr H (2008) Customer segmentation, allocation planning and order promising in make-to-stock produc-
tion. OR Spectrum (in press)
Meyr H, Rosič H, Seipl C, Wagner M, Wetterauer U (2008a) Architecture of selected APS. In: Stadtler H,
Kilger C (eds) Supply chain management and advanced planning, 4th edn. Springer, Heidelberg, pp
349–366
Meyr H, Wagner M, Rohde J (2008b) Structure of advanced planning systems. In: Stadtler H, Kilger C
(eds) Supply chain management and advanced planning, 4th edn. Springer, Heidelberg, pp 109–116
Mild A, Natter M, Reutterer T, Taudes A, Wöckl J (2006) Retail revenue management. In: Schnedlitz P,
Buber R, Reutterer T, Schuh A, Teller C (eds) Innovationen in Marketing und Handel. Linde, Wien,
pp 124–143
Neslin SA (2002) Sales promotion. In: Weitz BA, Wensley R (eds) Handbook of marketing.. Sage Publi-
cations, London pp 310–338
Optiant (2007) http://www.optiant.com
Oracle (2007) Strategic acquisitions. http://www.oracle.com/corporate/acquisition.html
Pak K, Piersma N (2002) Airline revenue management: an overview of OR techniques 1982–2001. Econo-
metric Institute Report 256, Erasmus University, Rotterdam
Petruzzi NC, Dada M (1999) Pricing and the newsvendor problem: a review with extensions. Oper Res
47(2):183–194
Pibernik R (2005) Advanced available-to-promise: classification, selected methods and requirements for
operations and inventory management. Int J Prod Econ 93–94:239–252
Pibernik R, Yadav P (2008) Inventory reservation and real-time order promising in a make-to-stock system.
OR Spectrum (in press)
Porteus EL (2002) Foundations of stochastic inventory theory. Stanford University Press
Rohde J, Wagner M (2008) Master planning. In: Stadtler H, Kilger C (eds) Supply chain management and
advanced planning, 4th edn. Springer, Heidelberg, pp 161–180
SAP (2003) Profitable-to-promise: The next step in the evolution of order promising. http://www.sap.com/
solutions/business-suite/scm/pdf/BWP_profit-to-promise.pdf
Scarf H (1960) The optimality of (s,S) policies for the dynamic inventory problem. In: Arrow KJ, Karlin
S, Suppes P (eds) Mathematical methods in the social sciences. Stanford University Press, Stanford,
California, pp 196–202
Sharman G (1984) The rediscovery of logistics. Harvard Bus Rev 62(5):71–79
Silver EA (1981) Operations research in inventory management: a review and critique. Oper Res 29(4):628–
645
Silver EA, Pyke DF, Peterson R (eds) (1998) Inventory management and production planning and sche-
duling, 3rd edn. Wiley, New York
Smartops (2007) http://www.smartops.com
Spengler T, Rehkopf S, Volling T (2007) Revenue management in make-to-order manufacturing—an
application to the iron and steel industry. OR Spectrum 29(1):157–171
Stadtler H (2008) Purchasing and material requirements planning. In: Stadtler H, Kilger C (eds) Supply
chain management and advanced planning, 4th edn. Springer, Heidelberg, pp 217–230
Stadtler H, Kilger C (eds) (2008) Supply chain management and advanced planning, 4th edn. Springer,
Heidelberg
Talluri KT, van Ryzin GJ (2004) The theory and practice of revenue management. Kluwer, Dordrecht
88 R. Quante et al.
Tempelmeier H (2006) Inventory management in supply networks. problems, models, solutions. Books on
Demand, Norderstedt
Vakali A, Angelis L, Pournara D (2001) Internet based auctions: a survey on models and applications. ACM
SIGecom Exchanges 2(1):6–15
Vollmann TE, Berry WL, Whybark DC (2005) Manufacturing planning and control for supply chain
management, 5th edn. Mcgraw-Hill, New York
Weatherford LR, Bodily SE (1992) A taxonomy and research overview of perishable-asset revenue mana-
gement: yield management, overbooking, and pricing. Oper Res 40(5):831–844
Whitin TM (1955) Inventory control and price theory. Manage Sci 2(1):61–68
Yano CA, Gilbert SM (2003) Coordinated pricing and production/procurement decisions: a review. In:
Chakravarty AK, Eliashberg J (eds) Managing business interfaces: marketing, engineering and manu-
facturing perspectives. Kluwer, Norwell, International Series in Quantitative Marketing, pp 65–103
Active demand management for substitute products
through price optimisation
Aaron A. Levis · Lazaros G. Papageorgiou

OR Spectrum (2007) 29:551–577
DOI 10.1007/s00291-006-0064-1
Abstract This paper presents a systematic mathematical programming

approach for active demand management in process industries. The proposed
methodology aims to determine optimal pricing policies as well as output levels
for substitute products, while taking into consideration manufacturing costs,
resource availability, customer demand elasticity, outsourcing and market com-
petition. First, profit maximisation analytical formulae are derived for deter-
mining Nash equilibrium in prices for a duopolistic market environment where
each company produces only one product. An iterative algorithm is then pro-
posed so as to determine the decision-making process by solving a series of non-
linear mathematical programming (NLP) models before determining the Nash
equilibrium in prices for the competing companies. The proposed algorithm is
extended in order to accommodate the case of multi-product companies, each
one selling a set of substitute products at different prices. The applicability of the
proposed methodology is demonstrated by a number of illustrative examples.
Keywords Active demand management · Substitute products · Price

optimisation · Nash equilibrium · Non-linear mathematical programming
1 Introduction
In today’s global marketplace, process industries no longer compete as inde-

pendent entities but rather as integral part of supply chain links. The ultimate
success of a firm depends on its managerial ability to integrate and co-ordinate
A. A. Levis · L. G. Papageorgiou (B)

Centre for Process Systems Engineering, Department of Chemical Engineering,
UCL (University College London), London WC1E 7JE, UK

90 A. A. Levis, L. G. Papageorgiou
the complex network of business relationships among supply chain members

(Min and Zhou 2002). The recent wave of mergers and acquisitions (M & A) has
led a number of smaller companies to consolidate into a few giant supply chain
firms (e.g. Unilever, Procter & Gamble) that provide close substitute products
(e.g. fast moving consumer goods FMCG) to a wide range of customers. The
intense competition among different companies is evident and occurs in almost
every market sector nowadays.
In this competitive environment, the customer demand is usually satisfied by
a small number of companies, each one manufacturing and selling its individual
subset of products. The goal of every company is to obtain the highest possible
profit by determining optimal price levels for its portfolio products. In that case
an oligopolistic price competitive market environment is established that needs
to account for both competitors’ activities and customers’ willingness to buy.
Duopolistic market competition is the natural starting point for investigat-
ing the behaviour of oligopolies. Consider two companies, company A and
company B, each one manufacturing its own subset of substitute products. By
substitute products we mean slightly differentiated product brands that belong
to the same product-class (e.g. lubricants, detergents, cosmetics, carbonated soft
drinks, etc.). In that case, increased sales of one product result in reduced sales
of another, thus forming a market environment where products brands compete
with each other over a common customer base.
The manufacturing of products usually takes place in production sites owned
by the company (in-house manufacturing). Every site has a limited amount
of available resources used for production. Alternatively, each company may
have the option to allocate manufacturing of a certain amount of products to
a third-party company (outsourcing). As shown in Fig. 1, final products from
each company (in-house manufactured and outsourced) are then transported
to the marketplace in order to satisfy the anticipating customer demand at given
product prices.
A crucial precondition of effective price competition is that customers are
inclined to search for lower-priced substitute products. Low prices however
can kill profit margins and jeopardise the overall company profitability. On the
other hand, high prices will drive away potential customers and inevitably put
company’s market share at risk. Pricing decisions are of crucial importance and
unless taken seriously, they can pose a major threat to the sustainability of the
company.
Traditional approaches for customer demand management assume fixed
product prices and usually rely on forecasting tools, trying to predict customer
demand based on historical sales patterns (Markidakis and Wheelwright 1982).
Passive demand management (PDM) approaches ignore the importance of
flexible product pricing and usually lead to poor customer demand satisfaction.
Modern industrial enterprises are typically multi-product, multi-purpose and
multi-site facilities operating in different countries and dealing with a global-
wide international clientele. In such enterprise networks, the issue of optimal
product pricing policy plays a key role in business performance and necessitates
the appropriate attention. A new trend towards active demand management
Active demand management for substitute products 91
OUTSOURCE OUTSOURCE
SITE 1 MARKET SITE 3
SITE 2 SITE 4
COMPANY A COMPANY B
Fig. 1 Duopolistic market environment
(ADM) has recently emerged focusing on how to actually drive customer

demand away from traditional baseline forecasts so as to maximise both over-
all business performance and customer satisfaction. Many companies rang-
ing from the automotive industry (Ford Motor Co.) to the Internet dotcoms
(Amazon.com) have recently realised the potential benefits of adopting such a
marketing-based concept of smart pricing (Coy 2000). Some firms (e.g. those
operating petrol forecourts) do not hesitate to go even a step further ahead and
employ such clever “dynamic pricing” strategies almost on a daily basis. Accord-
ing to Manugistics (Manugistics Website 2003), a leading company in pricing
and revenue optimisation, pricing is the next battleground for competitiveness.
However, product pricing is not a trivial task. Successful pricing strategies
should consider simultaneously rapidly changing customer expectations, fast-
reacting competitors, complex product interactions and fluctuating manufactur-
ing capacity constraints. Accelerating product lifecycles and increasing product
mix diversity further magnify the complexity of capturing an accurate understat-
ing of the pricing environment and managing a comprehensive strategy around
it (Rapt Website 2003). Lanning et al. (2000) allow demand to be determined
by prices via a constant-elasticity demand function. Prices are then optimised
jointly with capacity investment decisions. Optimal capacity levels and prices for
substitutable products are considered by Birge et al. (1998) in a single-period
model while joint co-ordination of production and marketing decisions are
investigated by Eliashberg and Steinberg (1987). In market-oriented program-
ming, Kaihara (2001) proposes negotiation mechanisms that can lead to pareto
optimal resources allocation in supply chain management. In a subsequent
research paper, Kaihara (2003) formulates the supply chain model as a discrete
resource allocation problem under dynamic environment and demonstrates
the applicability of the virtual market concept and the analysis of the system
behaviour in economic terms.
Although the problem of product pricing is not new in the applied economics
and operational research literature, previous studies adopt a somehow simplis-
tic approach to the problem. They focus their attention on single-product firms
and therefore cannot accommodate the nature of multi-product firms which
are predominant nowadays. Another common drawback is that many studies
consider product pricing in isolation of the market competition, thus ignoring
the reaction effect of rival companies offering substitute products in the

marketplace. Even in the case where competition between single-product firms
is addressed, the number of firms is restricted to two (duopolistic competi-
tion) while joint production and pricing decision-making is based on unreal-
istic assumptions such as unlimited manufacturing resources that render the
proposed solution of impractical value and inapplicable to real-life business
problems.
There exists a clearly identified need to address product pricing issues in a
more realistic context that is able not only to consider simultaneously multi-
product firms competing in an oligopolistic market environment but also pro-
pose alternative pricing policies and production modes such as outsourcing
options. Our proposed methodology of active demand management through
price optimisation is able to capture the dominant trade-off between product
price and product market share so as to deliver value to the customer while
ensuring high profitability for the company.
The rest of the paper is structured as follows. In the next section, the role of
price as a marketing tool is briefly described, while the main characteristics of an
efficient pricing strategy are also discussed. Section 3 presents the case of single-
product price competition between two firms. Analytical formulae are derived
for determining Nash equilibrium in prices while we propose an iterative algo-
rithm validated by a motivating example. In Sect. 4, we extend the proposed
algorithm in order to accommodate the case of multi-product firms operating
in an oligopolistic market environment and address customer demand forecasts
while also considering outsourcing options. A number of illustrative examples
are then studied to demonstrate the applicability of the proposed approach.
Finally, some concluding remarks are drawn in Sect. 5.
2 Pricing strategy for active demand management
The marketing mix is defined as the set of controllable tactical marketing tools
that the firm blends to produce the response it wants in the market place. The
marketing mix consists of everything a firm can do to influence the demand for
its product. The many possibilities gather into four groups of variables known
as the “four P’s”: product, price, place and promotion (Kotler et al. 1996).
In the narrowest sense, price is the amount of money charged for a product or
service. More broadly, price is the sum of all the values that consumers exchange
for the benefits of having or using a specific product or service. Price is the only
element in the marketing mix that produces revenue, while all other elements
represent costs (Kotler et al. 1996). Product, promotion and place are value-cre-
ating activities while pricing can be viewed as the firm’s attempt to capture some
of the created value in the profits earned (Nagle and Holden 1995). Therefore,
pricing is identified as the most flexible element of the marketing mix, since it
is the fastest and most cost-effective way to enhance company profits.
Every company nowadays is operating with a different set of business objec-
tives. Many companies for example set profit maximisation as their ultimate
goal. Other companies however, seek to increase their market share or even
try to augment their customer satisfaction levels. Different business objectives
can be achieved through the employment of alternative pricing strategies such
as skim pricing, penetration pricing, neutral pricing, etc.
Irrespective of the business objectives, an effective pricing strategy should
consider simultaneously the following three main aspects: costs, customers and
competition. Integrating cost management, customer behaviour and market
competition into a unified framework is the key in developing a successful
pricing strategy for active demand management.
2.1 Costs
Costs play a significant role in formulating an efficient pricing strategy. There

can be variable and/or fixed costs. Manufacturing costs are usually variable
costs depending on the sales volume. Traditional pricing strategies are based on
a cost-driven approach as shown in Fig. 2. According to the cost-based pricing
strategy, every product is priced so as to cover its own costs plus make a fair mar-
ginal profit. Although such a strategy seems as a simple guide to profitability, in
practice it does not deliver the desired results. The fundamental problem with
cost-driven pricing is that unit costs cannot be calculated before determining
the product price. The reason for that is that pricing affects sales volumes and
sales volumes in turn affect unit costs (Nagle and Holden 1995).
2.2 Customers
In order to capture the trade-off between price and sales volume, a value-based
pricing strategy can be employed as shown in Fig. 2.
The main difference in this case is the inverse order of decision-making
allowing for a value-based pricing strategy that is more customer-oriented.
Unlike, cost-based pricing, customer’s perceived value is now the driving force
for product pricing. Conjoint-analysis is a market research tool concerned
with understanding how customers perceive product value and how they make
Product Cost Price Value Customer
Cost-based Pricing
Customer Value Price Cost Product
Value-based Pricing
Fig. 2 Cost-based vs. value-based pricing strategies
choices between products based on their individual attributes. BPTO (brand-

price trade-off) is a variation of the conjoint analysis used for testing price
sensitivity in the context of brands available on the market so as to assess brand
preference at any given price scenario.
Price sensitivity can be measured by using the concepts of demand elas-
ticity and cross-elasticity. Price elasticity measures the percentage change in
the quantity demanded relative to the percentage change in price (Pashigian
1998). When there exists a certain degree of substitution between differenti-
ated products, cross-elasticity can be used to measure the percentage change in
the quantity demanded relative to the percentage change in price of another
product, so as to quantify the competition effect between close substitutes
brands (Pindyck and Rubinfield 1992). Estimating elasticity and cross-elasticity
parameters is an active research area while many market research companies
are developing their own methodologies. A paper by Stavins (1997) adopted
from the differentiated-product literature is an illustrative example of demand
elasticity estimation in the personal computer (PC) market while Acutt and
Dodgson (1996) present a method for calculating cross-elasticities between
different public transport modes. Shankar and Krishnamurthi (1996) relate
price sensitivity and price policy from a retailer point of view. Alves and Bueno
(2003) estimate the cross-elasticity between gasoline and alcohol while Tellis
(1988) confirms the negative sign of elasticity parameters. Besanko et al. (1998)
provide an example of price elasticity parameter estimation for two product
categories (yogurt and catsup). Their approach is based on weekly sales data
analysis with main focus on prices and market shares for a 102-week period.
Their framework provides explicit estimates of customers’ willingness to pay
for a brand while taking into account product price responses.
2.3 Competition
Oligopolistic competition has received a great deal of attention in the research

literature (Varian 1992). However, the “oligopoly problem” has proved to be
one of the most resilient problems in the history of economic thought (Vives
1999).
A very early paper written by Hotelling (1929) describes competition among
a small number of firms. His work focuses on spatial competition where the
locations of the products differ. He also makes reference of two earlier devel-
oped models that proceed from different assumptions, namely the Cournot
(1838) and the Bertrand (1883). According to those models, competing firms
only act once and also act simultaneously to determine the outcome of compe-
tition among them. The Cournot model treats output (quantity) as the strategic
decision variable of each firm while the Bertrand model focuses on price as the
strategic decision variable to be determined by each firm.
Smithies (1941) generalised the theory of spatial competition by assuming an
elastic linear demand function at every point of the market and compared differ-
ent cases ranging from monopoly to full competition. Smithies also considers
the effect of the magnitude of the freight rates and changes in marginal costs
for one or both producers.
Kreps and Scheinkman (1983) considered a two-stage duopolistic game. In
the first stage, the two firms determine their capacities while in the second stage
they engage into a Bertrand-type of price competition subject to the capacities
constrains determined previously. Their study emphasises not only the impor-
tance of strategic variables selection (quantity vs price) but also the context of
the (game form) in which those variables are employed.
Apart from the aforementioned Cournot and Bertrand models, oligopolistic
competition formulations include the repeated and the sequential games. The
repeated games can be viewed as series of Cournot-type or Bertrand-type mod-
els not related to each other and solved independently. The sequential games
on the other hand, involve a sequence of decision-making between the firms
where the outcome of competition derives from the interaction of logic-based
firm policies. The Stackelberg model (1934), also known as the leader-follower
model, constitutes an extension of the Cournot model that can be classified as
a sequential game of oligopolistic competition. Output decisions are taken in
turns with the leader-firm making the first move and the follower-firm acting
upon observation of the previous move, resulting in a two-stage game.
A well-respected solution concept for non-cooperative games in oligopolistic
competition is the Nash equilibrium point (Nash 1951) which is defined as the
point where all players in the game do their best given the choice of all other
players. Sherali et al. (1983) study the supply side of an oligopolistic market
supplying an homogeneous product noncooperatively. They characterise the
nature of Stackelberg–Nash–Cournot equilibria and they prescribe methods
for their computation. Sherali and Leleno (1988) present a mathematical pro-
gramming approach for Nash–Cournot equilibrium analysis of oligopolies and
derive equilibrium solutions in various market structures.
The coordination of pricing and production decisions in the face of price
competition is studied by Zhao and Wang (2002). They examine a supply chain
that consists of a manufacturer and a retailer in a leader-follower setting where
both firms try to maximise their respective profits. According to their analysis,
the Stackelberg solution itself will not lead in general to channel optimality and
they provide managerial insight on how to achieve a channel-optimal pricing
policy where both competing parties can benefit from. More recently, Parlar and
Weng (2006) study the effect of coordinating pricing and production decisions
on the improvement of a firm’s position in a price-competitive environment.
They formulate game-theoretical models in order to analyse duopolistic com-
petition between firms facing price-sensitive demand and manage to quantify
the effects of coordinating price and production decisions.
Choi et al. (1990) present a product pricing and positioning methodology in
the face of price competition. They propose both an analytical and a numerical
approach in order to provide qualitative and quantitative solutions respectively.
Despite the difficulty to derive closed form solutions for multi-firms competi-
tion, they suggest a numerical solution approach for single-product firms that
results in an oligopolistic Stackelberg–Nash equilibrium in prices.
A comprehensive review of theories of oligopolistic behaviour (Shapiro

1989) suggests that for similar firms with constant marginal costs and homog-
enous products, the only Nash equilibrium in prices (Bertrand equilibrium)
exists where each firm prices its product at marginal costs. In our case how-
ever, firms are not similar and furthermore their products are not perfect sub-
stitutes and therefore the Bertrand equilibria involve prices above marginal
costs.
Our proposed methodology can be classified as a sequential Bertrand-type
price optimisation approach that aims to determine optimal price levels and
production plans. Our analysis examines the duopolistic case where two firms
produce only one product each. Apart from the derived analytical formulae, a
comprehensive algorithm is also developed. Furthermore, our proposed meth-
odology is extended from single-product to multi-products firms. In addition
to market competition, our proposed mathematical is taking into account cus-
tomer demand forecasts and also considers outsourcing options. Outsourcing
options in conjunction with market competition provide an interesting game-
theoretical insight into the price competition problem as described in Sect. 4.2.5
of this paper.
3 Single-product price competition
In this section, we focus our attention on the specific case of single-product

firms operating in a duopolistic marketplace. Analytical formulae are derived
for that special oligopolistic case, while we propose an iterative algorithm
for determining optimal product prices. A motivating example is then solved
in order to validate the applicability of both formulae and the proposed
algorithm.
3.1 Analytical formulae
Consider two firms 1 and 2, each one offering a single product to the market.
Suppose their products are close substitutes and compete with each other over
the same customer base. However, there is at least some degree of differentia-
tion between the two products and therefore each firm faces different demand
curves (Q1 , Q2 ) and different variable (VC1 , VC2 ) and fixed (FC1 , FC2 ) man-
ufacturing costs while the products are sold for different prices (P1 , P2 ). The
sales volume for every firm is defined as a linear function of its own price (P1 )
and the competitor’s price (P2 ):
Firm 1: Q1 = a1 − b1 · P1 + c12 · P2 (1a)
and
Firm 2: Q2 = a2 − b2 · P2 + c21 · P1 (1b)

where a1 , a2 are demand coefficients, b1 , b2 are demand elasticity parameters

and c12 , c21 are demand cross-elasticity parameters. All parameters in our for-
mulation take positive values.
Note that the quantity each firm can sell decreases when the firm raises its
own price, but increases when its competitor charges a higher price. If both
firms set their prices at the same time, we can use a Bertrand-type model to
determine the resulting equilibrium. Each firm will choose its own price, taking
the competitor’s price as fixed. The profit of firm 1 equals its revenue minus the
variable and fixed manufacturing costs:
1 = (P1 − VC1 ) · min(Q1 , Cap1 ) − FC1 (2)
where Cap1 is the available capacity. Depending on the values of Q1 and Cap1 ,
the profit of Firm 1 equals to

(P1 − VC1 ) · Q1 − FC1 when Q1 < Cap1 ,
1 = (3)
(P1 − VC1 ) · Cap1 − FC1 otherwise
If Q1 equals Cap1 , then
a1 − b1 · P1 + c12 · P2 = Cap1 (4)
and the critical value for the price of Firm 1 equals to
a1 − Cap1 + c12 · P2
P1c = (5a)
b1
Similarly for Firm 2:
a2 − Cap2 + c21 · P1
P2c = (5b)
b2
3.1.1 Case 1: Unconstrained–unconstrained
In that case both firms have unlimited resource capacity, meaning that Q1 <
Cap1 , Q2 < Cap2 , P1 > P1c and P2 > P2c . Using Eq. (2) and substituting Q1
from Eq. (1a), the profit for Firm 1 is calculated as follows:
1 = a1 · P1 − b1 · P12 + c12 · P2 · P1
− a1 · VC1 + b1 · P1 · VC1 − c12 · P2 · VC1 − FC1 (6)
Firm’s 1 profit is maximised when the incremental profit from a very small
increase in its own price is zero. Taking P2 as fixed, Firm 1’s profit is concave in
P1 and therefore the optimal price is given by
∂1 /∂P1 = a1 − 2 · b1 · P1 + b1 · VC1 + c12 · P2 = 0 (7)

This can be rewritten to give the following pricing rule or reaction curve for
Firm1:
a1 + b1 · VC1 + c12 · P2
P1 = (8a)
2 · b1
This equation dictates the price Firm 1 should set, given the price P2 that Firm 2
is setting. Similarly, we can derive the pricing rule (reaction curve) for Firm 2:
a2 + b2 · VC2 + c21 · P1
P2 = (8b)
2 · b2
The point where the two reactions curves cross determines the Nash equilib-
rium in prices. At that point, each firm is doing the best it can, given the price
its competitor has set and therefore, neither firm has the incentive to change its
price.
By substituting Eq. (8b) in (8a), the Nash equilibrium in prices is determined
at point (P1∗ , P2∗ ):
2 · b2 · (a1 + VC1 · b1 ) + c12 · (a2 + VC2 · b2 )

P1∗ = (9a)
4 · b1 · b2 − c12 · c21
and
a2 + b2 · VC2 + c21 · P1∗

P2∗ = (9b)
2 · b2
3.1.2 Case 2: Constrained–constrained
In that case both firms have limited capacity resources meaning that Q1 ≥ Cap1 ,
Q2 ≥ Cap2 , P1 ≤ P1c and P2 ≤ P2c . The resulting profit for Firm 1 is calculated
as follows:
1 = (P1 − VC1 ) · Cap1 − FC1 (10)
In that case, the firm’s profit is a monotonically increasing function of price P1 .

Therefore, reaction curve for Firm 1 is given by the critical price P1c .
a1 − Cap1 + c12 · P2
P1 = (11a)
b1
and similarly for Firm 2 we have
a2 − Cap2 + c21 · P1
P2 = . (11b)
b2
Again, Nash equilibrium in prices is determined at the point where the two
reactions curves cross each other.
By substituting Eq. (11b) in (11a) in Firm 1’s reaction curve, we get
a1 · b2 − Cap1 · b2 + c12 · a2 − c12 · Cap2

P1∗ = (12a)
b1 · b2 − c12 · c21
and
a2 − Cap2 + c21 · P1∗

P2∗ = (12b)
b2
Nash equilibrium in prices is determined at point (P1∗ , P2∗ ).
3.1.3 Case 3: Unconstrained–constrained
In that case it is assumed that Firm 1 has unlimited capacity resources (uncon-
strained) while Firm 2 has a limited amount of capacity resources (constrained),
meaning that Q1 < Cap1 , Q2 ≥ Cap2 , P1 > P1c and P2 ≤ P2c . The reaction curve
for Firm 1 is given by Eq. (8a) as in Case 1.
While the reaction curve of Firm 2 is calculated is given by Eq. (11b) as in
Case 2.
By substituting Eq. (11b) in (8a) the Nash equilibrium in prices is determined
at point (P1∗ , P2∗ ):
a1 · b2 + b2 · VC1 · b1 + c12 · a2 − c12 · Cap2

P1∗ = (13a)
2 · b1 · b2 − c12 · c21
and
a2 − Cap2 + c21 · P1∗

P2∗ = (13b)
b2
3.2 Algorithm A1
In the previous section, the analytical form of the Nash equilibrium was derived
for the case of price competition between two firms that manufacture and sell
two substitute products. Based on the capacity resource levels of each company,
three different cases were studied, namely the unconstrained-unconstrained,
the constrained-constrained and the unconstrained–constrained case, respec-
tively. For each case, the closed form of the resulting Nash equilibrium in prices
was calculated.
In this section, we propose an iterative algorithm [Algorithm A1] able to
accommodate all the aforementioned cases and derive the Nash Equilibrium
point by employing mathematical programming techniques. In any iteration of
the algorithm, each company f decides on its individual pricing policy while
taking into account the price its competitor is currently charging (Pf0 ). The
pricing decision-making process for each firm is formulated as a non-linear
programming (NLP) mathematical model that tries to maximise the company
profit given the competitor’s price. The algorithm terminates when the price
changes become infinitesimal so that a convergence criterion is satisfied for the
prices of both firms. The optimisation problem for company f and rival company
f in iteration m is mathematically formulated as follows:
[Model M1]
max m m m
f = (Pf − VCf ) · Qf − FCf
Subject to:
Demand constraints
m−1
Qm m
f = af − bf · Pf + cff · P
f
Capacity constraints
Qm
f ≤ Capf
The proposed algorithm [Algorithm A1] comprises the following steps:
[Algorithm A1]
Step 1. Set price levels to current market prices and initialise

iterations counter m:=0.
Step 2. Set iterations counter to m:= m+1. If m > mmax then STOP.
Step 3. For every company f compute the prices for all products of the
company using Model M1.
m
f
−m−1
f
Step 4. If m
≤ ε for all companies then STOP. Otherwise, go to
f
Step 2.
The proposed algorithm determines the pricing decision-making process be-

tween two competing firms. It should be noted that the equilibrium product
prices are computed by a central decision-making by applying algorithm A1.
It is also assumed that the central decision-maker has knowledge of the costs
and demand functions of all companies involved. Each firm decides on its
optimal pricing policy while taking into account the observable current price
charged by its competitor firm. Therefore, the algorithm is able to capture the
game-theoretical nature of the pricing problem and successfully determines the
sequential decision-making process between the two firms. The algorithm ter-
minates at a point where neither company wants to change its pricing policy
given the price of its competitor. At that point both companies are doing their
best, therefore neither company wants to deviate from that point and that is
by definition, the Nash equilibrium point in prices. The applicability of the pro-
posed algorithm is demonstrated by solving a motivating example as described
in the following section.
Table 1 Additional input

Parameter Firm 1 Firm 2
data for the motivating
example
Variable cost (VC) 0.5 0.4
Fixed cost (FC) 20 25
Capacity (Cap) in Case 1 Unlimited Unlimited
Capacity (Cap) in Case 2 60 50
Capacity (Cap) in Case 3 Unlimited 50
3.3 Example for single-product firms
Consider two firms that offer two differentiated products that are close substi-
tutes to each other. Suppose that the two companies are facing the following
demand curves:
Firm 1: Q1 = 160 − 30 · P1 + 4 · P2 and
Firm 2: Q2 = 180 − 40 · P2 + 3 · P1
The additional input data concerning the two companies is shown in Table 1.
Three different cases, namely Case 1, Case 2 and Case 3 are examined based on
the capacity resource availability. In Case 1, both firms have unlimited capacity
resources. According to the analytical form equations derived in the previ-
ous section, the Nash Equilibrium in prices is determined at point (P1∗ , P2∗ ) =
(3.09, 2.57). Also, Algorithm A1 successfully predicts the same Equilibrium
point within less than three iterations depending on the starting point (initial
price vector1 ) as illustrated in Fig. 3. Most importantly, product prices converge
to the same equilibrium point irrespective of the starting point, thus illustrating
the robustness of the proposed methodology.
In Case 2, both firms have limited amounts of capacity resources, therefore
their output levels are restricted by the resource availability of every firm.
Consequently, the Nash equilibrium in prices is also influenced by the lack of
unlimited resources. According to the theoretically derived equations, Nash
equilibrium in prices is now determined at point (P1∗ , P2∗ ) = (3.80, 3.54). The
proposed algorithm derives the exact same equilibrium point irrespective of
the initial price vector employed as shown in Fig. 4. It is very interesting to
notice that the equilibrium prices in this case are slightly higher than the equi-
librium prices in Case 1. This is mainly attributed to the fact that the outputs
in Case 2 are restricted to the available resource levels. At the equilibrium
point, both companies make full utilisation of their resources, producing 50
and 60 units of product respectively which are less compared to the equilib-
rium outputs in Case 1 (77.6 and 86.6, respectively). In order to compensate
for the decreased output levels, both firms are now forced to raise their prices
so as to maximise their profits. Finally in Case 3, Firm 1 has unlimited amount
of capacity resource while Firm 2 has a finite level of capacity resource. The
closed form equations predict that the Nash equilibrium in prices lies at point
1 For all cases examined, four different initial price vectors (P0 , P0 ) were used as follows (1, 1), (2, 2),
1 2
(3, 3) and (4, 4).
4.5
4 4
3.5
3.18
3.09 3.09 3.09 3.09 3.09 3.09 3.09
3
Price level
2.57 2.57 2.57 2.57 2.57 2.57 2.57 2.57

2.5
1.5 P1_1 P2_1
P1_2 P2_2
1
P1_3 P2_3
0.5 P1_4 P2_4
0
0 1 2 3 4 5 6 7 8
Iteration
Fig. 3 Nash equilibrium in prices (Case 1)
4.5
4 4
3.87 3.81 3.8 3.8 3.8 3.8 3.8 3.8
3.5 3.54
3.54 3.54 3.54 3.54 3.54 3.54 3.54
3
Price level
2.5
2
P1_1 P2_1
1.5
P1_2 P2_2
1 P1_3 P2_3
P1_4 P2_4
0.5
0
0 1 2 3 4 5 6 7 8
Iteration
(P1∗ , P2∗ ) = (3.15, 3.49). The proposed algorithm converges at the exact same
equilibrium point as shown in Fig. 5.
Unlike Cases 1 and 2, the equilibrium price for Firm 1 is now slightly lower
than the price charged by Firm 2. Firm 2 has a limited capacity resource and
therefore its equilibrium output is restricted to 50 product units. The lack of
resources for Firm 2 is inevitably reflected on the resulting high price. On the
other hand, Firm 1 is able to produce a larger output and charge a lower price
4.5
4 4
3.49 3.49 3.49 3.49 3.49 3.49 3.49 3.49

3.5
3.18 3.15 3.15 3.15 3.15 3.15 3.15 3.15
3
Price level
2.5
1.5 P1_1 P2_1
P1_2 P2_2
1
P1_3 P2_3
0.5 P1_4 P2_4
0
0 1 2 3 4 5 6 7 8
Iteration
for its product so as to benefit from the economy of scale. The results from Case
3 clearly illustrate what takes place in a real life marketplace, where it is very
common that big companies supply large quantities in relatively low prices and
consequently outrun the small companies who struggle to cover their costs by
charging high prices.
The Nash equilibrium points for Cases 1, 2 and 3 are summarised in Table 2.
In all three cases the proposed algorithm A1 successfully determines the same
Nash equilibrium point as the one predicted from the closed form equations.
Algorithm A1 is further extended so as to accommodate the case of multi-prod-
uct competing firms trying to satisfy the anticipated customer demand forecast
while considering outsourcing options as described in the next section.
4 Multi-product price competition
In the previous sections, we investigated the case of price competition in a

duopolistic market environment where each company is producing only one
product. However, process industries nowadays usually operate multi-product
Table 2 Nash equilibrium points for the motivating example
Case 1 Case 2 Case 3
Firm 1 Firm 2 Firm 1 Firm 2 Firm 1 Firm 2
Price 3.09 2.57 3.80 3.54 3.15 3.49

Output 77.6 86.6 60 50 79.47 50
Profit 180.89 162.63 178.28 131.77 190.53 129.31
plants producing a set of differentiated products (e.g. different paints, deter-

gents, carbonated drinks, etc.). These products belong to the same family of
products (product class) and they share a number of common characteristics
(e.g. water-based paints). On the other hand, products are differentiated from
each other in such a way so as to cover a broad range of customer preferences
(e.g. different paint colours/quality).
Market segmentation is a widely used marketing strategy that recognises
the different ways customers perceive product value and make their purchase
choices accordingly. In order to deliver value to the different existing customer
segments, most companies decide to market launch a wide variety of slightly
differentiated products so as to attract customers via a tailor-based market-
ing approach. Each one of the company products is a unique brand name with
unique features that clearly differentiates itself from the rest of the family prod-
ucts. The unique attributes of each product appeal to a very distinct customer
base that is choosing to buy that specific product over the entire range of prod-
ucts present in the marketplace. Customers are willing to buy their preferred
product as long as the product price charged by the company reflects their
perceived value of the product. Product brand loyalty is expressed by repeat
purchases of the installed customer base. Alternatively, the customer may well
switch to a lower-priced substitute product offered by the same company or a
rival company.
Pricing in a multi-product competitive market environment is not an easy
task. The analytical formulae presented in Sect. 3 for the two products pric-
ing problem cannot be applied to the multi-product pricing problem so as to
derive a meaningful Nash equilibrium in prices. In the multi-product case, price
competition exists not only between company and competitor products but also
between differentiated brands belonging to the very same company. Moreover,
company products are manufactured by utilising a common pool of available
resources. Family products are therefore competing with each other for scarce
and shared manufacturing resources. Therefore, product cannibalisation effects
have to be seriously taken into account when determining an optimal pricing
policy. On the same time though, the company has to account for the market
competition by considering the pricing policy adopted by the rival company for
its products also present in the marketplace. On top of that modern process
industries have recently realised the benefits of adopting outsource manufac-
turing policies in an attempt to drive manufacturing costs further down and
avoid any unnecessary capacity expansion overheads. Such outsource options
should be addressed in a proper manner before deciding on a comprehensive
pricing strategy.
Reaction curve analysis cannot be applied in a straightforward way as in the
previous two-products case. However, in order to capture the trade-off between
product price and market share in a multi-product environment, an extension of
the previously developed non-linear programming (NLP) mathematical model
is proposed. Based on that mathematical model, Algorithm A1 is extended in
order to determine optimal pricing polices for multi-product competing com-
panies.
4.1 Mathematical model
The following nomenclature is used in our mathematical model formulation:
Indices
f companies
i, j products
s production sites
r resources
Sets
SPf set of products i for company f
Sf set of production sites s for company f
Rf set of resources r for company f
Zf set of products i using resource r at site s in company f
Parameters
ai demand coefficient for product i
bi demand elasticity coefficient for product i
cij demand cross-elasticity coefficient between products i and j
rfcs relative fixed cost coefficient for site s
rtcs relative transportation cost coefficient for site s
rvcs relative variable cost coefficient for site s
VCi variable manufacturing cost for product i
FCi fixed manufacturing cost for product i
TCi unit transportation cost for product i
OCi unit outsource cost for product i
ρir unit consumption coefficient for product i using resource r
Ars availability level of resource r at site s
DF total market demand forecast
Variables
Pi price for product i
Vi sales volume for product i
Qis amount of product i manufactured at site s
Oi amount of product i outsourced
f total profit for company f
The derivation of the general mathematical model for company f [Model

M2] is described next. The sales volume for every product i is a monotonically
decreasing function of its price and a monotonically increasing function of the
price of all other competing products, including substitute products belonging
to company f as well as competitor products. The sales volume for every product
i is given by the following linear function:

Vi = ai − bi · Pi + cij · Pj ∀i (14)
j=i
The total sales volume of every product i equals to the amount manufactured
in-house at all production sites belonging to company f plus the amount manu-
facturing from outsourcing:

Vi = Qis + Oi ∀i ∈ SPf (15)
s∈Sf
The amount of company products manufactured in-house at every production

site s is limited by the availability of the shared company resources. The follow-
ing constraints safeguard that the resource availability levels are not exceeded:

ρir · Qis ≤ Ars ∀r ∈ Rf , s ∈ Sf (16)
i∈Zf
Market research surveys are conducted periodically so as to assess the current

trends and predict future customer demand of a specific product class. The total
sales volume of all products present in the marketplace should be greater or
equal to the forecasted customer demand:

Vi ≥ DF (17)
i
The objective function employed in our mathematical model corresponds to

the net profit generated by the subset of the products belong to company f.
The net profit is calculated as sales revenue minus the different costs, namely
variable and fixed manufacturing costs, transportation and outsourcing costs.
Mathematically we have
max

f = Pi · Vi − rvcs · VCi · Qis − rfcs · FCi
i∈SPf i∈SPf s∈Sf i∈SPf s∈Sf

− rtcs · TCi · Qis − OCi · Oi (18)
i∈SPf s∈Sf i∈SPf
4.1.1 Summary of the mathematical model
In the general case, the optimisation problem for company f is mathematically

formulated as follows:
[Model M2]
max

f = Pi · Vi − rvcs · VCi · Qis − rfcs · FCi
i∈SPf i∈SPf s∈Sf i∈SPf s∈Sf

− rtcs · TCi · Qis − OCi · Oi
i∈SPf s∈Sf i∈SPf
Subject to:

Vi = ai − bi · Pi + cij · Pj ∀i
j=i

Vi = Qis + Oi ∀i ∈ SPf
s∈Sf

ρir · Qis ≤ Ars ∀r ∈ Rf , s ∈ Sf
i∈Zf

Vi ≥ DF
i
Clearly, the restrictions imposed by the analytical formulae are now alleviated
and Model M2 is able to accommodate the case of oligopolistic market compe-
tition where more than two firms are competing. Furthermore, every company
present in the marketplace is allowed to manufacture in-house and/or outsource
more than one product. The Nash equilibrium in prices is obtained by employing
the same steps as the previously proposed algorithm [Algorithm A1] with the
difference being the optimisation model [Model M2] used for each company.
The proposed methodology is able to determine optimal production policies
and prices for all products, as it is demonstrated by the illustrative example
described in the next section.
4.2 Illustrative examples for multi-product firms
Consider two firms, namely company A and company B that manufacture and
sell products P1, P2, P3, P4 and P5, P6, P7, respectively, as shown in Fig. 1.
Products P1–P7 are close substitutes to each other and therefore each product
has a unique demand function curve associated with it, as described by Eq. (14).
Demand coefficients include parameter αi , elasticity bi and cross-elasticity cij
parameters as shown in Tables 3 and 4. Every company has two available man-
ufacturing sites (sites 1 and 2 for company A and sites 3 and 4 for company B).
The products can be manufactured in-house by using shared manufacturing
resources available at each site (in-house manufacturing). Resource utilisation
coefficients for every product are given at Table 5 while resource availability
levels for every resource at each site are given at Table 6. Manufacturing sites
are geographically distributed facilities, therefore relative manufacturing cost
and transportation cost coefficient are used so as to capture the effect of differ-
ent manufacturing locations (see Table 7). Final products are transported from
the manufacturing sites to the end-customers at a given transportation cost TCi .
Alternatively, a certain amount of production can be outsourced to a third-party
company at a given outsource cost OCi . Note that since we are dealing with
products belonging to the same product class, fixed costs are assumed to be
the same for all products, therefore they are not considered explicitly in the
illustrative example.
Table 3 Product input data

Product ai bi VCi TCi OCi Pi0
for the illustrative example
P1 160 25 4 1 5.3 8
P2 200 30 3 1 5.3 10
P3 150 25 4 1 5.3 9
P4 120 20 5 1 5.3 11
P5 170 30 3 1 4.2 9
P6 110 25 4 1 4.2 10
P7 180 30 3 1 4.2 8
Table 4 Cross-elasticity
Product P1 P2 P3 P4 P5 P6 P7
parameters (cij ) for the
illustrative example
P1 – 4 3 2 6 7 4
P2 2 – 5 3 5 4 2
P3 3 3 – 2 3 2 5
P4 4 2 4 – 2 6 3
P5 2 4 3 5 – 3 2
P6 5 3 4 3 2 – 3
P7 3 3 2 2 4 3 –
Table 5 Resource utilisation

Product res1 res2 res3 res4 res5 res6
coefficients
P1 1 1.1 0.8 – – –
P2 0.7 1.2 0.7 – – –
P3 1.2 1.4 0.9 – – –
P4 1.1 1.3 0.4 – – –
P5 – – – 0.9 1.2 0.7
P6 – – – 0.8 1.4 0.6
P7 – – – 1 1.6 0.8
Table 6 Initial resource

Resources Site 1 Site 2 Site 3 Site 4
availability levels across
manufacturing sites
Res1 200 340 – –
Res2 300 370 – –
Res3 150 270 – –
Res4 – – 140 150
Res5 – – 210 220
Res6 – – 110 130
Table 7 Manufacturing site

Manufacturing. Site rvcs rtcs
related data
Site 1 1 1
Site 2 0.8 1.2
Site 3 0.7 1.4
Site 4 0.9 1.1
Table 8 Initial state of

Products Site 1 Site 2 Outsource
company A
P1 – 205 –
P2 95 – –
P3 68 20 –
P4 – 90 –
Total 163 315 0
Table 9 Initial state of

Products Site 3 Site 4 Outsource
company B
P5 84 – –
P6 41 – –
P7 32 68 –
Total 157 68 0
Given an initial price vector (current market prices Pi0 ) for all products, the
problem is to determine optimal product prices, output levels and outsource
amounts so as to derive a comprehensive Nash equilibrium point for companies
A and B that neither company would wish to deviate from.
Initially, both companies A and B manufacture their products in-house by
only relying on the manufacturing capabilities of their production sites while no
outsourcing is considered. In particular, the allocation of production between
the different sites is shown in Tables 8 and 9. The total amount of sales for
the specific product class equals the combined manufacturing volume of both
companies (703 units). Given the initial price vector and output levels for all
products, the initial profit is 2,194 rmu2 and 1,129 rmu for company A and B,
respectively.
A recent market research survey has estimated future customer demand for
the product family under investigation to be equal to 712 units and therefore
the two companies are competing over the anticipated customer demand. Every
company has the strategic choice to consider outsourcing options or rely entirely
on its own in-house manufacturing capabilities, thus resulting in four distinct
cases as explained in the following sections. All four cases were implemented
in GAMS (Brooke et al. 1998) using the CONOPT NLP solver (Drud 1985)
while all runs were performed on an IBM RS/6000 workstation.
4.2.1 Case 1: In-house/in-house
In this case both companies A and B manufacture their products in-house while
no outsourcing is allowed to take place. Model M2 is solved with the outsource
variable fixed to zero for both companies. As shown in Fig. 6, Nash equilibrium
is reached after five iterations resulting in profits 2,259 rmu and 1,262 rmu for
2 rmu = relative monetary units.

3000
2500 2405
2194 2250 2259 2259 2259
2000
Profit Level
1500
1211 1261 1262 1262 1262
1129
1000
500
Company A
Company B
0
0 1 2 3 4 5
Iteration
Fig. 6 Nash equilibrium for Case 1
12 Initial Price
Equilibrium Price
10
8
Price Level
0
P1 P2 P3 P4 P5 P6 P7
Product
Fig. 7 Product prices for Case 1
company A and B, respectively. Optimal product price levels are determined

as illustrated in Fig. 7. More specifically, equilibrium prices for P1 and P3 lie
above their original levels while a price decline is suggested for products P2, P5
and P6. Finally, the optimal prices of products P4 and P7 are very close to their
original values.
Table 10 Nash equilibrium results across all cases
Company A Company B
Profit P1 P2 P3 P4 Profit P5 P6 P7
Initial 2194 8 10 9 11 1129 9 10 8

Case 1 2259 10.46 8.52 9.50 11.06 1262 8.09 8.74 7.78
Case 2 2241 10.45 8.54 9.51 11.05 1284 8.07 8.57 7.78
Case 3 2288 10.49 8.55 9.53 10.92 1259 8.08 8.74 7.77
Case 4 2269 10.48 8.57 9.54 10.90 1281 8.06 8.56 7.77
4.2.2 Case 2: In-house/outsource
In this case company A manufactures its products entirely in-house while com-
pany B has the option to outsource a certain amount of production. Model M2
is solved with the outsource variable fixed to zero only for company A. Start-
ing from the same initial state as in case 1, Nash equilibrium results in profits
2,241 rmu and 1,284 rmu for company A and B, respectively. In this case, com-
pany B outsources 84 units of product P6. Optimal product prices are shown in
Table 10.
4.2.3 Case 3: Outsource/in-house
This case is the exact inverse of case 2. Company B manufactures all of its prod-
ucts in-house while company A has the option to outsource a certain amount
of production. According to the results, profits of 2,288 rmu and 1,259 rmu for
company A and B, respectively, are achieved at the Nash equilibrium point.
Company A outsources 91 units of product P4 while optimal product prices are
given in Table 10.
4.2.4 Case 4: Outsource/outsource
In this case both companies A and B have the option to manufacture products
in-house and/or outsource a certain amount. Nash equilibrium results in profits
2,269 rmu and 1,281 rmu for company A and B, respectively. In this case, com-
pany A outsources 90 units of product P4 and company B outsources 84 units
of product P6. Optimal product prices can be found in Table 10. The alloca-
tion of production between manufacturing sites and outsourcing in Case 4 are
given as pie charts in Fig. 8 for both companies. Notice that the largest share of
production is allocated to site 2 and site 3 since they both offer low variable man-
ufacturing cost compared with sites 1 and site 4, respectively. According to the
obtained results, outsourcing activity constitutes over 20% of total production
for both companies.
Company A Company B
Site 1
5%
Outsource
22%
Outsource
28%
Site 3
52%
Site 4
Site 2 20%
73%
Fig. 8 Allocation of production in Case 4
4.2.5 Game-theoretical insight
In the previous subsections we examined four different cases of duopolistic

competition. By comparing the derived Nash equilibria, useful insight can be
gained from a game-theoretical point of view. The duopolistic game under inves-
tigation is defined as follows. The competing companies are regarded as two
players. Each player in the game has a number of possible strategies, courses of
action that he may choose to follow. In our particular case, companies have the
choice to either produce their products entirely in their own manufacturing sites
(in-house strategy) or produce a certain amount in-house while outsourcing a
certain percentage of production (outsource strategy). The strategies chosen by
each player determine the so-called outcomes of the game. In our example, we
end up with four different outcomes, namely in-house/in-house, in-house/out-
source, outsource/in-house and outsource/outsource, each one representing a
case examined in the previous sections. In every formally stated game, there is
a collection of numerical payoffs, one to each player, associated with every pos-
sible outcome of the game. Those payoffs represent the value of the outcome to
the different players. In our example, Nash equilibrium profits can play the role
of companies payoffs for every particular case examined. Overall, we are deal-
ing with a two-person game with two strategies per player and a game payoff
matrix as shown in Table 11. The values in parentheses are the Nash equilibrium
profits determined previously for all four cases, with the first number being the
profit for company A and the second one the profit of company B.
Game theory is the study of how players should rationally play games. Each
player would like the game to end in an outcome which offers him the largest
possible payoff. He has some control over the outcome, since his choice of
strategy will influence it. However, the outcome is not determined by his choice
alone, but also depends upon the choices of all other players. In general, there
Table 11 Game payoff matrix
Company A
In-house Outsource
Company B In-house Case 1 (2,259;1,262) Case 3 (2,288;1,259)

Outsource Case 2 (2,241;1,284) Case 4 (2,269;1,281)
might be conflict because different players value outcomes differently (Straffin

1993).
In our example companies are faced with the question of which strategy
to adopt in order to reach the Nash Equilibrium associated with the highest
profit for the company under investigation. First, let us consider company A.
Company A does not have any indication of which policy rival company B will
adopt. If company B adopts a strictly in-house manufacturing policy, then com-
pany A has a choice between Case 1 and 3. Since the profit for company A in
Case 3 is higher than the one in Case 1 (2,288 vs 2,259), company A decides to
adopt an outsourcing strategy. In case company B adopts an outsourcing policy,
company A has a choice between Case 2 and 4. Case 4 offers company A with a
profit of 2,269 which is higher that the Case 2 profit (2,241). So, in both scenar-
ios, company A is better off by choosing to outsource a certain amount of its
production, irrespective of the production policy adopted by rival company B.
Similarly, we can prove that the exact same rule applies for company B as well.
Without any prior knowledge of the production policy adopted by company A,
company B always earns a higher profit by adopting an outsourcing strategy.
It is very interesting to notice that Case 3 provides the highest profit for
company A while Case 2 provides the highest profit for company B. However,
Case 4 is considered to be the most likely outcome of the game since the out-
source/outsource policy guarantees higher profits for both companies no matter
what policy the rival company decides to adopt, thus providing a robust Nash
equilibrium for both companies.
5 Concluding remarks
A systematic mathematical programming approach for active demand manage-

ment through price optimisation was presented in this paper. First, we derived
analytical formulae for calculating Nash equilibrium points in a duopolistic
market environment where each company produces and sells only one product.
An iterative algorithm was then proposed that derived the exactly same equi-
librium points as predicted by the closed-form formulae. Following that, the
proposed algorithm was further extended in order to accommodate the case
of multi-product firms and also consider additional features such as customer
demand forecast and mixed in-house and outsourcing production policies. An
illustrative example was solved in order to demonstrate the applicability of the
proposed methodology across four different case studies. Finally, a comparison

among the different cases provided us with valuable game-theoretical insight
concerning the problem of duopolistic competition coupled with outsourcing
options.
References
Acutt MZ, Dodgson JS (1996) Cross-elasticities of demand for travel. Transp Policy 4:271–277
Alves DCO, Bueno RD (2003) Short-run, long-run and cross-elasticities of gasoline demand in
Brasil. Ener Econ 25:191–199
Bertrand J (1883) Theorie mathematique de la richesse sociale. J Savants 499–508
Besanko D, Gupta S, Jain D (1998) Logit demand estimation under competitive pricing behaviour:
An equilibrium framework. Manage Sci 44:1533–1547
Birge JR, Drogosz J, Duenyas I (1998) Setting single-period optimal capacity levels and prices for
substitutable products. Int J Flexible Manuf Syst 10:407–430
Brooke A, Kendrick D, Meeraus A, Raman R (1998) GAMS: a user’s guide. GAMS Development
Corporation, Washington
Choi SC, Desarbo WS, Harker PT (1990) Product positioning under price competition, Manage
Sci 36: 175–199
Cournot A (1838) Recherches sur les principes mathematiques de la theorie des richesses. Hachette,
Paris
Coy P (2000) The power of smart pricing. Economics April 10th, Business Week
Drud AS (1985) CONOPT: a GRG code for large sparse dynamic nonlinear optimization problems.
Math Progr 31: 153–191
Eliashberg J, Steinberg R (1987) Marketing-production decisions in an industrial channel of distri-
bution. Manage Sci 33: 981–1000
Hotelling H (1929) Stability in Competition. Econ J 39: 41–57
Kaihara T (2001), Supply chain management with market economics. Int J Product Econ 73: 5–14
Kaihara T (2003), Multi-agent based supply chain modelling with dynamic environment. Int J
Product Econ 85: 263–269
Kotler P, Armstrong G, Saunders J, Wong V (1996) Principles of marketing. Prentice-Hall, Engle-
wood Cliffs
Kreps D, Scheinkman J (1983) Quantity pre-commitment and Bertrand competition yield Cournot
outcomes. Bell J Econ 14: 326–337
Lanning S, Mitra D, Wang Q, Wright M (2000) Optimal planning for optical transport networks.
Philos Trans Roy Soc Lond Ser A–Math Phys and Engi Sci 358: 2183–2196
Manugistics Inc., http://www.manugistics.com/
Makridakis S, Wheelwright SC (1982) The handbook of forecasting: a manager’s guide. Wiley,
Newyork
Min H, Zhou G (2002) Supply chain modeling: past, present and future. Comput Indus Eng 43:
231–249.
Nagle T, Holden R (1995) The strategy and tactics of pricing: A guide to profitable decision making,
2nd edn., Prentice-Hall, Englewood Cliffs
Nash JF (1951) Non-cooperative games. Ann Math 54: 286–295
Parlar M, Weng ZK (2006) Coordinating pricing and production decisions in the presence of price
competition. Euro J Operl Res 170: 211–227
Pashigian BP (1998) Price theory and applications, 2nd edn. McGraw-Hill, Boston
Pindyck RS, Rubinfield DL (1992) Microeconomics, 4th edn. Prentice-Hall, Upper Saddle River
Rapt Inc., http://www.rapt.com/
Shankar V, Krishnamurthi L (1996) Relative price sensitivity to retailer promotional variables and
pricing policy: an empirical analysis. J Retail 72: 249–272
Shapiro, AC (1989) Theories of oligopoly behavior. In: Schmalensee R, Willig RD (eds.)Handbook
of industrial organization 329–414
Sherali HD, Soyster AL, Murphy FH (1983) Stackelberg-Nash-Cournot equilibria:characteriza-
tions and computations. Oper Res 31: 253–276
Sherali HD, Leleno JM (1988) A mathematical programming approach to a Nash-Cournot equi-

librium analysis for a two-stage network of oligopolies. Oper Res 36: 682–702
Smithies A (1941) Optimum Location in Spatial Competition. J Polit Econ 49: 423–439
Stackelberg VH (1934) Marktform und Gleichgewicht. Springer, Vienna
Stavins J (1997) Estimating demand elasticities in a differentiated product industry: the personal
computer market. J Econ Busin 49: 347–367
Straffin PD (1993) Game theory and strategy. The Mathematical Association of America, Wash-
ington
Tellis GJ (1988) The price elasticity of selective demand: a meta-analysis of econometric models of
sales. J Market Res 25: 331–341
Varian HR (1992) Microeconomic analysis, 3rd edn. W.W. Norton, New York
Vives X (1999) Oligopoly pricing. The MIT Press, Cambridge
Zhao W, Wang Y (2002) Coordination of joint pricing-production decisions in a supply chain. IIE
Trans 34: 701–715
Customer segmentation, allocation planning and order
promising in make-to-stock production
Herbert Meyr

OR Spectrum (2009) 31:229–256
DOI 10.1007/s00291-008-0123-x
Abstract Modern advanced planning systems offer the technical prerequisites for
an allocation of “available-to-promise” (ATP) quantities—i.e. not yet reserved stock
and planned production quantities—to different customer segments and for a real
time promising of incoming customer orders (ATP consumption) respecting allocated
quota. The basic idea of ATP allocation is to increase revenues by means of customer
segmentation, as it has successfully been practiced in the airline industry. However,
as far as manufacturing industries and make-to-stock production are concerned, it is
unclear, whether, when, why and how much benefits actually arise. Using practical data
of the lighting industry as an example, this paper reveals such potential benefits. Fur-
thermore, it shows how the current practice of rule-based allocation and consumption
can be improved by means of up-to-date demand information and changed customer
segmentation. Deterministic linear programming models for ATP allocation and ATP
consumption are proposed. Their application is tested in simulation runs using the
lighting data. The results are compared with conventional real time order promising
with(out) customer segmentation and with batch assignment of customer orders. This
research shows that—also in make-to-stock manufacturing industries—customer seg-
mentation can indeed improve profits substantially if customer heterogeneity is high
enough and reliable information about ATP supply and customer demand is available.
Surprisingly, the choice of an appropriate number of priority classes appears more
important than the selection of the ATP consumption policy or the clustering method
to be applied.
H. Meyr (B)
Chair of Production and Supply Chain Management,
Technical University of Darmstadt,

118 H. Meyr
Keywords Available-to-Promise (ATP) · Advanced planning systems · Clustering ·

Integer and linear programming · Order promising
1 Introduction
One of the biggest challenges in airline industry is to avoid that a hasty, high margin
business class customer cannot get a seat because a low price economy customer
has booked the last one a few minutes ago. Revenue management has developed
techniques to treat such problems adequately, e.g. to establish and fence off customer
segments in form of booking classes and to determine booking limits. The situation is
different in make-to-stock (MTS) supply chains of consumer goods industries where
final item stocks are built up on basis of forecasts and customer requests are served from
this stock. But not too different. Here, too, exist more important and less important
customers yielding higher and lower profit margins. Here, too, occur shortages. And
a service level of 98 percent also implies that two percent of the customers have not
been served as desired. This may concern several dozens of orders per day, for a single
item only. “Not as desired” not necessarily means that the customers are not supplied
at all. However, late deliveries lead to customer annoyance and customer migration in
the long term. Thus here, too, it is important to consider carefully who gets its goods
on time and—even more crucially—who does not.
Actually “order promising”, i.e. communicating the customer a reliable and
hopefully soon delivery date, is the planning task to be considered. However, in
MTS situations order promising also means deciding about— and for short-term
orders simultaneously releasing—delivery (see Fleischmann and Meyr 2003a). Thus,
these decisions about actual deployment can hardly be re-thought. In order to promise
reliable delivery dates, modern enterprise resources planning (ERP) systems or advan-
ced planning systems (APS) build on up-to-date information about stock on hand and
planned supply of the distribution centers that both not yet have been assigned to
customers. Such unreserved quantities are called “available-to-promise”(ATP). Since
production has to be planned on basis of forecasts (push concept implied by MTS),
unused production capacity, sometimes called “capable–to–promise”, and stock re–
filling are no more concern at this point in time. The information about the planned
supply of the distribution centers either stems from the short–term master production
schedule of a single, corresponding production plant or—for a longer preview—even
from a mid–term production and delivery plan (“master plan”) of the overall supply
chain (see e.g. Kilger and Meyr 2008).
Usually two different modes of promising ATP to incoming customer orders are
distinguished, “batch order processing” and real time “single order processing” (see
e.g. Ball et al. 2004; Fleischmann and Meyr 2003a; Pibernik 2005). In batch mode,
an order is not promised immediately upon request, but held back. It is then assigned
to ATP inventories together with several other orders in a “batch”. Thus, there must
be enough time to gather these orders and a customer must be willing to wait for an
answer. Often, this “batching horizon” comprises several hours or a whole day.
Sometimes customers expect an immediate answer for their order query. In this
case batching of orders is not possible. Thus, each single order has to be processed in
real time and ATP is consumed in a first-come-first-served (FCFS) manner.
Customer segmentation, allocation planning 119
As addressed in the airline example above, in shortage situations, where demand

is higher than capacities—i.e. in this case than ATP inventories—single order proces-
sing entails the danger of promising scarce inventory to the wrong customers, e.g. to
less important customers or to customers showing smaller profit margins. Allocation
planning, as propagated by APS vendors like i2 and SAP (see Kilger and Meyr 2008),
promises to be a way to improve real time single order processing by reserving shares
of the ATP, the so-called “quotas” or “allocated ATP”, for important customers in the
medium term and afterward promising orders with respect to these allocated quotas
in the short term. That means ATP is held back in anticipation of later arriving, more
profitable orders even if a less profitable order already requests this stock. Such an
allocation of quotas shall take advantage of a customer segmentation into low and
high priority customers as it has shown to be successful in airline industries. This
leads to a two step ATP allocation and ATP consumption process, in the following
called “allocation planning and ATP consumption” (AP&C).
It is important to note that such a segmentation already appears useful if the same
product is sold for different profits or with different priorities. For example, various
sales channels might generate different profit margins because sales prices vary due to
country-specific tax levels or due to differing transport costs. Or in-house customers
might show other strategic importance for a company than external customers. All in
all, the AP&C approach promises to be useful for companies which produce storable
standard products in high volume on an MTS basis and whose multitude of customers
are heterogeneous in the above sense. Then, there is the hope that the same or even
better profits as in batch mode can be achieved, even though a customer gets his answer
immediately.
The intention of this paper is to structure the AP&C process and to reveal the
potential benefits of allocation planning as compared to the common practices of FCFS
single order processing or batch order processing. However, before the contribution
of the paper can be specified in some more detail, a brief review of current practices
and existing literature is necessary.
1.1 Literature review
For a literature review we will concentrate on ATP support for commercial (ERP and)
advanced planning systems and especially discuss papers which tackle ATP allocation
or consumption in more detail. Note that we focus on MTS situations, i.e. the ATP
supply of finished items is assumed to be fixed because it bases on the stock on hand
and on the production quantities that have already been planned in the short-term
production scheduling module of the APS and/or a mid–term master planning module
(see e.g. Meyr et al. 2008). This rules out literature on make–to–order (MTO) and
assemble–to–order (ATO) supply chains, which most of the due date setting (see e.g.
Keskinocak and Tayur 2004) and batch order promising models (see e.g. Chen et al.
2001, 2002) have been developed for. In these situations customers are usually willing
to wait longer for an order promise than in MTS supply chains. This also rules out
inventory rationing (see e.g. de Vericourt et al. 2002), which explicitly allocates stocks
on hand to several customer classes, but assumes that the refilling of the stock can still
be influenced by means of orders. Finally, it also excludes revenue management (see
120 H. Meyr
e.g. Talluri and Van Ryzin 2004), where “capacities” are assumed to be perishable
and thus stocks cannot be held at all. A deeper discussion of the relationship between
these various, but similar types of models and their applications in industry would
go beyond the scope of this paper. Instead, the reader who is interested is referred to
Quante et al. (2008).
Demand fulfillment and order promising on the basis of ATP information is one
of the most popular planning tasks (see Kilger and Wetterauer 2008, Table 16.1)
covered by commercial APS. A general overview regarding APS and the role of ATP
therein is given by Fleischmann and Meyr (2003b) and Stadtler and Kilger (2008).
Fleischmann and Meyr (2003a) classify different situations of demand fulfillment with
respect to the three order penetration points MTO, ATO and MTS. They also point
out that—as opposite to MTO and ATO — in MTS situations it often is sufficient
to consider each product separately. Pibernik (2005) also characterizes different ATP
applications and models. He implicitly uses a similar categorization by distinguishing
the operating mode (real time/batch), the availability level of goods and the interaction
with manufacturing planning, where the two latter ones are usually used to characterize
the different order penetration points. ATP software modules of several APS vendors
are presented by Meyr et al. (2008). Dickersbach (2004, Sect. 11) and Knolmayer
et al. (2002, Sect. 3.1.5), however, put a special emphasis on the Global ATP module
of SAP’s advanced planner and optimizer (APO).
The paper of Kilger and Meyr (2008) is basic for the following sections because it
presents the implementation of demand fulfillment in APS in a sufficiently high detail.
Kilger and Meyr (2008) especially describe the simple rules that are usually applied
in APS for both allocation planning (Sect. 9.4) and ATP consumption (Sect. 9.5).
Whereas their argumentation mainly bases on experiences with software of the APS
vendor i2, Dickersbach (2004, Sects. 11.2 and 11.3) shows that a similar approach
has also been favored by SAP/APO. Allocation planning rules, for example, quote
an overall ATP quantity to different customer classes on basis of priority rankings,
with respect to some pre-defined fixed shares or proportional to the original forecasts
of different customers or markets. ATP consumption rules, for instance, allow access
to allocated ATP of an order’s corresponding class or to ATP of classes showing
lower priority. If customers have not been segmented—and thus the above allocation
planning is useless—ATP that has been assigned to other time buckets, to substitute
products or to other locations (e.g. distribution centers or regional warehouses) is
searched for in an user–defined sequence.
Fischer (2001) compares such ATP consumption rules for single order processing
with a linear programming (LP) based batch order processing for a practical case
of the lighting industry and shows advantages of the batch mode. It is interesting to
note that this lighting company originally distinguished eight classes of customers
showing different importance, which have—for sake of simplicity — been reduced to
three by Fischer. In a similar MTS environment Pibernik (2006) compares different
ATP consumption rules for managing the stock outs of a pharmaceutical company.
He suggests to change from a single order to a batch order processing mode only if
shortage is foreseeable. Even though this company also segments their customers into
five priority groups, allocation planning is tested only rudimentarily by Pibernik, using
a “naive” allocation scheme reserving stock for the two most important groups only.
As mentioned above, the APS allocation rules either make no assumptions about
demand (for example priority rankings) or use short–term demand forecasts in a rather
doubtful manner, e.g. by allocating production quantities and ATP proportionally to
the demand forecasts, which has been shown to increase the bullwhip effect within
supply chains (see Lee et al. 1997). Instead, Ball et al. (2004, Chap. 15.4.2) propose an
LP based deterministic allocation model. Basically, it summarizes linear and mixed
integer programming models of hierarchical production planning that are used to
allocate aggregate inventory of product families and/or limited production capacity to
various items within a family. Obviously, this general idea can be transferred to allocate
ATP to different customer classes. Although the model proposed by Ball et al. ought
to be applied in an MTS environment, it rather fits ATO supply chains because it also
decides about raw material and capacity usage. A more convenient MTS application
of this type of models is presented below in Sect. 2.3.
Summing up, modern APS offer the technical prerequisites for ATP allocation and
ATP consumption, thus hoping to gain similar advantages in manufacturing industries
as have been achieved by revenue management principles in airline or hotel indus-
tries. However, they only provide very simple allocation and consumption rules, and
furthermore do not give advices how and when to apply them. Thus, overall benefits
are doubtful. Looking through scientific literature is hardly helpful in this specific
situation because either the model assumptions do not fit (e.g. stochastic inventory
rationing) or the overall performance of both allocation and consumption policies
has not been tested for potential alternatives of customer segmentation (for example,
(Fischer 2001; Pibernik 2005) take the segmentation for granted).
1.2 Contribution and organization of the paper
The basic idea of this paper is to improve demand fulfillment in MTS supply chains
by making use of the heterogeneity of different customers through AP&C order pro-
mising. The fundamental steps are:
• To segment customers with respect to their importance and profitability into several
priority classes,
• to allocate ATP to these classes on basis of a deterministic profit maximization
process taking advantage of short–term demand information, and
• to promise customer orders, i.e. to consume ATP, in real time with respect to these
customer hierarchies.
In order to demonstrate the usefulness, all steps will be executed in a holistic simulation
experiment exploiting practical data of the lighting industry. To our knowledge, such
a comprehensive test, including customer segmentation and allocation, is missing so
far. The aim is to structure the planning tasks concerned with AP&C and to gain
ideas whether and how a preceding allocation process—making use of the short–term
information provided by APS—may be advantageous compared to the traditional
first-come-first-served single order processing.
The next section introduces appropriate LP models for demand fulfillment in MTS
supply chains. Numerical experiments with data of the lighting industry are run in
122 H. Meyr
Sect. 3. A summary of the methodology proposed and of the managerial insights

gained concludes the paper.
2 Model formulations
The following section describes the modeling environment that allows to compare the
different ways of order promising and ATP assignment. LP models for single and batch
order processing without customer segmentation are proposed in Sect. 2.2, whereas
Sect. 2.3 introduces the allocation planning model making use of segmentation. All
models aim at profit maximization. Their outcome can be compared directly with
the optimal profit that would result from a simultaneous ex–post assignment of all
orders arriving within the planning horizon, which is called “global optimization” in
the following.
2.1 Modeling environment
The different order promising alternatives verbally described in the introduction will
now be represented by mathematical models. Figure 1 shows the modeling environment
a-c without customer segmentation: d) with customer segmentation:
once for supply planning once supply planning demand planning

planning (e.g. production
horizon T planning)
ATP forecasts
ATP
once ATP allocation

a-c ATP consumption (“allocation planning”, AP)
(“order promising”)
allocated ATP
(= quotas
order(s) commit- for customer
ment(s) classes k)
real- real-
“customer” ATP consumption (SOPA)
time time
a) GO: once for all orders of the single single

planning horizon T order commitment
b) BOP: several times for all orders
of a batching horizon B<<T real-
c) SOP: in real-time for each single order time “customer”
Fig. 1 Modeling environment for the models “Global Optimization” (GO), “Batch Order Processing”
(BOP) and “Single Order Processing” (SOP) without customer segmentation and “Allocation Planning”
(AP) and “SOP after allocation planning” (SOPA) with customer segmentation
that was chosen to do this. The models (a–c) that do not distinguish customer segments
shall be compared with the AP&C models (d) which put the revenue management idea
of the introductory example into practice by differentiating different customer classes
k, allocating ATP to these customer classes and satisfying customer demand only if
enough allocated ATP of the customer’s corresponding class is available.
For this, the finite, overall planning horizon T is subdivided into discrete time
buckets t = 1, . . . , T . Once, at the beginning of planning (t = 0), on the basis of
supply information—e.g. from the master production schedule or master plan—it is
calculated how much ATP becomes available in each period t (for this calculation see
e.g. Fleischmann and Meyr 2003a,b). Customer orders i arrive one after each other
at different arrival dates ai . For each order i it is known, how much the customer
wants to get (“requested delivery quantity” qi ) and when he wants to get this quantity
(“requested delivery date” di , i.e. the time bucket t, for which the customer requests
his order i to be delivered). The limited availability of ATP necessitates that not all
orders can be served on time. The “order promising” or “ATP consumption” problem
is to decide whether, when and to which degree each order will be served from the
ATP. Not fulfilling an order on time or not filling an order at all will be punished by
penalty costs diminishing the original profit the order would leave. ATP is assumed to
be known deterministically at t = 0, for the whole planning horizon T . Thus it needs
only to be updated when orders are accepted but not because its supply has changed
unexpectedly.
The models (a–c) without customer segmentation differ according to the number of
orders that are gathered before the orders are processed, i.e. assigned to the different
periods’ ATP by means of an LP model maximizing the profit of all incoming orders.
The “Single Order Processing” model SOP processes each order immediately in real
time and thus is trivial to be solved. The “Batch Order Processing” model BOP gathers
all orders arriving within a batching horizon B T . The “Global optimization” model
GO gathers all orders of the whole planning horizon T . Of course, since T is a quite
long time span (e.g. a month) it is not realistic that customers will wait so long until
getting a promise. However, because all orders of the whole planning horizon are
covered and optimized simultaneously, this model can serve as a benchmark to judge
the performance of an iterative application of the other models.
Situation (d) is modeled by a sequence of an “Allocation Planning” model AP that
is executed once at t = 0 and several single order processing models—now denoted as
“Single Order Processing After allocation planning” (SOPA)—which are executed in
real time when each new customer order arrives. The AP model once allocates ATP to
the different, a priori known customer classes k by means of linear programming. For
this, up-to-date forecasts of customer demand within each customer class are necessary.
Like in (c) each single order is processed in real time, but it is only allocated to the
desired delivery date if enough allocated ATP (aATP) of its respective customer class
is available and can be consumed. This corresponds to the revenue management and
inventory rationing idea that some portion of scarce stock should be held back for
more important orders which might arrive later on.
The motivation for this kind of deterministic, mathematical modeling originates
from current practice of APS usage (see e.g. Kilger and Meyr 2008). ATP and demand
forecasts are calculated in APS anyway and can be aggregated for different customer
124 H. Meyr
classes. Also basic allocation and consumption rules are used. Thus, the fundamen-
tal technical framework for its application already exists. Furthermore, LP as a more
sophisticated allocation method could probably easily be implemented because it is
used for mid–term master planning and strategic network design, anyway (Fleisch-
mann and Meyr 2003b).
Of course, very simplifying assumptions are made in this modeling environment as
compared to practice. For example partial delivery of orders is assumed to be possible,
no bargaining about delivery dates is allowed, customer service can only be expressed
in terms of money and uncertainties of demand and supply are excluded. The latter
problem, for instance, could be tackled by introducing a rolling horizon planning on
at least two planning levels: a mid-term (e.g. weekly rolling) level for updating supply
information and executing allocation planning and a short-term (e.g. daily rolling for
BOP or real-time for SOP/SOPA) level for ATP consumption. In this case ATP updates
would be necessary weekly after each update of supply information, but also daily
or for each order (see e.g. Fleischmann and Meyr 2003a, for ATP re-calculation).
AP runs would also be necessary weekly, after each supply and subsequent ATP
update, and would base on the latest forecasts on customer demand. However, a more
detailed discussion of these application issues would go beyond the scope of this
paper because, first, structural insights on the negative impacts of the decomposition
of the GO problem into subsequent BOP, SOP or AP/SOPA models should be gained.
Thus, the restrictive assumptions are necessary to exclude side effects, e.g. due to bad
forecasting of supply and demand. Of course, in a next step, these assumptions should
be weakened (see Sect. 4).
In the following, the situation (a–c) without customer segmentation is described in
more detail by introducing a single, “basic” order promising model that is applied in
different ways to gain the models GO, BOP and SOP.
2.2 Models without customer segmentation
The basic order promising model is a simple network flow problem where the requested
quantities qi —in the following also called “demand”—of certain customer orders
i = 1, . . . , I have to be satisfied by ATP inventories AT Pt that become available
in discrete periods t = 1, . . . , T , e.g. days or weeks. In order to ensure feasibility
even if demand is higher than ATP inventory, a fictitious period T + 1has been
I
introduced being able to serve the surplus demand by setting AT PT +1 := i=1 qi −
T
t=1 AT Pt . The goal is to find the part oit of order i that has to be satisfied by
ATP of period t so that the overall profit is maximized for a given per unit profit
pit . This per unit profit can, for example, be computed by subtracting the per unit
costs ci from the per unit revenues ei of the order i and by punishing the use of ATP
from periods earlier (necessitating storage) or later (backlogging) than the customer’s
requested delivery date di . ATP of the fictitious period T + 1 models non-delivery and
thus cannot generate any profit ( pi, T +1 = 0). It may even cause a loss of goodwill
being punished by negative profits pi, T +1 < 0. Note that costs and revenues of
different customers/orders may vary individually, e.g. due to different transportation
Table 1 Indices, data and variables of the basic order promising model
Indices
s = 1, . . . , S Iterations
i, j = 1, . . . , I Orders
t = 1, . . . , T Periods
t = T +1 Dummy period with “infinite” supply
Is Set of orders that are promised in iteration s
Data
ai Arrival date of order i (i.e. when customer requests a promise)
di Date, the customer requests order i to be delivered
qi Quantity, the customer requests to get delivered by order i [SKU]
ei Per unit revenue of order i [$/SKU]
ci Per unit supply costs of order i (e.g. transportation costs) [$/SKU]
AT Pts Not yet assigned supply that becomes available in period t and [SKU]
can still be promised to customers during iteration s
pit Per unit profit of order i if satisfied by ATP of period t [$/SKU]
= ei − ci
- “low holding costs” if t < di , and
- backlogging costs if di < t ≤ T , respectively
= 0 if t = T + 1
(or -penalty costs for loss of goodwill)
Variables
s ≥0
oit Part of order i which is served by ATP of period t and promised [SKU]
during iteration s (only defined for i ∈ I s )
costs and customized sales prices, which have already been negotiated in the medium
term.
In practice, orders arrive successively with a continuous arrival date/time ai . This
dynamic situation will later on be modeled by a simulation run with successive itera-
tions s = 1, . . . , S. At a certain point in time, i.e. in a certain iteration s, only a limited
subset I s of all orders i = 1, . . . , I is usually known and has not yet been promised,
e.g. a single order in the SOP case or a batch of all orders of a single day in the BOP
case. Thus, the LP formulation of the basic order promising model shown below is
restricted to this subset I s of orders for a given iteration s. For ease of readability,
Table 1 summarizes the indices, data and variables of the LP model. The superscripts
s of the data AT Pt indicate that—after consumption in iteration (s − 1)—the ATP
remaining for iteration s had to be reduced, accordingly. Whereas, the superscripts s
of the variables oit indicate in which iteration s the corresponding order i has been
promised.
Basic order promising model of iteration s:

T +1
maximize pit oits (1)
i∈I s ,t=1
126 H. Meyr
subject to

T +1
oits = qi ∀ i ∈ I s (2)
t=1

oits ≤ ATPst ∀ t = 1, . . . , T (3)
i∈I s
The overall profits of satisfying the orders i ∈ I s from ATP inventory are maximized
by the objective function (1). The requested quantity qi of each order i has to be met
exactly, either by “real” supply of a regular period t = 1, . . . , T or by the “fictitious
supply” modeling non–delivery (2). Constraints (3) ensure that the supply capacity
cannot be exceeded, i.e. that only the still available ATP of period t can be assigned
to yet unpromised orders i ∈ I s .
This basic order promising model will be applied for simulating the three scenarios
(a), (b) and (c) of Fig. 1. With î(s) := argmin i {ai : i ∈ I s } denoting the order i ∈ I s
having the earliest arrival date during iteration s, the following three situations—just
differing by the cardinality |I s | of the subsets I s —can be distinguished:
(a) GO: all orders of the planning horizon T are known in advance and are considered
in a single optimization run, i.e. I s := {1, . . . , I } and S := 1.
(b) BOP: only subsets of orderswithin a “batching
horizon” of B periods areconsi-

dered, i.e. I := i : aî(s) ≤ ai < aî(s) + B and S := T /B with aî(s)
s
denoting the period t the arrival of order aî(s) is assigned to (assuming that T is
an integer multiple of B).
(c) SOP: only a single order is considered during an iteration s, i.e. I s := {î(s)} and
S := I . This is the case for real time due date assignment on an FCFS basis.
Since the degree of freedom decreases, it is expected that the overall objective function
T /B I
values of these models decrease, too, i.e. GO ≥ s=1 BOPs ≥ s=1 SOPs with

a denoting the optimal solution of a model. As already mentioned, GO can serve
as a benchmark (“first best solution”), showing what profit would be optimal if there
were perfect knowledge of customer demand for the whole planning horizon T . The
I T /B
values SOP := s=1 SOPs and BOP := s=1 BOPs are directly comparable to
GO . They show the loss of profit that has to be accepted if, for the sake of customer
service, real time order promising or a short batching horizon B have to be realized.
To compute SOP and BOP in a simulation experiment, the remaining ATP has to
be updated according to ATPs+1 t := ATPst − i∈I s oits ∀t = 1, . . . , T in-between
the iterations s and s +1. This corresponds to the inventory netting and ATP calculation
procedure, more generally described by Fleischmann and Meyr (2003a), for the special
case that supply is assumed to be deterministically known in advance. AT Pt1 can
be initialized by inventory on hand (t = 0) and the projected supply (according
to the master production schedule or master plan of the supply chain) of periods
t = 1, . . . , T . Without loss of generality, AT P0s = 0 ∀s is assumed in the following.
2.3 Models with customer segmentation
The above formulas give rise to the suspicion that BOP can be brought closer to G O
by simply increasing the batching horizon B. This behavior has already been confirmed
by the experiments of Chen et al. (2002, 2001). However, customer expectations of
short order promising response times set a natural limit to an increase of B. Thus
modern APS follow another approach to close the gap to G O while simultaneously
offering the real time single order response times of SOP. As described by Kilger
and Meyr (2008), they adapt ideas of revenue management for industrial purposes:
scarce capacity (in this case ATP) is allocated to certain customer classes with different
priorities (or profits). Incoming customer orders are allowed to consume capacity of
their own or a lower priority class only. By doing this, it shall be prevented that a
lower priority customer order can consume capacity that would later on be needed for
a higher priority order gaining higher profits.
Thus, single order promising can still be applied, but it is preceded by an earlier
allocation (sometimes called “quoting”) process, reserving ATP for distinct priority
classes. It is the aim of this paper to model the planning problems arising in such a
context and to demonstrate and quantify the potential benefits of such a procedure.
Therefore, the ATP allocation and ATP consumption processes of situation (d) in Fig. 1
have been put into the same modeling and simulation environment as GO, BOP and
SOP in a–c of Fig. 1 and the LP models AP and SOPA have been designed to represent
both partial problems: The allocation planning model AP first assigns ATP to a pre-
defined number K of customer (or more generally: priority) classes k = 1, . . . , K . The
subsequent single order consumption SOPA of the class-specific ATP is also modeled
and solved by LP, even if APS usually apply simpler and faster rule-based algorithms
for the ATP consumption. Section 2.4 finally demonstrates how orders can be assigned
to priority classes.
Table 2 shows the indices, data and variables of the AP model. As can be seen,
agreements on how much has to be sold at a minimum (lower bound on sales quantity)
to a respective priority class k in a certain period t and forecasts on how much can at
most be sold (upper bound on sales quantity) are needed in order to quote ATP with
respect to the expected profits of the respective classes. The lower bounds usually
represent strategic sales targets or mid-term commitments which ensure that certain
customer groups get a minimum level of service. The upper bounds are estimates of
the aggregate customer demand of the respective class in a certain period, i.e. forecasts
on what all customers of this class will buy at a maximum. The degree to which the
demand of a certain class should (in terms of overall profits) actually be satisfied will
be determined by the model. Thus, with respect to the limited ATP capacity, the model
further restricts potential sales to certain customer classes by allocating ATP to the
most profitable ones.
In detail, the AP problem can be formalized as follows:
Allocation planning problem (AP):

T +1
T
maximize p̄ktτ · z ktτ (4)
k,t=1 τ =1
128 H. Meyr
Table 2 Indices, data and variables of the allocation planning problem (AP)
Indices
k = 1, . . . , K Priority (or profit) classes of orders/customer groups
(The number of classes K has to be pre–defined in advance).
k Set of orders i belonging to priority class k
Data
min (≥ 0)
dkt Lower bound on sales to priority class k in period t [SKU]
max (≥ d min )
dkt Estimated (maximum) customer demand of class k in period t [SKU]
kt
(= upper bound on sales quantity to priority class k in period t)
p̄ktτ Per unit profit if ATP of period t (= 1, . . . , T + 1) satisfies [$/SKU]
demand of priority class k in period τ (= 1, . . . , T ), e.g.
= Per unit revenue ēk in priority class k
-supply costs c
-“low holding costs” if t < τ , and
-backlogging costs if τ < t ≤ T , respectively
= 0 if t = T + 1
(or -penalty costs for loss of goodwill)
Variables
z ktτ ≥ 0 Part of demand of priority class k in period τ (= 1, . . . , T ) [SKU]
which is satisfied by ATP in period t (= 1, . . . , T + 1)
ft ≥ 0 Still unallocated part of ATP in period t [SKU]
subject to

T +1
min
dkτ ≤ z ktτ ≤ dkτ
max
∀ k, τ = 1, . . . , T (5)
t=1

T
z ktτ + f t = ATP1t ∀ t = 1, . . . , T (6)
k,τ =1
ATP is allocated to the priority classes k so that the overall profit is maximized (4).
The per unit profits p̄ktτ of a class k can, for example, be computed as the average
profits pit of the orders i ∈ k that have been assigned to class k. The totally reserved
ATP has to be within the upper and lower sales bounds of the respective priority
class (5). If, due to the upper bounds dkτ max , ATP cannot be assigned to one of the
classes, it remains unallocated (6) and thus can be used by any class in the later SOPA
consumption.
As already explained in Sect. 2.1, when facing supply and demand uncertainty, AP
should be done on a rolling horizon basis. However, since supply uncertainty should
not matter in the simulation experiments of Sect. 3, AP only needs to be executed
once at the beginning of planning in t = 0. Further, to exclude forecast errors (demand
uncertainty), the aggregate demand forecast dkτ max of class k is initialized with the (later

on) actually requested quantities, i.e. dkτ := i∈k :di =τ qi ∀k, τ with k denoting
max
Table 3 Indices, data and variables of the SOPA problem in iteration s
Indices
classi Priority class order i belongs to
i Set of priority classes which can be consumed by order i
Data (ATP that can be consumed by order î(s) in iteration s)
aATPsktτ ATP that becomes available in period t and has been allocated to orders [SKU]
in priority class k with a requested delivery date in period τ
uATPst ATP that becomes available in period t but has not yet been allocated to [SKU]
any priority class or planned delivery date
Variables
s ≥0
ōkt Part of allocated ATP of priority class k in period t (= 1, . . . , T + 1) [SKU]
which is in iteration s assigned to order î(s) showing a requested delivery
date dî(s)
xts ≥ 0 Part of unallocated ATP of period t, which is in iteration s assigned to [SKU]
order î(s) showing a requested delivery date dî(s)
the priority class which order i belongs to and di denoting the requested delivery
period of order i. For ease of simplicity, the lower bounds on sales are set to zero,
min := 0 ∀k, τ .
i.e. dkτ
of AP allows a very detailed allocation of ATP, not only
The optimal solution z ktτ
specifying the period t, the ATP becomes available, but also specifying which priority
class k it should be reserved for and in which period τ it should be consumed. Let
aATPsktτ denote allocated ATP that has been defined on the same level of granularity
and remains available for consumption in iteration s. Then, the allocated ATP of the
first period after the allocation procedure AP can be defined according to

aATP1ktτ := z ktτ ∀k, t, τ, (7)
thus allowing a very restrictive reservation for important classes. This appears useful
if the forecasts of customer demand are very reliable. Of course, if forecast accuracy
is low, also a more aggregate allocation could be applied, e.g. by

T

aATP1kt := z ktτ ∀k, t. (8)
τ =1
The quantities uATP1t := f t ∀t remain unallocated in case the expected ATP inven-
tories are higher than estimated demand. If, on the other hand, estimated demand is
expected to be higher than total ATP inventories, the portion of demand of period t in

class k that has been allocated to z kt,T +1 > 0 by (5) cannot be served later on.
The LP model (9)–(12) uses these allocated and unallocated ATP quantities as an
input for real time single order processing after allocation planning. The variables of
this SOPA model are explained in Table 3. Since the SOPA models of the subsequent
simulation iterations consider a single order, each, the only order of iteration s is
denoted by î(s) in the following:
130 H. Meyr
“SOP after allocation planning” model of iteration s (SOPAs ):
+1
T
T
maximize s
pî(s),t ōkt + pî(s),t xts (9)
k∈î(s) t=1 t=1
subject to
+1
T
T
s
ōkt + xts = qî(s) (10)
k∈ î(s) t=1 t=1
s
ōkt ≤ aATPsktd ∀ k ∈ î(s) , t = 1, . . . , T (11)
î(s)
xts ≤ uATPst ∀ t = 1, . . . , T (12)
In (9) the
original profits pit of Table 1 are maximized. Thus, the simulation result
SOPA := s=1 S
SOPAs of a preceding AP optimization, followed by S := I ite-
rations of SOPA (with an optimal objective function value SOPAs of iteration s), is
directly comparable to GO , BOP and SOP as computed in Sect. 2.2. The Eqs. (10)
ensure that the requested quantity of order î(s) is either met by (un)allocated ATP
or assigned to the fictitious period T + 1 and thus denied, however generating no
profit or even incurring penalty costs. The capacity constraints (11) and (12) limit
the use of allocated and unallocated ATP to their predefined values. An order î(s)
can only consume ATP in some dedicated classes î(s) . For example, by setting
î(s) := {k : classî(s) ≥ k ≥ K } it can be ensured that an order î(s) ∈ l can
only consume ATP of its own priority class l := classî(s) or other classes k > l sho-
wing lower priorities. Thus, also for the AP problem, it is assumed that the classes
k = 1, . . . , K have been sorted according to decreasing priorities, e.g. defining k > l
if the average profits fulfill

t, i∈k pit t, i∈l pit
≤ . (13)
|k | |l |
Such a strategy of allowing access to lower priority ATP has, for example, been applied
by Fischer (2001)—there called “hierarchical cumulated quoting”—or by Kilger and
Meyr (2008) using customer hierarchies.
Analogously to the SOP procedure described in Sect. 2.2, in the following simula-
tion experiments the (un)allocated ATP remaining after iteration s for use in iteration
s + 1 can easily be calculated by (14) and (15):
s+1
a AT Pktd := a AT Pktd
s
− ōkt
s
∀ k, t = 1, . . . , T, (14)
î(s) î(s)
u AT Pts+1 := u AT Pts − xt ∀ t = 1, . . . , T. (15)
As already mentioned in Sect. 2.1, this is possible because demand and supply are
assumed to be known in advance. Such a data update is more complicated if demand
and supply are uncertain and if AP is executed on a rolling horizon basis. In this case
inventory netting and ATP calculation as described by Fleischmann and Meyr (2003a)
are necessary. Note that in MTS situations late delivery or cancellation of orders is only
possible for newly arriving orders but not for orders that have already been promised
(and thus delivered!). This is opposite to order promising in ATO or MTO situations.
Applying AP/SOPA instead of GO can be seen as a kind of problem decomposition
because the single problem GO has to be decomposed into the two subproblems
allocation planning and SOPA, which have to be solved subsequently and iteratively.
Due to this decomposition, a gap between the GO and SOPA may result, even if
all orders were known with certainty. This gap is generated by aggregating individual
orders to priority classes. However, if demand was known in advance and each order
i was assigned to its own priority class (classi = {i}, K = I ), the final objective
function values GO and SOPA would be identical. Thus, the overall problem is to
find a decomposition that brings the result of AP/SOPA as close as possible to the (in
reality only ex post known) result of GO.
Summarizing these structural insights, the following conclusions can be drawn: In
practice, the result of GO (“first best solution”) cannot be realized because of two
reasons:
• There are demand and supply uncertainties, i.e. orders and supplies cannot be
known in advance. Schneeweiss (2003) denotes a problem decomposition, which
is caused by such a missing information, “time decomposition”.
• For real time order promising, an aggregation of individual orders to priority
classes is necessary. The impacts of this will further be analyzed in Sect. 3.
However, before, the still open problem of determining priority classes has to be
discussed.
2.4 Identification of customer classes
In the above sequence of AP and SOPA an assignment of orders i to priority classes

k was assumed to be predefined, which is expressed by the order sets k and class
indices classi . Usually, such an assignment of orders to classes is not obvious, it may
even be hard to define a useful number K of classes k. This assignment task is a
mid-term planning task because the allocation planning AP has also to be done in the
medium term. It may sound confusing that an order i can be assigned to a class before
it actually arrives at date ai . But usually there are quite stable relationships between
vendors and their customers so that an order can directly be linked to the customer
sending it and thus the problem reduces to assigning customers to priority classes k in
the medium term. For ease of simplicity, the notation will not further be complicated
by distinguishing between customers and their orders. The reader should just keep this
1:n-relationship in mind.
The profits pit as introduced in the above tables usually originate from a time-
independent indicator vali of the “value” of order i (or its corresponding customer)
and a time-dependent, discrete function p·t that punishes non-delivery or earliness and
lateness with respect to di . A piecewise-linear example for such a function, which will
132 H. Meyr
be applied in the following experiments, is given by (16):

help
pit := vali · 1 − (16)
(T − 1) · late
with
⎧
⎪
⎨(di − t) · early if t < di
help := (t − di ) · late if di ≤ t ≤ T
⎪
⎩
(T − 1) · late if t = T + 1
and with penalty costs early for being early and late for being late (usually early <<
late). In the experiments of Sect. 3 early := 1 and late := 10 are used.
One should be aware that usually vali is only an artificial measure describing the
overall importance of order i. Besides the per unit profit ei − ci other non-monetary
factors may contribute to vali as well, for instance, the strategic power of the cus-
tomer ordering i. An example for such a procedure is given by Fischer (2001) and
in Sect. 3.1. Thus, quantifying the measure vali is a crucial task, depending on the
practical application under consideration.
Knowing the vali , for the assignment of a given set of orders to a predefined number
of classes standard clustering methods can be used. They group all such orders i and j
into the same class which are “similar” according to a certain distance measure disti j ,
for example,
disti j := dist ji := vali − val j . (17)
Thereby, “similarity” can be expressed by different types of objectives. For example
Meyr (2007) introduces two alternative clustering models, CS and CM, minimizing
the sum of the distances between any pair of orders within the same class and the sum
of the maximum distances of each class, respectively. To solve the CS problem, he
proposes three alternative local search heuristics basing on steepest descent (called
Sum-DE), threshold accepting (Sum–TA) and tabu search (Sum–TS). For the CM model
a simple rule–based heuristic is applied (called MinMax).
Clustering models, including CS and CM, usually assume that the number of classes
K is known in advance (see Meyr 2007). This was also the case for the AP and SOPA
models of the previous section. Obviously, the optimal objective function values of
CS and CM both will decrease to 0 if K is increased to I . This is because, assuming
complete demand information, in the extreme case K = I the allocation problem
AP reserves the necessary ATP for each single order i, separately. Thus it seems to
be useful to choose the number of classes as large as possible. However, one has to
be aware that increasing the number of classes is not only advantageous. First, also
the complexity of AP and of the clustering problem is increased. Second, and more
crucially, in practice demand information is uncertain. Thus, missing information
about not yet known orders has to be substituted by demand forecasts. Following
the law of large numbers, forecast accuracy is the better, the higher the number of
orders per class is, i.e. the lower the number of classes is. Altogether, a trade off
between better allocation/reservation capabilities and lower forecast accuracy has to
be balanced, which can hardly be formalized. Section 3.5 will give some hints how a
hopefully good compromise can be found.
3 Experiments
The described models are tested with a practical example of the lighting industry. The
case itself and the corresponding data are described in Sect. 3.1. Then, a first overview
of the benefits of allocation planning is given. Different ways of defining the ATP
search space î(s) and ATP consumption rules are discussed in Sect. 3.3. The effects
of varying K are tested in Sect. 3.4. The final subsection of Sect. 3 evaluates the overall
impact of clustering on the finally decisive SOPA outcome.
The allocation and ATP assignment problems GO, BOP, SOP, AP and SOPA can
all be interpreted as classical transportation problems. Thus, standard LP software or
specialized network flow solvers (see e.g. Ahuja et al. 1993) can be applied without
any problems. SOP and SOPA show an even simpler structure because of considering
a single order i only. Thus, they can be solved to optimality with fast backward and
forward-oriented, rule-based algorithms, which start in period di and class classi and
proceed in sequence of descending per unit profits. Similar real time ATP search rules
are usually implemented in APS (as heuristics for more complicated variants of SOP
and SOPA). However, for ease of simulation in the following experiments, which have
been coded with Microsoft Visual C++ 6.0, the standard linear programming solver
CPLEX 9.0, the modeling language ILOG OPL Studio 3.7 and its C++ component
libraries interface (ILOG 2007) have been used for all ATP models, including the
simpler SOP and SOPA problems. The computational tests have been executed on a
personal computer with an Intel Pentium M 1.3 GHz processor and 512 MB RAM,
operated by the Microsoft Windows XP Professional system.
3.1 Problem data
The experiments of the following sections use practical data that have been introduced
by Fischer (2001) in a case study of lighting production. This business is a classical
MTS-environment where customer orders arrive at the distribution centers and have to
be served from the stock which is already available or at least projected to arrive soon.
Six different problems, denoted as P1, …, P6 in the following, have been considered
by Fischer. These problems reflect the demand for six different final items—also called
P1, …, P6 in the following—during one month, i.e. a period of T = 30 days. Note,
even if 30 days are simulated by Fischer and in the experiments of Sects. 3.2–3.5,
orders usually arrive between day 1 and day 26. The only exceptions are P3, where
the last order arrives at day 23, and P5, where the first order arrives at day 6.
The characteristics of the problems P1,…, P6 are shown in Table 4. The four
problems P1, P2, P3 and P5, with less than 40 orders arriving, are rather small. Due to
the infrequent arrival of orders and the resulting low average number of orders per day
between 0.9 and 2.1, a BOP-horizon of a single day is expected to show only weak
impacts. This might be different for the two larger problems P4 and P6 with 1,305 and
509 orders, respectively, and with 72.5 or 28.3 orders per day, on the average.
134 H. Meyr
Table 4 Data used by Fischer (2001)
P1 P2 P3 P4 P5 P6
Total no. of orders 37 29 25 1305 17 509

Orders per class 8/14/15 29// 24//1 725/440/140 //17 500/7/2
No. of supplies 19 13 12 19 19 19
Lost sales (percent) 11.3 17.9 17.3 19.9 4.5 22.6
Orders per day 2.1 1.6 1.4 72.5 0.9 28.3
Aver. distance disti j 4.7 0.7 0.0 4.3 0.0 0.9
The number of supplies, i.e. refillings of ATP inventory, within the planning horizon
varies between 12 and 18. The four products P1, P4, P5 and P6 start with positive initial
inventory that has been modeled as an additional 19th ATP supply at day t = 0 (see
row “no. of supplies” in Table 4).
The total order quantity within the planning horizon exceeds the total supply signi-
ficantly by 4.5–22.6%. The respective shares have been denoted as “lost sales” in
Table 4. This indicates that there are indeed shortage situations in this kind of busi-
ness. However, usually not all of these sales are really “lost” because some of the
orders might be satisfied by supply arriving after the planning horizon of T = 30
days. Nevertheless, it seems that the customer service level has been poor for these
six products.
Table 4 also shows the average distance disti j between all pairs of orders i and
j for a certain product. Note that this is not the original distance measure used by
Fischer. Fischer used up to three priority classes, as indicated in the row “orders per
class” of Table 4, to differentiate customers/orders showing various importance when
computing order–specific costs. The original data have been normalized in order to
allow the application of general clustering models, like CS and CM, also for K = 3.
The distances disti j have been calculated as follows:
Two major attributes contribute to the value indicator vali of a certain customer
order i:
• The normalized per unit profit pr o f itinor m of order i has been calculated by means
of
( pr o f iti − pr o f it min )
pr o f itinor m :=
pr o f it max − pr o f it min
with pr o f iti := ei − ci denoting the per unit profit of order i and pr o f it min :=
mini { pr o f iti } and pr o f it max := maxi { pr o f iti } denoting the minimum and
maximum profit of any order i. The resulting normalized profits are in a range
0 ≤ pr o f itinor m ≤ 1.
• According to the varying importance of different customers, Fischer assigned all
customers and their respective orders to the three priority groups mentioned above.
Therefore, each customer order has a priority index priorit yi ∈ {1; 2; 3}. These
priority indices have also been normalized to a range between 0 and 1 by using
( priorit yi − 1)
priorit yinor m := .
3−1
Both attributes have been aggregated into the single value indicator vali of order i by
weighing them with weights w1 and w2 according to
vali := w1 · pr o f itinor m + w2 · priorit yinor m . (18)
For the experiments in the following sections identical weights w1 := w2 := 10 have

been used, resulting in an indicator range 0 ≤ vali ≤ 20.
The profits pit of GO, SOP, BOP, and SOPA and the distance measure disti j of CS
and CM (see Meyr 2007) have then finally been calculated by (16) and (17) with penalty
costs early := 1 for being early and late := 10 for being late. Since early ≤ late
and |t − di | ≤ (T − 1), the profits pit also range between 0 and 20. Note that the
customers’ and orders’ priorities of the problems P2, P3, P5 and P6 seem to be quite
similar because their average distance is small. In P2 and P5 all orders even have the
same priority values priorit yinor m , but the average distance disti j = 0.7 of P2 is
caused by varying per unit profits. However, P3 contains a single order with lower
priority, but the higher profit pr o f itinor m of this order causes all value indicators vali
of P3 to be equal. Altogether, no real advantage of the allocation process underlying
SOPA can be expected for P3 and P5.
3.2 Benefits of allocation planning
Using the notation of Sects. 2.2 and 2.3, GO , SOP , BOP and SOPA denote the
overall objective function value of a complete, raw-data driven simulation run over
T time periods. The SOPA run is preceded by the allocation planning problem AP
as described in Sect. 2.3 and uses the original priority classes of Fischer.
Table 5 shows the percentage deterioration of SOP , BOP and SOPA as com-

pared to GO , e.g. GO −SOP · 100. It can be interpreted as the percentage profit
GO
loss of a short–range order acceptance compared to the ex–post optimal solution. The
BOP results are varied over a batching horizon of B = 1, . . . , 5 days (and T mod B
for the last periods, respectively). SOPA results are shown in two different variants:
SOPA a aggregates allocated ATP according to (8). SOPA d uses disaggregate aATP
as defined by (7), thus also allowing a reservation of ATP becoming available in
period t for use in another period τ = t. Therefore, S O P A a demonstrates the
“pure” effect of allocating ATP to the three customer classes pre–defined by Fischer
(2001), whereas S O P A d combines this effect with an additional “temporal” reser-
vation of ATP quantities for the periods of their expected use, thus assuming a high
forecast quality. The computation times of a single run are negligible, e.g. solving
GO for the biggest problem P4 takes just a few seconds. However, since only the
standard C++ libraries and data conversion routines of the LP software OPL (ILOG
2007) are used, a complete SOP- or SOPA-simulation run of P4 may last several
hours.
136 H. Meyr
Table 5 Percentage profit loss of S O P , B O P (B = 1, . . . , 5) and S O P A (disaggregate, aggregate)

as compared to G O
P1 P2 P3 P4 P5 P6 Average
SOP* 15.0 21.1 0.0 12.4 1.2 6.9 9.4

BOP*1 14.8 21.0 0.0 11.6 1.2 6.8 9.2
BOP*2 14.1 21.0 0.0 11.0 1.2 6.8 9.0
BOP*3 14.1 13.3 0.0 10.7 0.6 6.7 7.6
BOP*4 13.6 20.8 0.0 6.3 1.2 6.7 8.1
BOP*5 12.1 11.2 0.0 10.0 0.3 6.7 6.7
SOPA*a 5.5 21.1 0.0 1.5 1.2 0.3 4.9
SOPA*d 0.5 0.1 0.0 0.2 0.0 0.3 0.2
As expected, batching orders and increasing the batching horizon B is advantageous

when compared to the FCFS single order processing SOP. But even for a batch horizon
of a whole week, the overall improvement is rather disappointing. Astonishingly, this
holds especially true for the problems P4 and P6, which show a high degree of freedom
because of their large number of orders per day. Of course, a simulation horizon of 30
days is actually too short for such experiments. However, studying Table 5 it seems
likely that increasing the simulation horizon would stabilize the results, but not really
change the overall picture.
SOPA a shows a significant improvement for problems P1, P4 and P6. The diffe-
rence to S O P is only caused by the allocation planning on basis of the three priority
classes used by Fischer (see Sect. 3.1, Table 4). These results can further be improved
by S O P A d, which allows a temporal reservation of ATP, too. In this case, near–
optimal profits can be gained for all six scenarios. Thus, if companies are able to
realize a high forecasting accuracy, defining disaggregate ATP seems reasonable.
SOP solves P3 and P5 almost to optimality because of their corresponding custo-
mers’ homogeneity and the distances disti j = 0 between every two orders i and j. The
profit loss of 1.2% for P5 is caused by inventory or backlogging costs as a consequence
of an unfavorable temporal assignment of ATP and can thus additionally be avoided
by S O P A d. As opposite to P3 and P5, P2 not only shows a small number of orders,
but also a non–zero heterogeneity. This might be the reason for the exorbitant advan-
tage of temporal reservation for P2. On the other hand, temporal reservation seems to
have no impact on P6 (0.3 for both SOPA a and SOPA d). Summing up both SOPA
variants clearly profit from clustering effects.
3.3 Variation of the ATP search space and consumption rules
Both SOPA variants of the last section assumed that ATP can only be consumed in the
priority class classî(s) , the order î(s) belongs to. The subsequent experiments allow
a more flexible consumption of ATP by varying the ATP search space î(s) in the
following way:
cc ⇔ ATP can only be consumed in the class classî(s) , order î(s) has been assigned
to, i.e. î(s) := {classî(s) } (as done in Sect. 3.2).
cK ⇔ ATP can be consumed in the order’s original class or in classes with lower
priority, i.e. î(s) := {classî(s) , . . . , K } (see Sect. 2.3 regarding the sorting of
classes).
1K ⇔ ATP can be consumed in all classes, i.e. î(s) := {1, . . . , K }.
1c ⇔ ATP can be consumed in the order’s original class or in classes with higher
priority, i.e. î(s) := {1, . . . , classî(s) }.
Intuitively, the last variant does not seem to make much sense, but has been imple-
mented for ease of validation and comparison.
Note that the constraints above do not specify a sequence for searching this space. By
solving the SOPA model (9) – (12) using linear programming, ATP can be consumed
freely within the search space î(s) because—for a given period t—the order’s original
profit pî(s),t remains the same independently of the class, the ATP quantities actually
come from. In order to guide the search through various priority classes in an intended
manner (while still applying LP methods), fictitious gains and losses have been defined
the following way: The objective function (9) is extended by

(K − k) · 0.01 · pî(s),t ōkt
s
(19)
t,k≥classî(s)
for the search space cK and by (19) plus

(k − classî(s) ) · 0.01 · pî(s),t ōkt
s
(20)
t,k<classî(s)
for the search space 1K. Thus, the allowed classes are searched in an order of descen-
ding priorities first, starting with the original class classî(s) . If no such ATP has been
found for a search space 1K, higher priority classes are then searched in a sequence
of ascending priorities, starting with classî(s) − 1. Of course, the loss of profit shown
in Table 6 has been calculated on basis of the regular profits (9), only. This way, ATP
search rules, as proposed by Kilger and Meyr (2008) and Fischer (2001) and used in
most APS, can also be simulated within the LP framework of this paper.
Table 6 shows the percentage profit losses for a variation of aATP aggregation
(aggregate, disaggregate), of the search space (cc, cK, 1K, 1c) and of the search
sequence (free allocation, search sequence predefined). The two rows marked in bold
correspond to the respective SOPA results of Table 5.
When comparing the four a/·/f scenarios among themselves, the best results are
achieved for the cc search space, i.e. when staying within an order’s original priority
class. Access to lower class ATP is only reasonable if search rules are used (a/cK/s).
In this case, the a/cc/f results can be equalized but not improved. Free access to higher
priority ATP (a/1K/· and a/1c/f) is indeed proven to be nonsense. A variation of the
search space or the introduction of search rules (a/·/·) do not show any effects on P2 and
P5. For these products a profit increase can only be achieved by temporal reservation
(d/·/·). The situation is actually the same for P3. Its anomalies for a/cK/· only occur
138 H. Meyr
Table 6 Percentage profit loss of SOPA as compared to GO for varying temporal reservation (aggregate,
disaggregate), ATP search space (cc = original class, cK = lower priority, 1K = all classes, 1c = higher
priority) and ATP search rules (f = free allocation, s = search sequence predefined)
P1 P2 P3 P4 P5 P6 Averagea
a/cc/f 5.5 21.1 0.0 1.5 1.2 0.3 5.9

a/cK/f 5.5 21.1 (24.4) 4.3 1.2 0.3 6.5
a/1K/f 15.0 21.1 0.0 12.5 1.2 6.9 11.3
a/1c/f 18.0 21.1 0.0 18.9 1.2 6.9 13.2
a/cK/s 5.5 21.1 (17.7) 1.5 1.2 0.3 5.9
a/1K/s 15.0 21.1 0.0 12.4 1.2 6.9 11.3
d/cc/f 0.5 0.1 0.0 0.2 0.0 0.3 0.2
d/cK/f 0.5 0.1 0.0 0.2 0.0 0.3 0.2
d/1K/f 0.6 0.1 0.0 6.8 0.0 0.8 1.7
d/1c/f 12.8 0.1 0.0 13.8 0.0 0.8 5.5
d/cK/s 0.5 0.1 0.0 0.2 0.0 0.3 0.2
d/1K/s 0.6 0.1 0.0 6.8 0.0 0.8 1.7
a Without P3
because the penalty holding costs “early = 1” have turned out to be too low for this
product. Thus, the average values in the corresponding column of Table 6 have been
calculated without considering P3.
The comparison of the a/·/· with their respective d/·/· scenarios emphasizes the
advantages of a temporal reservation, again. Altogether the picture is similar for the
disaggregate scenarios. The search spaces cc and cK show equal quality, whereas 1K
and 1c compare badly. A positive effect of search rules cannot be recognized, here.
3.4 Variation of the number of classes
All S O P A results presented so far are based on Fischer’s original assignment of

customers to three priority classes (see Sect. 3.1). It will now be investigated whether
a variation of the number of priority classes might be advantageous. At the same
time the various ATP consumption alternatives will be compared again. The following
experiments will be limited to P1. Product P1 has been chosen because
• it comprises only 37 orders and thus can be simulated in short computation times,
• Fischer’s assignment of orders to classes showed balanced proportions for P1 (8/14/
15, see Table 4), and because
• the SOPA allocation achieved significant and non–identical profit increases for both
variants—those with (0.5% loss) and those without (5.5%) temporal reservation—
as compared to the standard SOP (15%) procedure (see Table 5).
Up to 20 priority classes have been generated using the clustering models and heuristics
of Meyr (2007). Table 7 shows the average of the corresponding percentage SOPA
profit losses.
Table 7 Percentage profit loss of SOPA as compared to G O for P1 with respect to different ATP
consumption rules (see Table 6) and a varying number of priority classes K (missing entry = 0.0)
Reser.: Aggregate Disaggregate
Space: cc cK 1K 1c cK 1K cc cK 1K 1c cK 1K
Search: Free Sequ. Free Sequ.
K=1 15.0 15.0 15.0 15.0 15.0 15.0 0.9 0.9 0.9 0.9 0.9 0.9
2 10.1 10.1 15.0 15.1 10.1 15.0 0.6 0.6 0.6 1.6 0.6 0.6
3 6.4 6.7 15.0 17.2 6.4 15.0 0.5 0.5 0.5 10.5 0.5 0.5
4 5.7 5.7 15.0 18.0 5.7 15.0 0.4 0.4 0.5 10.6 0.4 0.5
5 4.5 5.2 15.0 20.8 4.5 15.0 0.4 0.4 0.5 10.6 0.4 0.5
6 3.6 4.3 15.0 23.8 3.6 15.0 0.1 12.0 0.1
7 2.7 3.2 15.0 24.7 2.7 15.0 0.2 0.2 0.3 12.1 0.2 0.3
8 2.9 3.5 15.0 26.1 2.9 15.0 0.1 12.0 0.1
9 2.4 2.7 15.0 26.7 2.4 14.9 0.1 12.0 0.1
10 2.4 2.7 15.0 26.2 2.4 14.9 0.1 12.0 0.1
11 2.4 2.7 15.0 26.2 2.4 14.9 0.1 12.0 0.1
12 2.4 2.6 15.0 26.6 2.4 14.9 0.1 12.0 0.1
13 2.4 2.6 15.0 26.6 2.4 14.9 0.1 12.0 0.1
14 2.5 2.6 15.0 26.7 2.5 14.9 0.1 12.0 0.1
15 2.8 3.0 15.0 26.8 2.8 14.9 0.1 12.2 0.1
16 3.0 3.2 15.0 26.7 3.0 14.9 0.1 11.9
17 3.0 3.3 15.0 26.8 3.0 14.8 0.1 12.1
18 2.2 2.7 15.0 26.8 2.2 14.8 0.1 12.1
19 2.2 2.7 15.0 26.9 2.2 14.8 0.1 12.3
20 2.2 2.9 15.0 26.8 2.2 14.7 0.1 12.1
Average 4.05 4.38 15.0 24.0 4.05 14.9 0.16 0.16 0.24 10.7 0.16 0.20
Fischer 5.5 5.5 15.0 18.0 5.5 15.0 0.5 0.5 0.6 12.8 0.5 0.6
The row “average” of Table 7, containing average results of all 20 classes for each
ATP search alternative, confirms the findings of the last section. Within the aggregate
aATP scenarios (left part of Table 7) the search spaces cc and cK perform best again,
also for a varying number of classes K . If access to lower priority classes is allowed
(a/cK/·), sequential search rules should be applied instead of a free ATP consumption.
The results of the disaggregate aATP (right part of Table 7) show a similar structure.
However, the overall solution quality is better. Due to the limited degree of freedom
left after the temporal reservation, the d/1K/· scenarios also behave well. All in all,
the a/cK/s–rules for ATP consumption, as proposed by most APS, seem justified by
these experiments. However, simply staying within the original class (a/cc/·) would
perform equally.
The number of classes K appears more important than the search space and search
rule. This can be seen when studying the profit improvement resulting from increasing
K for all ·/cc/· and ·/cK/· scenarios. The absurdity of an 1c search space becomes
140 H. Meyr
particularly clear in Table 7 where the profit loss even increases for a higher number
of customer classes. The row K = 1 shows the results for a single class only, i.e. the
SOP performance without allocation planning. The a/·/· values coincide with the SOP
value of P1 in Table 5. The d/·/· values for K = 1 illustrate the improvement possible
by solely introducing temporal reservation, without additionally building customer
classes. A profit loss of 0.9% still remains because all orders of the same period are
considered as being equal. However, for P1 this affects only 2.1 orders on the average
(see Table 4).
The two lines marked in italics allow a comparison of the clustering methods (K = 3)
of Meyr (2007) with the original customer segmentation of Fischer. There seems
to be a small advantage for the automatic methods. Nevertheless, in general both
segmentations lead to similar results.
Note that the results are based on a single product only and thus can hardly be gene-
ralized. Nevertheless, the example shows that profit can be increased by introducing
priority classes. Even if there is no obvious, natural customer segmentation, a cluste-
ring into several price classes is valuable, as long as different customer orders show
various per unit profits. Thus, it seems more important whether a clustering is done
than how it is done. To what extent this assumption is true will be further investigated
in the next section.
3.5 Effects of clustering on SOPA
Table 8 shows the percentage profit loss of single order processing after allocation
planning, as compared to the global optimization result G O , for each of the clustering
alternatives MinMax, Sum-DE, Sum-TA and Sum-TS of Meyr (2007), individually (see
Sect. 2.4). The results are presented for the products P1, P4 and P6 comprising the
largest number of orders (37, 1,305 and 509, respectively) and showing the largest
inhomogeneity of distances disti j (see Table 3). For ease of clarity, the simulation
has been restricted to the single d/cK/s scenario, one of the best-performing scenarios
of Sect. 3.4. Missing entries in the table indicate a profit loss of 0.00, i.e. that G O
has been reached. Note that the MinMax results and the results of the CS heuristics
Sum-DE, Sum-TA and Sum-TS would not have been directly comparable because they
solve the two different problems CM and CS. However, each product’s profit losses of
Table 8 can immediately be compared with each other, since the clustering heuristics
influence SOPA only indirectly by the different ways of cluster building.
Looking at row “aver.”, containing the results averaged over all 20 classes, gives
a quick overview of the overall performance of the four heuristics. However, results
appear nonuniform. While P1 and P6 are dominated by MinMax, the CS heuristics out-
perform the CM algorithm clearly for P4. Thus there does not seem to be a significant
correlation between the clustering objectives, the solution quality of different heuristics
and the profits generated by the respective clusters.
Interestingly, the profit losses of Sum-DE (for P6) and Sum-TA (for P4 and P6)
decrease first, but then increase again. A reason for this might be found in a bad overall
solution quality of the CS heuristics, particularly for large problems with many orders
and classes (Meyr 2007). This is, besides forecast accuracy, a second argument for
choosing a not too large class number K .
Table 8 Percentage profit loss of the SOPA clustering alternatives MinMax (MM), Sum-DE, Sum-TA and
Sum-TS as compared to GO for P1, P4 and P6 in the d/cK/s scenario (missing entry = 0.00)
K P1 P4 P6
MM DE TA TS MM DE TA TS MM DE TA TS
1 0.89 0.89 0.89 0.89 11.03 11.03 11.03 11.03 11.39 11.39 11.39 11.39
2 0.49 0.66 0.66 0.66 11.00 10.12 10.12 10.12 0.30 4.13 4.13 4.13
3 0.25 0.52 0.52 0.52 10.10 3.50 3.45 3.50 0.30 2.71 2.71 2.71
4 0.25 0.50 0.50 0.50 0.15 0.15 0.15 0.15 0.29 2.72 2.72 2.72
5 0.20 0.52 0.44 0.50 0.15 0.15 0.15 0.15 0.29 1.77 1.77 1.77
6 0.09 0.03 0.03 0.03 0.15 0.34 0.10 0.34 0.29
7 0.03 0.03 0.03 0.69 0.15 0.10 0.10 0.10 0.29
8 0.03 0.03 0.03 0.03 0.15 0.10 0.07 0.10 0.29
9 0.03 0.03 0.03 0.03 0.15 0.07 0.07 0.07 0.28
10 0.03 0.03 0.03 0.03 0.15 0.07 0.07 0.07 0.03
11 0.03 0.03 0.03 0.03 0.13 0.07 0.04 0.07 0.03
12 0.03 0.03 0.03 0.03 0.12 0.08 0.03 0.08 0.01 0.39
13 0.03 0.03 0.03 0.03 0.12 0.04 0.03 0.04 1.64
14 0.03 0.03 0.03 0.12 0.03 0.03 0.03 1.35
15 0.03 0.14 0.03 0.03 0.03 2.65
16 0.08 0.03 0.03 0.03 0.97 0.67
17 0.08 0.03 0.03 0.03
18 0.08 0.03 0.02 0.03 1.39 0.49
19 0.08 0.03 0.06 0.03 0.91
20 0.06 0.03 0.06 0.03 0.58 1.96
Aver. 0.12 0.17 0.16 0.20 1.71 1.30 1.28 1.30 0.69 1.58 1.34 1.14
The clustering of Fischer often shows better results (0.51 for P1, 0.21 for P4 and
0.33 for P6) than the automatic clustering methods for K = 3. However, for K = 4
already the MinMax clustering outperforms Fischer’s profits for all three products.
Starting with K = 7 the same holds true for all CS heuristics as well. On the whole,
all four heuristics show promising results when four or more classes are used.
Summing up this section, SOPA indeed seems not to be very sensitive with respect
to the clustering method used. An increase of the number of classes K leads to higher
profits if orders of the same product are inhomogeneous enough. Considering the
examples of this section at least 4, but better 6–7 classes should be used. However, the
number of classes should not be chosen too large in order to reduce forecasting errors
and a bad performance of clustering heuristics, especially for CS.
4 Summary, managerial insights and outlook
The exemplary tests of the paper have shown that a first-come–first-served processing
of arriving customer orders is hardly the best way of demand fulfillment in shortage
situations if reliable forecasts are available. Gathering data for a certain period of time
142 H. Meyr
and processing them in a batch can improve the situation. However, often customer
service sets a natural limit to such a procedure because customers increasingly expect
short order confirmation lead times. Another way of improvement can be to precede
the FCFS single order processing by a further allocation planning step. Here, priority
classes for customer orders are built, available inventory (ATP quantities) is “allocated”
to these classes and reserved for later consumption by their respective customers. Such
a customer segmentation has proven its potentials when introducing booking classes in
airline yield management. Thus, the basic idea is not new and has also been supported
by advanced planning systems where simple ATP allocation and consumption rules are
offered. However, until now it was largely unclear—in science and practice—whether,
when, why and to what extent such a proceeding might be useful in manufacturing
industries, too.
First answers to these questions have been given using an example from the lighting
industry where bulbs, fluorescent lamps etc. are made to stock on the basis of forecasts,
first, and then sold from stock as soon as customer orders arrive. In order to demonstrate
its potentials the following planning tasks had to be structured, discussed and solved
first:
1. Determination of a reasonable number of priority classes,
2. clustering, i.e. assignment of customers and customer orders, respectively, to these
classes,
3. allocation planning, i.e. allocation of available inventory on hand and planned
production quantities (ATP) to the priority classes, and
4. ATP search, i.e. successively consuming this allocated ATP for each incoming
order. In this case both the search space (classes allowed) and the search sequence
have to be specified.
(1) has been tackled by means of simulation by varying the number of classes in a
reasonable range and (2) by applying standard clustering methods. For (3) and (4)
linear programming models have been proposed and solved to optimality. All in all, it
was not intended to discuss each of these planning tasks in all detail and to solve it in
the best possible manner (even though this has not satisfactorily been done in science
up to now). The primary goal was to bring all four tasks together in a single simulation
experiment to give an impression of the overall potential of allocation planning in
make–to–stock industries of this or similar types.
Since practical data have been used and the test bed was limited one has to be aware
that the results are only exemplary and more general statements would need further
experiments. Nevertheless, some interesting insights have been gained by the lighting
example and also common views have been confirmed: Introduction of priority classes
and allocation planning can indeed increase revenues and profits, substantially. The
more heterogeneous the customers and their orders, e.g. with respect to the revenues
made or to the strategic importance of the customers, the higher the advantages are.
The number of customer classes plays an important role. Too few classes cause a
loss of profits, too many classes make forecasting and clustering difficult. A temporal
reservation of stocks, for use in a specific period, would generally be advantageous,
but its practical application is only reasonable if customer demand can be forecast
reliably enough.
At least in the lighting case, ATP consumption policies and the clustering method
itself are not as crucial as the choice of an appropriate number of classes. Although
LP methods have been applied for ATP consumption, simple ATP search rules would
perform equally for this example. Such rules should either stay within an order’s
original class or, as often claimed in Revenue Management and by APS, also allow
access to lower priority classes. In the latter case, lower classes should better be
searched for in order of descending priorities. However, note that LP methods or more
sophisticated rules are required in more complex supply chains, e.g. in make-to-stock
supply chains with several stocking points and/or product substitution or in assemble-
to-order supply chains with multi–stage bills of materials.
Thus, also in manufacturing industries managers should pay additional attention to
their customers’ varying nature and try to increase their overall customer service by
allocating their scarce resources—in make–to–stock environments more specifically:
their limited finished item stock—with higher priority to their more important cus-
tomers. As the example of the lighting industry has shown, for this, not even active
or for the customer visible measures of customer segmentation (like fencing strate-
gies or longer response times for order promises) are necessary. It is sufficient to take
advantage of the already existing customer heterogeneity by applying standard clus-
tering methods for identifying priority classes and by introducing well-coordinated
ATP allocation and consumption processes.
Of course, there are still a lot of research challenges. Each of the planning tasks
introduced above should be investigated in more detail for prerequisites of application
and fitting solution methods. First of all, the sensitivity of the results with respect to
less reliable supply information, e.g. concerning the viability of production plans, and
demand information, i.e. to lower forecast accuracy, has to be tested. Furthermore,
similar simulation experiments should be executed for more complex types of supply
chains with other order penetration points. On the one hand, APS support allocation
planning and ATP consumption in resource- and capacity-constrained manufacturing
industries by offering the deterministic rules mentioned above. On the other hand,
there is an obvious affinity to inventory rationing for several customer classes and
to quantity-based revenue management, as defined by Talluri and Van Ryzin (2004)
and practiced in many service industries like airline, hotel or car rental. Most of their
methods are of a stochastic nature. Thus the most challenging prospect for future
research is to find out whether and how these worlds can learn from each other.
Acknowledgments The author is grateful to Markus E. Fischer and Bernhard Fleischmann for providing
the data, Matthias Mann for supporting the experiments and the Wiener Wissenschafts-, Forschungs- und
Technologiefonds (WWTF) for funding the research.
References
Ahuja R, Magnanti T, Orlin J (1993) Network flows: theory, algorithms and applications. Prentice Hall,
Englewood Cliffs
Ball M, Chen C-Y, Zhao Z-Y (2004) Available to promise. In: Simchi-Levi D, Wu S, Shen Z-J (eds)
Handbook of quantitative supply chain management: modelling in the eBusiness era. Kluwer, Boston
pp 447–482
144 H. Meyr
Chen C-Y, Zhao Z-Y, Ball M (2001) Quantity and due date quoting available to promise. Inform Syst
Frontiers 3(4):477–488
Chen C-Y, Zhao Z-Y, Ball M (2002) A model for batch advanced available–to–promise. Product Oper
Manage 11(4):424–440
de Vericourt F, Karaesmen F, Dallery Y (2002) Optimal stock allocation for a capacitated supply system.
Manage Sci 48(11):1486–1501
Dickersbach J (2004) Supply chain management with APO. Springer, Berlin
Fischer M (2001) Available-to-Promise: Aufgaben und Verfahren im Rahmen des Supply Chain Manage-
ment. S. Roderer Verlag, Regensburg
Fleischmann B, Meyr H (2003a) Customer orientation in Advanced Planning Systems. In: Dyckhoff H,
Lackes R, Reese J (eds) Supply Chain Management and Reverse Logistics. Springer, Berlin pp 297–
321
Fleischmann B, Meyr H (2003b) Planning hierarchy, modeling and Advanced Planning Systems. In: de
Kok A, Graves S (eds) Supply chain management: design, coordination, operation, Vol 11 of handbooks
in operations research and management science. Elsevier, Amsterdam pp 457–523
ILOG (2007) ILOG OPL development studio. URL: http://www.ilog.com/products/oplstudio/
Keskinocak P, Tayur S (2004) Due-date management policies. In: Simchi-Levi D, Wu SD, Shen Z-J (eds)
Handbook of quantitative supply chain Analysis – Modeling in the E-Business era. Chap 12. Kluwer,
Dordrecht pp 485–547
Kilger C, Meyr H (2008) Demand fulfilment and ATP. In: Stadtler H, Kilger C (eds) Supply Chain Mana-
gement and Advanced Planning 4th edn, Chap 9. Springer, Berlin pp 181–198
Kilger C, Wetterauer U (2008) The selection process. In: Stadtler H, Kilger C (eds) Supply Chain Mana-
gement and Advanced Planning 4th edn, Chap 16. Springer, Berlin pp 309–323
Knolmayer G, Mertens P, Zeier A (2002) Supply Chain Management based on SAP Systems. Springer,
Berlin
Lee H, Padmanabhan V, Whang S (1997) The bullwhip effect in supply chains. Sloan Manage Rev Spring
1997:93–102
Meyr H (2007) Clustering methods for rationing limited resources. Schriften zur Quantitativen Betriebswirt-
schaftslehre 4/2007, Technische Universitaet Darmstadt, Hochschulstr. 1, 64289 Darmstadt, Germany
Meyr H, Rosič H, Seipl C, Wagner M, Wetterauer U (2008) Architecture of selected APS. In: Stadtler H,
Kilger C (eds) Supply chain management and Advanced Planning 4th edn, Chap 18. Springer, Berlin
pp 349–366
Meyr H, Wagner M, Rohde J (2008) Structure of advanced planning systems. In: Stadtler H, Kilger C (eds)
Supply chain management and advanced planning 4th edn, Chap 5. Springer, Berlin pp 109–115
Pibernik R (2005) Advanced available–to–promise: classification, selected methods and requirements for
operations and inventory management. Int J Product Econ 93(94):239–252
Pibernik R (2006) Managing stock–outs effectively with order fulfilment systems. J Manufact Technol
Manage 17(6):721–736
Quante R, Meyr H, Fleischmann M (2008) Revenue management and demand fulfilment: Matching applica-
tions, models, and software. Technical report, Institute for Production Management, Vienna University
of Economics and Business Administration, Nordbergstr. 15, 1090 Wien, Austria (OR Spectrum to
appear)
Schneeweiss C (2003) Distributed decision making, 2nd edn. Springer, Berlin
Stadtler H, Kilger C (2008) Eds. Supply Chain Management and Advanced Planning, 4th edn. Springer,
Berlin
Talluri K, Van Ryzin G (2004) The theory and practice of revenue management. Kluwer, Boston
Managing product availability in an assemble-to-order
supply chain with multiple customer segments
Thomas R. Ervolina · Markus Ettl ·

Young M. Lee · Daniel J. Peters

OR Spectrum (2009) 31:257–280
DOI 10.1007/s00291-007-0113-4
Abstract In this article, we propose a novel availability management process called

Available-to-Sell (ATS) that incorporates demand shaping and profitable demand
response to drive better supply chain efficiency. The proposed process aims at finding
marketable product alternatives in a quest to maintain a financially viable and prof-
itable product portfolio, and to avoid costly inventory overages and shortages. The
process is directly supported by a mathematical optimization model that enables on
demand up-selling, alternative-selling and down-selling to better integrate the sup-
ply chain horizontally, connecting the interaction of customers, business partners and
sales teams to procurement and manufacturing capabilities of a firm. We outline the
business requirements for incorporating such a process into supply chain operations,
and highlight the advantages of ATS through simulations with realistic production
data in a computer manufacturing environment. The models featured in this paper
have contributed to substantial business improvements in industry-size supply chains,
including over $100M of inventory reduction in IBM’s server computer supply chain.
T. R. Ervolina · M. Ettl · Y. M. Lee (B)

IBM T.J. Watson Research Center,
P.O. Box 218, Yorktown Heights, NY 10598, USA
T. R. Ervolina
M. Ettl
D. J. Peters
Real Estate Operations, 294 Route 100, Somers, NY 10589, USA

146 T. R. Ervolina et al.
Keywords Availability management · Assemble-to-order · Demand shaping ·

Configure-to-order
1 Introduction
In today’s competitive and dynamic business environment, companies need to contin-

ually evaluate the effectiveness of their supply chain and look for ways to transform
business processes to achieve superior customer service and higher profitability. Imbal-
ances between supply and demand are the primary reason for degraded supply chain
efficiency, often resulting in delinquent customer orders, missed revenue, and excess
inventory. This paper describes a novel availability management process called Avail-
able-To-Sell (ATS) that incorporates demand shaping and profitable demand response
to drive better operational efficiency of the supply chain. The proposed methodol-
ogy aims at finding marketable product alternatives that replace demand on supply-
constrained products while minimizing expected stock-out costs for unfilled product
demand and holding costs for left-over inventory. While most prior related literature
focuses on the concept of Available-To-Promise (ATP) where a scheduling system
determines a particular product’s availability, this paper proposes a new approach
where product substitutions and up-sell opportunities are considered in the planning
phase. The ATS business process is most effective in an assemble-to-order manufactur-
ing environment where end products are configured from pluggable components, such
as computer manufacturing where computer systems are assembled from standardized
components such as hard disks, microprocessors, video cards, etc.
Industry best practices for demand shaping and demand response include identify-
ing entry level products suitable for up-selling, changing marketed products based on
supply position, providing product alternatives, and methods of continuous up-selling
and cross-selling to meet financial objectives (O’Marah and Souza 2004; Cecere et al.
2005). While every industry struggles with instabilities in demand and supply syn-
chronization, real-time supply chain software provides continuous visibility to the
supply position of every part. If a part is in short supply due to greater demand, the
sales department is informed to check if demand can be moved to alternative products.
For example, if a 3.0 GHz processor is in short supply marketing teams could offer an
up-sell product with a 3.2 GHz processor to the technologically savvy customer, or a
down-sell model with a 2.8 GHz processor to a price-sensitive customer. In a consumer
society driven by having a product in next to real time, improved shipment or arrival
lead times can be a compelling factor in a purchase decision.
Consider the following simple example that illustrates the benefits of ATS over
a conventional ATP planning process. A firm sells two types of computer products,
denoted M1 and M2 , to its customers over a given planning period, e.g., a quarter. Each
product is assembled from one of two processors, hard drives and memory modules as
shown in Table 1. The demand for each product is 1,000 units. The component quan-
tities that would be required in order for all customers to receive their choice of prod-
uct are 1,000 units of every component. Now, suppose that the firm’s micro supplier
has a short-term manufacturing constraint on 3.0 GHz processors and committed to
deliver 500 units less than the required 1,000 units. To compensate for the shortfall, the
Managing product availability 147
Table 1 Example of availability management with ATP and ATS
Components Price Profit Component Product portfolio

supply
Group Technology M1 M2 M3
SYSTEM 3.0 GHz/800 MHz Xeon 391 51 500 1 – –

PROCESSORS 3.2 GHz/800 MHz Xeon 504 84 1,500 – 1 1
HARD DRIVES 60 GB 4200 RPM 216 36 1,000 1 – 1
120 GB 7200 RPM 312 72 1,000 – 1 –
MEMORY 512 MB SDRAM 192 32 1,000 1 – 1
1.0 GB SDRAM 260 60 1,000 – 1 –
Demand forecast 1,000 1,000 –
ATP schedule 500 1,000 –
ATS schedule 500 1,000 500
supplier promised additional 500 units of 3.2 GHz processors. Given these component
quantities, a conventional ATP model would match the available component supplies
to the demand and create a so-called ATP schedule as shown at the bottom of the
Table 1. As a result of the constrained supply of 3.0 GHz processors, the firm incurs
500 unfilled orders of product M1 , and 500 units of unallocated inventory of 60 GB
hard drives and 512 MB memory modules. The gross profit from sales under the ATP
schedule is $275,500.
To effectively deal with supply and demand imbalances, the proposed ATS model
aims at finding marketable product alternatives that replace demand on supply-
constrained products while minimizing inventory holding costs from left-over com-
ponent supplies. In the above example, the ATS model creates an alternative product
M3 shown in the shaded column of the table that can be substituted for the con-
strained product M1 (possibly in conjunction with a price discount offered to customers
for accepting the substitute). The ATS schedule eliminates any unfilled orders and
excess component inventories. The gross profit from sales under the ATS schedule is
$351,500 which corresponds to a percentage profit gain of 27% over the conventional
ATP approach.
The ATS capabilities described in this paper can easily be imbedded into supply
chain operations to improve day-to-day flexibility. Direct sales businesses that deal
with customers directly through external websites or telesales systems can use the pro-
posed models to highlight featured products on-the-fly based on current component
availability and steer customers towards product configurations that they can supply
easily and profitably. Business benefits are increased revenue, profitability, and mar-
ket share, and improved client satisfaction. Additional financial benefits that directly
impact the profit and loss statement are the cost avoidance of brokering, scrapping
and inventory obsolescence, reduction of inventory carrying costs, return cash for
additional investments, and improved cash-to-cash cycle times.
The contributions of this paper are twofold. First, we present a linear program-
ming model for ATS that determines the optimal planned allocation of products and
components to market segments based on customer demand. Secondly, we develop a

simulation model to evaluate the benefits of the ATS model in an order execution envi-
ronment where products and components must be allocated to customer orders over
time. The simulations demonstrate how ATS enables on demand up-selling, alterna-
tive-selling and down-selling to better integrate the supply chain horizontally, connect-
ing the interaction of customers, business partners and sales teams to the procurement
and manufacturing capabilities of a company.
The remainder of this paper is organized as follows. In Sect. 2 we review the related
literature. In Sect. 3 we present the underpinning principles of ATP and ATS, and dis-
cuss the advantages and disadvantages of each management approach. In Sect. 4 we
propose an ATS planning model that captures customer preferences to effectively mit-
igate supply and demand imbalances. In Sect. 5 we present a simulation framework
for availability management in assemble-to-order supply chains. Numerical findings
and discussions of results are presented in Sect. 6. These produce several insights
into how advanced availability management can help proactively coordinating supply
and sales, and quantify several business benefits in the context of assemble-to-order
manufacturing. Finally in Sect. 7 we present concluding remarks and suggestions for
future research.
2 Literature review
There are two streams of research that are related to our work: (1) models from the pro-
duction planning and operations literature that deal with Available-to-Promise (ATP)
systems for order promising and fulfillment, and (2) models from the operations man-
agement literature that consider inventory problems with configurable products and
product substitution. We provide an overview of both research streams.
There is an extensive literature in the production planning area dealing with real-
time order promising and ATP (e.g., Kilger and Schneeweiss 2000; Moses et al. 2004;
Hoop and Roof 1999). Ball et al. (2004) develop a general modeling framework for
availability promising and present examples of ATP business practices from electron-
ics companies including Dell and Toshiba. Chen et al. (2002) present a mixed integer
programming model that provides an ATP order promising and fulfillment solution
for batch orders that arrive within a predefined time interval. Ervolina and Dietrich
(2001) describe an application of the implosion technology for ATP order promising
in assemble-to-order (ATO) and configure-to-order (CTO) manufacturing environ-
ments. The goal is to create a feasible production plan that can be used to schedule
or promise customer orders. Chen-Ritzo (2006) studies a similar availability manage-
ment problem in a CTO supply chain with order configuration uncertainty. Akcay and
Xu (2004) develop a two-stage stochastic integer program with recourse to allocate
constrained components so as to maximize the fraction of orders assembled within a
quoted maximum delay. The closest work in this stream is Dietrich et al. (2005) which
describes a deterministic implosion model that identifies suitable product configura-
tions for an Available-to-Sell process that consume the most surplus inventory and
require minimal additional component purchasing costs. The focus of this model is on
the perspective of the firm, independent of the customer’s propensity to buy alternative
products. Market demand, customer preferences, or product substitution policies are

not considered. In contrast, we explicitly model customer expectations in a dynamic
setting, utilizing a customer behavior model that determines how customers evaluate
product substitutions if their initial product selection is unavailable.
In the operations management literature, there are several papers in which product
substitutions or flexible customer requirements are important elements. Bassok et al.
(1999) study a multi-product inventory problem with full downward substitution where
excess demand for a product can be filled using a product of higher utility. Hale et al.
(2001) extend the analysis of the downward substitution problem to an ATO system
with two end-products where each product is composed of two components. Substitu-
tions are carried out at the component level. Gallego et al. (2006) consider downward
substitution to satisfy unmet demand for lower grade products in a semiconductor
production environment, and propose a heuristic allocation scheme for determining
near-optimal build plans. Swaminathan and Tayur (1998) determine optimal configu-
rations of semi-finished products (vanilla boxes) along with their inventory stocking
levels to enable late customization in an assemble-to-order supply chain for computer
manufacturing. Building upon these concepts, Yunes et al. (2007) apply a customer
migration model in conjunction with mixed-integer programming to determine the
optimal product portfolio at John Deere and Company. A customer migration list
contains alternative product configuration choices if a customer’s preferred product
selection is unavailable. Balakrishnan and Geunes (2000) study a production planning
problem with flexible bills-of-materials and component substitution. A dynamic pro-
gramming solution method is developed to find production and substitution quantities
that satisfy demands at minimum total cost, comprising setup, production, substitu-
tion, and inventory holding cost. Because supply is assumed to be unconstrained, the
model does not address matching of demand and supply. Balakrishnan and Geunes
(2003) consider a production planning problem faced by a steel manufacturer whose
customers allow flexibility in product specification. In a recent paper, Balakrishnan
et al. (2005) apply concepts from revenue management to investigate how a firm can
maximize profits by shaping demand through dynamic pricing.
3 Availability management
Availability management is the overarching task of coordinating the planning of prod-

uct availability with the real-time promising of customer orders. The most common
approach to Availability Management is the Available-to-Promise (ATP) process. The
goal of the ATP process is to generate a single, integrated supply plan that brings
together the business objectives of finance, sales, and marketing, with the reality of
the unbiased forecast and the capacity of the supply chain. This integrated supply
plan is called the “ATP schedule” that allocates component supplies to products and
customer segments. During execution, the ATP process deals with a real-time stream
of customer orders. As a customer request arrives, the order scheduling process must
promise an availability date to the customer. This task involves checking the contents
of the order against the ATP schedule, determining an availability promise date to the
customer, and decrementing the ATP to accurately reflect the supply committed to new
customer orders. The ATP process utilizes an analytical technique called implosion
to generate an optimized ATP schedule that takes into account supplier commitments
and limited manufacturing capacities (e.g., Dietrich et al. 2005). The ATP schedule
seeks to satisfy a fixed sales target that is based on a demand forecast. Because ATP
provides no mechanism for dealing with unallocated supply, a separate non-integrated
business process is often created to manage inventory excess and overages, e.g. by exer-
cising buy-back agreements with component suppliers or other procurement-related
techniques.
In today’s environment, customers expect that products are available in a large
variety of configurations, and, with this expanding variety, customers have become
increasingly flexible in what they will purchase. When the capacity of the supply chain
does not directly align with the sales target, imbalances often result in an ATP schedule
that falls short of customer demand. In this paper, we propose a new methodology
called Available-to-Sell (ATS) that more effectively balances supply and demand by
taking advantage of demand flexibility when determining the allocation of products to
customer segments. ATS is designed to find alternative product configurations that best
consume excess supply while minimizing additional procurement investments. ATS is
most effective in an assemble-to-order (ATO) manufacturing environment where end
products are configured from standard components, and where the simplified product
structure ensures that product substitutions will drive customer interest.
The main output of the ATS process is an “ATS schedule” that comprises opti-
mized availabilities of a firm’s core products as well as saleable product alternatives.
The ATS schedule takes advantage of up-sell, alternate-sell or down-sell opportunities.
An up-sell opportunity is where a customer is sold a more richly configured solution
above the customer’s initially selected price range; price incentives may be used to
entice the customer to agree to an up-sell. An alternative-sell relates to a sale of a sim-
ilar product that falls within the selected price range. An alternative-sell is performed
when an up-sell is not available or the customer opts for a similarly priced product.
A down-sell opportunity refers to a sale of a product that falls below the price range
selected by the customer. ATS can drive further efficiencies if the marketplace can be
subdivided into customer segments where customers in the same segment have similar
buying behavior and lifetime values to the company (customer lifetime value is the
present value of future cash flows attributed to a customer relationship).
4 ATS planning model
In this section, we define the underlying customer behavior model and formulate a
linear programming model for ATS. Before we state the problem, we define the rele-
vant notation that is used throughout the remainder of the paper.
Customer behavior model
• C: set of customer segments, or customer classes, indexed by c.

• αc ≥ 0 : Reservation price parameter for customers in segment c ∈ C.
• 0 ≤ βc ≤ 1 : Reservation quality parameter for customers in segment c ∈ C.
• γc : First-choice probability of customers in segment c ∈ C, denoting the probabil-

ity that a customer in segment c will only accept its first-choice product selection
and no product alternatives.
Products and components
• I : set of components, indexed by i.
• M : set of core products, indexed by m.
• N : set of alternative products, indexed by n.
• S : set of products, indexed by s where S := M ∪ N .
• u is : usage of component i ∈ I in product s ∈ S (bill-of-material); the usage
values are assumed to be binary.
Supply and demand
• Q i : supply of component i ∈ I .
• Dmc : demand for core product m ∈ M and customer segment c ∈ C; demand is
assumed to be deterministic.
Cost and profit
• h i : inventory holding cost of component i ∈ I per time period, e.g., week.
• rs : per-unit retail price of product s ∈ S.
• ps : per-unit marginal profit of product s ∈ S.
• qsc : quality of product s ∈ S for customer segment c ∈ C.
• c : backorder penalty of a customer order in segment c ∈ C for product m ∈ M;
bm
the backorder penalty is considered a one-time charge that is independent of the
duration of the backorder.
• wmnc : per-unit penalty cost for substituting one unit of product n ∈ N for one unit
of product m ∈ M (price discount).

Decision variables
• Xmc : ATS quantity for core product m ∈ M and customer segment c ∈ C.
• Ymn : ATS quantity of alternative product n ∈ N used as a substitute for core

c
product m ∈ M in customer segment c ∈ C.

c and Y c represent the optimized ATS schedule.
The two sets of decision variables X m mn
4.1 Customer behavior model
Similar to the customer migration model described in Yunes et al. (2007), we assume
that a known percentage of customers accept a substitute product if its price and quality
are within a certain range. The first-choice probability γc determines the proportion of
customers in segment c ∈ C that will not accept substitutions. The reservation price
determines the incremental price that a customer in segment c is willing to pay for an
alternative product. If the customer’s initial selection is product m and rm denotes the
price of product m, the customer’s reservation price is (1 + αc )rm . Similarly, if qmc
denotes the quality level of product m for a customer in segment c, the customer’s res-
ervation quality is (1−βc )qmc . A customer is willing to purchase an alternative product
n ∈ N if its price rn does not exceed the customer’s reservation price (1 + αc )rm , and
if its quality qnc is no less than the customer’s reservation quality (1 − βc )qmc . If no
alternative selections in the desired price and quality range are available, the customer
order is assumed to be backlogged.
4.2 Problem formulation
The ATS planning model is formulated as a linear programming problem. For analyt-
ical tractability, we choose a simple single-period model to accomplish the short-term
allocation of components and products to customer segments. The model recognizes
the unique customer preferences associated with a segmented market.
Consider a two-level product structure that consists of a set of components I where
each component i ∈ I has some finite supply Q i , and a product portfolio S = M ∪ N
that is the union of a set of core products M and a set of alternative products N . The set
of core products M contains currently featured products that are offered by the seller.
The set N contains alternative products that may be used to fill unsatisfied demand
for certain core products with additional substitution cost incurred. Since some core
products could be substituted for other core products, the sets M and N may not be
disjoint. The demand in customer segment c ∈ C for core product m ∈ M equals Dm c.
Given a component i, the number of units of the component in product s ∈ S equals

u is . The output of the model is the ATS schedule, i.e., the allocation of core and alter-
native products to customer segments. The model maximizes total supply chain profit,
consisting of net profit from sales less component holding costs, backorder costs and
substitution penalty costs.
Using the above notation, the objective function of the ATS planning problem is
formulated as follows:

Max Z (X m, Ymn ) =
c c
pm X m +
c c
pn Ymn
m∈M c∈C n∈N c∈C

− c
bm c
Dm − Xm
c
− c
Ymn
m∈M c∈C n∈N

− hi Qi − u im X m +
c c
u in Ymn
i∈I m∈M c∈C n∈N c∈C

− wmn
c c
Ymn (1)
m∈M n∈N c∈C
The first term is the net profit derived from sales of core and alternative products. The
second term represents backorder costs that are incurred when the allocation of core
and alternative products falls short of the customer demand Dm c . The third term repre-
sents the inventory holding cost incurred for unused components when the component
usage is less than the available supply. The last term represents product substitution
costs that are incurred as a result of discounts offered to customers for accepting a
substitute.
Next we formulate the constraints. Given the demand Dm c for core product m ∈ M,
the total build volume for this product, including the volume substituted by new alter-
native products, cannot exceed the demand for core product m:

c
Xm + c
Ymn ≤ Dm
c
for all m ∈ M, c ∈ C (2)
n∈N
The ATP schedule must be feasible with respect to the component supply, i.e., the
number of components consumed plus any unallocated inventory must be less than or
equal to the available component supply:

c
u im X m + c
u in Ymn ≤ Q i for all i ∈ I (3)
m∈M c∈C n∈N c∈C
The number of product substitutions for core product m ∈ M cannot exceed the
fraction of demand that can be filled with alternative products:

c
Ymn ≤ (1 − γc )Dm
c
for all m ∈ M, c ∈ C (4)
n∈N
Any alternative product n ∈ N that is used to substitute demand for core product
m ∈ M in customer segment c ∈ C must meet the reservation price and reservation
quality requirements qnc ≥ (1 − βc )qmc and rn ≤ (1 + αc )rm . These requirements are
c as follows (alterna-
expressed as linear constraints in the substitution quantities Ymn
tively, for combinations of m, n, and c that do not satisfy the above conditions, the
c variables could be eliminated during pre-processing):
Ymm
c
c
Ymn qn − (1 − βc )qmc ≥ 0 for all m ∈ M, n ∈ N and c ∈ C (5)
Ymnc
[(1 + αc )rm − rn ] ≥ 0 for all m ∈ M, n ∈ N and c ∈ C (6)
Finally, constraints (7) and (8) impose non-negativity constraints on the decision vari-
ables:
Xmc
≥ 0 for all m ∈ M, c ∈ C (7)
Ymn ≥ 0 for all m ∈ M, n ∈ N , c ∈ C
c
(8)
The optimization problem (1)–(8) is a linear program that can be solved very efficiently
for large problem sizes.
The single-period model described above has a few limitations which we discuss
next. First, the model does not capture situations in which demand and component
supplies are planned over multiple time periods. In many industrial applications,
component suppliers commit to the delivery of component supplies several weeks
into the future. Such multi-period deliveries might create correlations of component
allocations to customer segments that are not captured in a single-period model.
Under a time dimension, our model can be used to allocate available components
to customer segments in a myopic fashion using a rolling horizon. Although this is

an approximation, most practical applications of the model focus on maximizing the
firm’s ability to fill orders over a short-term planning horizon (1–2 weeks) in a myopic
fashion. The reason is that supply chain managers are often reluctant to reserve supply
for future weeks because demand forecasts become increasingly uncertain. More-
over, because many high-technology component suppliers operate inventory stocking
points near the OEM’s (original equipment manufacturer) production facilities they
are usually able to respond to supply changes on a weekly basis. If the likelihood of a
long-term supply constraint is small and supplier lead times are short, we suspect that
the effect of applying a myopic strategy on the optimal component allocations will be
small.
Second, we assume that all demand in a period is observed before products are
allocated to customer segments. In reality, customers orders are placed continuously
and components would need to be allocated to them on an order by order basis. Thus,
our model provides an optimistic estimate of the profitability of a firm under a given
demand, supply plan, and customer behavior model.
Third, the assumption that all customers within a market segment have the same
buying behavior (in terms of their willingness to upgrade to a higher-priced product)
may not be true in reality. It would be more realistic to model the customers’ reserva-
tion price and quality values as random variables. In our numerical study, we allowed
the customer behavior parameters to randomly deviate from their base values.
Finally, we assume that the component supply quantities are exogenous inputs to
the model that are determined prior to executing the allocation of components to cus-
tomer segments. It would be interesting to develop an integrated model that not only
determines product allocations but also the ideal component supply mix in terms of
maximizing the firm’s profit in view of customer’s propensities to upgrade to higher-
priced, more profitable products. Such an integrated decision model is beyond the
scope of this paper, but is a promising area for research.
5 Simulation framework
The simulation model was built using the Availability Management Simulation Tool
(AMST) which has been used at IBM to develop various availability management sim-
ulation models (Lee 2006). AMST was developed using the simulation capabilities
of IBM’s WBM (WebSphere Business Modeler) as a simulation modeling frame-
work for availability management processes. The simulation framework consists of
reusable components and methods which are easily adapted to different availability
management environments.
The goal of the simulation is to evaluate the potential of the ATS model in a real-
istic order execution environment where components must be allocated to products
on an order by order basis. The simulation model can be executed in single-period or
multi-period mode. In single-period mode, the ATS planning model is invoked once
to determine the planned allocation of products to customer segments. In multi-period
mode, the ATS planning model is invoked in a myopic fashion at pre-determined time
points (roll-forward events) after the remaining order backlog and unsold inventory
Fig. 1 ATS simulation model and data flow
from previous time periods is evaluated and carried forward to the next time period.
Subsequently, the simulator generates a randomized arrival sequence of customer
orders for a given demand statement. Each customer order is assigned a customer
segment, a core product selection, a first-choice probability, and a reservation price
and reservation quality pertaining to the assigned segment. The outputs of the sim-
ulation are statistical outcomes of the system performance metrics, including sales
profit, backorders, proportion of orders filled, and inventory holding costs. Figure 1
illustrates the data flow of the simulator and its interaction with the ATS model.
The ATS linear programming model assumes that all demand in a period is observed
before products are allocated to customer segments, whereas in reality customer orders
are placed continuously and components need to be allocated to each order as it arrives.
Thus, the results from the ATS model establish a theoretical upper bound of the prof-
itability under a given demand, supply, and customer behavior model. To assess the
true potential of ATS, the optimized allocation must be simulated under an operational
“order acceptance” policy. The simulation model therefore embodies different order
acceptance policies for allocating ATS quantities to customer orders. The simplest
order acceptance policy is First-Come-First-Serve (FCFS) shown in Fig. 2.
For each customer order, the FCFS order scheduler checks the ATS quantity of the
customer’s product selection. If the first-choice product is available, the order is sched-
uled and fulfilled. If the first-choice product is not available, the scheduler determines
whether the customer is willing to take a substitution, in which case the scheduler
looks for a substitution product that meets the customer’s price and quality tolerance.
In this case, the first available alternative product within the price and quality range
of a customer is selected. A more sophisticated implementation might consider prof-
itability or availability as selection criteria. When a substitution is found, the order is
scheduled and fulfilled. If the customer is not willing to take a substitution or a suitable
substitution is not found, the order is backlogged.
Under a second order acceptance policy called rationing, a substitute product may
be offered even when the preferred product selection is available in order to up-sell
an item that is available, acceptable, and may improve the profit from the sale. The
flow diagram is depicted in Fig. 3. If a customer is willing to take a substitution, the
Fig. 2 First-Come First-Serve (FCFS) order acceptance policy
Fig. 3 Rationing order acceptance policy
scheduler randomly searches for an alternative product that meets the price and quality
tolerance of the customer from the ATS schedule. Again, the first available substitute
product within the price and quality range of the customer is selected. If no suitable
product alternative is found and the customer’s first choice product is available, the
order is fulfilled; otherwise the order is backlogged.
In order to incorporate the benefits of risk pooling over multiple customer segments
into the simulation, the product allocation quantities X m c and Y c that are created by
m,n
the ATS planning model are transformed into an “aggregated” ATS schedule. The
allocated product quantities of the ATS schedule are assigned to common inventory
buckets, denoted X̄ m and Ȳn , that are derived from the original ATS allocation quan-
tities as follows:

X̄ m = c
Xm for all m ∈ M (9)
c∈C

Ȳn = c
Ym,n for all n ∈ N (10)
c∈C m∈M
In the simulation experiments presented in Sect. 6, both order acceptance policies use
the aggregated ATS allocation quantities X̄ m and Ȳn to determine whether to substitute
an alternative product for an item requested in a customer order.
6 Numerical study
The numerical study focuses on prescribing how the firm should adjust its sales strat-
egy when faced with different degrees of supply and demand imbalances. To address
this goal, we implemented the availability management models described in the previ-
ous sections in the context of a representative assemble-to-order (ATO) supply chain
for mid-range server computers. In Sect. 6.1 we describe the example scenario for the
numerical study. In Sect. 6.2 we test the conjecture that intelligent demand shaping
based on customer preferences can provide significant financial benefits over a tradi-
tional ATP-based approach, particularly in environments where inventory imbalances
exists. We apply the linear programming model to show how a firm can take advan-
tage of up-sell opportunities for increased profit, and analyze the effect of customers’
price sensitivity on profitability. In Sect. 6.3 we use the linear programming model
in conjunction with simulation to investigate the performance of ATS under different
order acceptance policies in a realistic order execution environment.
6.1 Supply chain model
The example scenario for the numerical study is derived from an industry-size assem-
ble-to-order supply chain of a server computer product line. The product portfolio con-
sists of eight mainstream server computer products that represent a whole spectrum
of price-performance points. The products and their bills-of-materials are depicted
in Table 2. Products M1 and M2 are entry level products, M3 to M6 are mid-range
systems, and M7 and M8 are high-performance computers. Each product is assembled
from components of six different commodity groups: system processors, memory,
hard drives, optical drives, video adapters and software preloads. For example, prod-
uct M1 is assembled from a 2.8 GHz system processor, a 30 GB hard drive, 128 MB
memory, a 48X CD-RW optical drive, an Extreme 3D video card and a system soft-
ware preload B. Although in reality each server product is assembled from dozens of
components, the major components represented in this study account for more than
80% of the cost of a product. The manufacturing operation is driven by an assemble-
to-order process. The table also shows the price, gross profit, and quality score of every
component. The quality score of a component depends on its parts worth relative to the
other components in the same commodity group. Components with the highest parts
worth are assigned the highest quality score, and they carry the highest gross profit.
Table 2 Bill-of-materials structure used in the numerical study
158
Components Price Profit Quality score Product portfolio
Group Technology Economy Value Performance M1 M2 M3 M4 M5 M6 M7 M8
SYSTEM 2.8 GHz/800 MHz Xeon 286 26 20 25 25 1 1 – – – – – –

PROCESSORS 3.0 GHz/800 MHz Xeon 391 51 30 50 50 – – 1 1 – – – –
3.2 GHz/800 MHz Xeon 504 84 40 75 75 – – – – 1 1 – –
3.4 GHz/800 MHz Xeon 650 150 50 100 100 – – – – – – 1 1
HARD DRIVES 30 GB 4200 RPM 110 10 10 20 20 1 – 1 – – – – –
40 GB 4200 RPM 161 21 20 40 40 – 1 – – 1 – – –
60 GB 4200 RPM 216 36 30 60 60 – – – 1 – 1 – –
120 GB 7200 RPM 312 72 40 80 80 – – – – – – 1 –
160 GB 7200 RPM 405 105 50 100 100 – – – – – – – 1
MEMORY 128 MB SDRAM 88 8 10 10 20 1 – 1 – – – – –
256 MB SDRAM 138 18 20 20 40 – 1 – – 1 – – –
512 MB SDRAM 192 32 30 30 60 – – – 1 – – – –
1.0 GB SDRAM 260 60 40 40 80 – – – – – 1 1 –
2.0 GB SDRAM 324 84 50 50 100 – – – – – – – 1
OPTICAL DRIVES 9.5 mm Slim DVD 77 7 30 30 30 – 1 1 – – – – –
48X CD-RW 104 14 40 40 40 1 – – – 1 – 1 1
48X CD-RW/DVD 120 20 50 50 50 – – – 1 – 1 – –
VIDEO ADAPTER ATI Performance 2D 138 13 30 30 30 – 1 1 – 1 – – –
Extreme 3D Express 173 23 40 40 40 1 – – 1 – 1 1 1
NVidia Advanced 3D 210 35 50 50 50 – – – – – – – –
SOFTWARE Preload A 110 10 10 10 10 – 1 – – – 1 – –
Preload B 132 12 20 20 20 1 – 1 – 1 – – –
Preload C 168 28 30 30 30 – – – 1 – – – –
Preload D 240 40 40 40 40 – – – – – – 1 –
Preload E 325 75 50 50 50 – – – – – – – 1
T. R. Ervolina et al.
Table 3 Product quality, demand and backorder costs by customer segment
Product Price Profit Economy segment Value segment Performance segment

rm pm
Quality Demand Backlog Quality Demand Backlog Quality Demand Backlog

1
qm 1
Dm 1 q2
cost bm 2
Dm 2 q3
cost bm 3
Dm 3
cost bm
m m
M1 960 92 47 800 44.6 39 200 133.8 37 – 178.4

M2 978 95 43 800 45.5 39 200 136.4 39 – 181.9
M3 1,002 101 43 500 46.8 40 400 140.3 38 100 187.1
M4 1,284 190 70 500 63.0 65 400 188.9 64 100 251.9
M5 1,218 161 57 300 58.8 56 400 176.4 54 300 235.2
M6 1,380 233 70 200 69.1 69 400 207.4 70 400 276.5
M7 1,656 358 83 200 86.9 85 300 260.7 84 500 347.6
M8 1,836 450 93 1000 99.0 95 300 297.0 96 600 396.0
The quality level of a product is computed as the (weighted) average of the quality
scores of the components used in its configuration. Each component in a commodity
group is assigned a quality score (a value between 0 and 100) based on its quality rela-
tive to all other components in the same commodity group. Higher scores are assigned
to components with higher parts worth, e.g., a 120 GB hard disk will score higher than
a 60 GB hard disk.
We assume that customers belong to one of three market segments, denoted econ-
omy, value and performance. Customers in the economy segment tend to purchase
entry-level products and are highly price sensitive (i.e., they have a low reservation
price). Customers in the value and performance segments are moderately price sensi-
tive and are more often willing to accept up-sells. The perceived quality of a product
depends on the customer segment. Customers in the economy segment have a balanced
valuation of components in the six commodity groups, with 50 being the highest quality
score in each group. Customers in the value segment place a higher relative importance
on system processors and hard drives, whereas customers in the performance segment
place their highest importance on processors, hard drives, and memory as indicated
by a top quality score of 100.
Table 3 summarizes the price and quality scores of the product portfolio for the
three customer segments. The price of product m, rm , is the sum of the prices of the
components used in its bill-of-materials. The quality score of product m in customer
segment c, qmc , is the average of the quality scores of all components used in its bill-of
material. The table also shows the backorder penalty, bm c , and the customer demand,
c
Dm , for each product and each customer segment. The backorder penalty is a fraction
of the price of a product; backorder penalties are lowest for economy customers and
highest for performance customers.
The reservation price and reservation quality parameters used in the customer
behavior model are assumed to be (α1, α2, α3 ) = (0.1, 0.2, 0.3) and (β1, β2, β3 ) =
(0.3, 0.2, 0.1). The parameter choice is driven by the fact that customers in the econ-
omy segment tend to be highly price sensitive and may compromise on product quality,
whereas customers in the performance segment are relatively price insensitive but
demand a high quality level of a product. In addition to utilizing a reservation price
and a reservation quality to determine whether a customer will consider an alternative
product, we assume that a fraction of customers are committed to their first product
choice and will not accept an alternative configuration. The first-choice probability
in the baseline scenario is γc = 0.5 for all three customer segments, i.e., 50% of
customer orders will not accept product alternatives if their initial product selection is
unavailable.
6.2 Comparisons between ATS and ATP
In practice firms often highlight entry-level products to customers to provide an inter-

esting price-performance point that will establish a sound brand image and elicit a
favorable customer response, i.e., a buy decision. The marketed entry-level products
are usually supplied at a lower rate than actual demand, driving longer product avail-
ability lead times (although the seller must have a reasonable supply line for the entry or
economy level products to meet regulatory and country specific business practices).
The goal is to have customers contact the seller which provides the opportunity to
up-sell the customer to a more richly configured solution, normally at a higher price-
performance point, usually thought of as the market “sweet spot” for the product
category. In this section, we apply the linear programming model of Sect. 4 to inves-
tigate how demand shaping with ATS helps improve the operational performance of
the supply chain when the component supply deviates from the ideal net component
requirements. We illustrate the impact of employing different supply schemes on a
firm’s profitability. This enables us to identify conditions under which benefits to the
firm are the most significant.
Supply and demand imbalances
To obtain a baseline supply plan for the demand scenario given in Table 3, we first
calculate the net component requirements by “exploding” the demand through the bill-
of-materials in a standard MRP-type calculation (e.g., Hopp and Spearman 2000). This
calculation yields supply quantities that would be required in order for all customers
to receive their first-choice product. This supply plan is labeled “unbiased component
mix” in Table 4. Component mix denotes the supply of a component relative to the
supply of the other components in the same commodity group. For example, 25%
of the first-choice products utilize a 30 GB hard drive, whereas 13% use a 160 GB
hard drive. Table 4 shows two biased supply scenarios, denoted “low skew” and “high
skew”. In both scenarios, high-value components are procured at a higher rate than
actual customer demand, and low-end components are purchased at a lower rate than
actual demand. For each supply scenario, we generated ten problem instances by
applying small random perturbations to the component mix shown in the table, and
solved each instance using the linear programming model described in Sect. 4. The
results presented next are averages over these ten problem instances.
Table 4 Component mix for unbiased, low, and high supply skew
Components Unbiased component Biased component Biased component

mix (no skew) mix (low skew) mix (high skew)
SYSTEM 2.8 GHz/800 MHz Xeon 25% 24% 19%

PROCESSORS 3.0 GHz/800 MHz Xeon 25% 26% 31%
3.2 GHz/800 MHz Xeon 25% 25% 20%
3.4 GHz/800 MHz Xeon 25% 25% 30%
HARD DRIVES 30 GB 4200 RPM 25% 10% 10%
40 GB 4200 RPM 25% 28% 28%
60 GB 4200 RPM 25% 31% 26%
120 GB 7200 RPM 13% 19% 21%
160 GB 7200 RPM 13% 13% 15%
MEMORY 128 MB SDRAM 25% 10% 10%
256 MB SDRAM 25% 28% 28%
512 MB SDRAM 13% 19% 19%
1.0 GB SDRAM 25% 28% 25%
2.0 GB SDRAM 13% 16% 19%
Fig. 4 Attainable net profit achieved by ATS under different supply scenarios
Figure 4 displays the attainable net profit, i.e., sales profit less backorder penalties
less inventory holding costs, under the ATS strategy where the firm takes advantage of
up-sell opportunities. We can see that the profit gain is monotone increasing with the
degree of supply skew for the two most profitable customer segments (performance
and value). The profit decreases slightly for the economy customer segment which is
caused by supply constraints for low-end components, combined with limited oppor-
tunities for up-selling or alternative-selling due to the low reservation prices of entry
level customers.
Figure 5 shows the percentage profit gain of ATS compared to an ATP-based
approach. This can be interpreted as the value of demand shaping when the firm
has advance knowledge about the price sensitivity and quality profiles of its customer
segments, and under the assumption that demand is known. The figure illustrates that
Fig. 5 Percentage profit gain of ATS over ATP under different supply scenarios
Fig. 6 Order fill rate and number of substitutions under different supply scenarios
the net profit derived from all customer segments increases as the degree of supply
skew increases. The firm can be significantly better off financially (up to 10%) by
recognizing opportunities for up-selling to the market sweet spots in its customer seg-
ments. The profit gain in the firm’s two most profitable customer segments (value and
performance) more than offsets the profit decrease in the entry-level (economy) seg-
ment. The modest positive profit gain of ATS under the no skew scenario is a result of
averaging over multiple random perturbations that were analyzed for each component
mix (i.e., the sampled component mix of each of the ten problem instances deviates
slightly from the unbiased component mix shown in Table 2 which results in a small
order backlog under ATP). Although not shown here, we note that the profit gain
would eventually decrease as the supply is progressively skewed towards high-value
components because the proportion of customers willing to accept up-sell products at
increasingly higher price points will eventually decline.
For the same experiment, Fig. 6 shows the number of product substitutions in the
different customer segments under the ATS regime as well as the proportion of orders
that are filled with either the customer’s first choice or an alternative product. The
proportion of orders filled is plotted against the secondary vertical axis.
We can make two key observations. First, the order fill rate decreases from 100% to
97%. This decline is driven entirely by backorders in the economy customer segment.
Second, the relative proportion of customers that purchase an alternative product is
highest in the value segment (43%), as is the incremental profit gain for this segment
as illustrated in Fig. 5. This observation is consistent with marketing analyses that
suggest that customers in the mid-range segment are most likely to respond favor-
ably to alternative product offerings. Firms should therefore focus their marketing
efforts on protecting revenues derived from the mid-range segment and employ effec-
tive demand shaping actions to grow the share of the wallet from these customer
accounts.
Effect of customer behavior
To investigate how customer buying preferences can affect a firm’s profitability, we

applied the heavily skewed supply scenario from the previous experiment and analyzed
the firm’s sales performance under different settings of the first-choice probability γc .
We examine three scenarios where γc = 0.5, 0.75 and 1.0. Higher values of γc imply
that customers are less willing to accept an alternative product if their initial product
choice is unavailable. In the last scenario, customers always prefer their initial product
choice. Figure 7 displays the effect of customers’ first-choice probability on the order
fill rate. We observe that as the first-choice probability increases the order fill rate
decreases. This is expected because a higher first-choice probability translates into
fewer customers accepting product substitutions when their first choice is stocked out.
The order fill rate in the economy segment decreases much more rapidly than in the
other segments because economy customers predominantly buy entry-level products
that are in short supply.
Table 5 depicts the net profit, backorder cost, and inventory cost for the differ-
ent first-choice probability values. In the extreme case where γc = 1.0, the average
profit penalty over the baseline scenario (γc = 0.5) is 37%. The reason for the large
penalty is a significant sales decline in the economy segment combined with increased
inventory costs that are driven by unsold high-value component supplies.
Fig. 7 Effect of customer first-choice probability on order fill rate

Table 5 Net profit, backorder cost and inventory costs under different customer behavior models
Customer First choice γ = 0.50 First choice γ = 0.75 First choice γ = 1.00
segment
Net Backorder Inventory Net Backorder Inventory Net Backorder Inventory
profit cost cost profit cost cost profit cost cost
Economy 48,1398 11,242 – 341,867 38,643 – (25,345) 100,707 –

Value 635,144 – – 584,505 3,162 – 551,621 234 –
Performance 705,195 – – 665,350 – – 619,300 – –
All segments 1,821,737 11,242 21,120 1,591,721 41,805 78,540 1,145,576 100,941 189,640
6.3 Impact of order acceptance policy
As discussed earlier, a key assumption of the linear programming model is that all
demand in a period is observed before products are allocated to customer orders. The
analytical results presented in the previous section should therefore be interpreted as
a theoretical upper bound of the firm’s profitability for a given supply and demand
scenario. In a realistic supply chain environment, the achievable profitability under
ATS would depend heavily on the order acceptance policy selected to execute the
ATS schedule. Since customer orders are placed continuously over time, the order
acceptance policy must match the inflow of customer orders against the optimized
ATS schedule and determine (on an order by order basis) which product to allocate to
an individual customer order without visibility of future orders.
We developed a discrete-event simulation model described in Sect. 5 to simulate
the performance of ATS under the two different order acceptance policies rationing
and FCFS. The simulation results reported here are derived from 120 independent
replications. The simulations are based on the assumption that the realized demands
in each replication are exactly equal to the demand values shown in Table 3, only
the sequence of order arrivals is varied. All simulation runs were conducted in sin-
gle-period mode. In each replication, the simulator executes the following steps: (a)
invoke the linear programming model to determine an optimal ATS schedule; (b) com-
pute an aggregated ATS schedule where the allocations are derived from Eqs. (9) and
(10); (c) generate a randomized arrival sequence of customer orders where the total
number of orders is equal to the demand quantities shown in Table 3; and (d) fulfill
customer orders on an order-by-order basis using the desired order acceptance policy.
Recall that throughout this paper demand is deterministic which permits analyzing the
effects of the order acceptance policy on profitability and fill rates without clouding
the analysis with uncertainty in aggregate demand.
Figure 8 demonstrates how a skewed component supply mix can drive profitability.
The results from the linear programming model are included for comparison. Not sur-
prisingly, the analytical model dominates the profitability metric and drives the largest
theoretical profit improvements attainable through integrated sales actions. Under the
rationing and FCFS order acceptance regimes, profitability decreases as the supply
mix is skewed to high-value components. The rationing policy seeks suitable up-sell
opportunities when customers indicate their receptiveness to other product choices.
Fig. 8 Expected net profit achieved under rationing and FCFS for different supply scenarios
Table 6 Order fill rate achieved under rationing and FCFS for different supply scenarios
Customer No supply skew Low supply skew High supply skew

segment
Analytical Rationing FCFS Analytical Rationing FCFS Analytical Rationing FCFS
Economy 100.0% 94.7% 93.2% 95.0% 80.5% 73.1% 92.9% 76.7% 70.2%
Value 100.0% 97.5% 96.0% 100.0% 88.9% 80.2% 100.0% 89.0% 79.0%
Performance 100.0% 99.1% 97.4% 100.0% 96.3% 85.9% 100.0% 96.1% 84.6%
All segments 100.0% 96.7% 95.2% 97.9% 87.1% 78.7% 97.0% 85.5% 76.6%
FCFS performs least favorably regardless of supply skew. FCFS may achieve a sense
of equality amongst customers, but when supply is skewed or constrained it decreases
the profitability of the enterprise. Therefore, the order acceptance policy implemented
in an enterprise’s application suite can have a profound effect on profitability when
an enterprise seeks to shape demand to market sweet spots.
Table 6 shows the order fill rates achieved under the different order acceptance poli-
cies by supply skew and customer segment. The details by customer segment show the
ATS engine driving sales towards value and performance products as supply is skewed
increasingly to those products. Overall, the order fill rates achieved under rationing
and FCFS demonstrate increasing gaps to the analytical fill rates as supply is skewed
to value and performance products, thereby leaving more backorders in each sales
cycle which has an undesirable effect on revenue and profitability. Under a stochastic
demand regime, the order fill rates would deteriorate unless the baseline supply plan
had a provision for extra safety stocks to protect against higher than expected demand
levels.
Figure 9 shows the effect on profitability based on the order acceptance regime
and customers’ flexibility on their first choice. Clients with a first-choice fixation (i.e.,
probability equals 100%) provide an overall drag on profitability. Clients that are more
flexible in their final product selections increase the profitability of the enterprise, and
are more valuable to the business. The linear programming model takes advantage of
Fig. 9 Expected net profit achieved under rationing and FCFS for different customer behavior models
Table 7 Order fill rate achieved under rationing and FCFS for different customer behavior models
Customer First choice γ = 0.50 First choice γ = 0.75 First choice γ = 1.00
segment
Analytical Rationing FCFS Analytical Rationing FCFS Analytical Rationing FCFS
Economy 92.9% 76.7% 70.2% 75.7% 73.9% 67.1% 36.8% 59.3% 59.3%
Value 100.0% 89.0% 79.0% 97.4% 85.7% 75.4% 99.8% 75.9% 75.9%
Performance 100.0% 96.1% 84.6% 100.0% 96.3% 84.8% 100.0% 90.3% 90.3%
All segments 97.0% 85.5% 76.6% 88.8% 83.3% 74.3% 73.1% 72.5% 72.5%
the given supply mix to drive the highest net profit regardless of first-choice probability,
followed by the rationing method and FCFS.
Table 7 provides detailed insight into Fig. 9 by evaluating the different order accep-
tance methods and customer segments on order fill rates. The analytical model consis-
tently outperforms the rationing and FCFS methods across order fill rates regardless
of customer segment.
7 Concluding remarks
In this paper we have developed and simulated an advanced availability management

process for assemble-to-order supply chains and have outlined the business require-
ments for incorporating such a process into supply chain operations. We have described
a mathematical model that aims at finding marketable product alternatives in a product
portfolio that best utilize inventory surplus and replace demand on supply-constrained
products, and have highlighted business benefits through simulations with realistic
production data. Our numerical results point out that more flexible customers are
more profitable customers. Market intelligence and data analytics can identify these
more flexible customers via market models. The integration of marketing insight with
the number and types of sales recommendations are the key to fully attaining these
results and are beyond the ability of this simulation construct. For example, a very
price-sensitive client may only be presented with two sales recommendations—both
of which are alternative-sells or one alternative sell and one down sell. A more price
insensitive client may be presented with five dynamic sales recommendations—three
are up-sells and two are alternative sells (no down sells). This stratification of clients
by price sensitivity and the approach to dynamic sales recommendations will be essen-
tial to achieving the business results we have identified. Moreover, the use of the ATS
model as an intelligent and dynamic engine for sales recommendations on the Internet
and for sales professionals, and the integration of the ATS output with sales activities
will be imperative for attaining a sustainable competitive advantage.
The models featured in this paper have already contributed to substantial business
improvements in real-world supply chains. In 2002 IBM has implemented a hybrid
ATS/ATP process in its complex-configured server supply chain. In this implementa-
tion, a conventional ATP model is executed first, and the ATS model is run subsequently
against the remaining unfilled demand and leftover component supplies. The hybrid
model is focused on managing inventory excesses and overages, i.e., finding saleable
product offerings that consume aged supplies that might otherwise be salvaged or sold
in secondary markets. The realized savings include a $100M reduction of inventory
write-offs in the first year of implementation, and over $20M reduction annually in the
subsequent years. While this hybrid implementation does not deliver the full potential
of ATS business benefits, it does provide management with a control lever ensuring
that there is sufficient supply for core products.
Future work requires the integration of the ATS engine with demand and supply
processes and market intelligence data and applications. This is important when large
product portfolios are in place and automation is necessary for speed and accuracy
of calculations. The benefit of an integrated process and application architecture is
to migrate ATS closer to sales execution and allowing automation to make the com-
munication and presentation of credible sales recommendations a push self-service
capability (instead of a pull) for customers. This will minimize the amount of effort
of the customer and supporting sales staffs as automation masks the complexity of
the product portfolio from the client and business considerations such as profitabil-
ity from the sales teams, and presents viable product alternatives that have attractive
price-performance characteristics.
The major prerequisite to integrate ATS into the process and application architec-
ture is a robust market intelligence capability. The rationale for this dependency is
to identify the customer’s flexibility concerning their first product selections or ame-
nability to another set of sales recommendations. As the customer’s flexibility range
is identified in the various types of market intelligence models such as propensity
to buy and share of the wallets profiles, this business insight into customer buying
behavior can be exploited by ATS modeling by proposing credible and dynamic sales
recommendations based on the customer’s buying characteristics. The flexibility of
customers and their spending patterns are highly relevant inputs into the ATS and
may determine the sequence and number of sales recommendations presented to the
customer based on enterprise business rules.
Acknowledgements The authors thank the two anonymous referees for their exceptional efforts which
helped significantly improve the content and presentation of the paper.
References
Akcay Y, Xu S (2004) Joint inventory replenishment and component allocation optimization in an assem-
ble-to-order system. Manage Sci 50:99–116
Balakrishnan A, Geunes J (2000) Requirements planning with substitutions: exploiting bill-of-materials
flexibility in production planning. Manuf Serv Oper Manage 2(2):166–185
Balakrishnan A, Geunes J (2003) Production planning with flexible product specifications: an application
to specialty steel manufacturing. Oper Res 51(1):94–112
Balakrishnan A, Xia Y, Zhang B (2005) Shaping demand to match anticipated supply. In: MSOM Confer-
ence 2005, Northwestern University
Ball MO, Chen CY, Zhao ZY (2004) Available to promise. In: Simchi-Levi D, Wu SD, Shen ZJ (eds)
Handbook of quantitative supply chain analysis—modeling in the e-business era. Kluwer, New York,
pp 447–480
Bassok Y, Anupindi R, Akella R (1999) Single-period multiproduct inventory models with substitution.
Oper Res 47(4):632–642
Cecere L, Hofman D, Martin R, Preslan L (2005) The handbook for becoming demand driven. AMR
Research
Chen C-Y, Zhao Z, Ball MO (2002) A model for batch advanced available-to-promise. Prod Oper Manage
11:424–440
Chen-Ritzo C-H (2006) Availability management for configure-to-order supply chain systems. PhD
Dissertation, Pennsylvania State University
Dietrich B, Connors D, Ervolina T, Fasano JP, Lougee-Heimer R, Wittrock R (2005) Applications of implo-
sion in manufacturing. In: An C, Fromm H (eds) Supply chain management on demand. Springer,
Heidelberg, pp 97–115
Ervolina T, Dietrich B (2001) Moving toward dynamic available-to-promise. In: Gass S, Jones AT (eds)
Supply chain management practice and research: status and future directions, pp 1–19
Gallego G, Katircioglu K, Ramachandran B (2006) Semiconductor inventory management with multiple
grade parts and downgrading. Prod Plann Cont 17(7):689–700
Hale W, Pyke DF, Rudi N (2001) An assemble-to-order system with component substitution. Working
Paper, Tuck School of Business at Dartmouth
Hopp WJ, Roof Sturgis M (2001) A simple, robust lead time-quoting policy. Manuf Serv Oper Manage
3(4):321–336
Hopp W, Spearman M (2000) Factory physics. 2nd edn. McGraw-Hill/Irwin, New York
Kilger C, Schneeweiss L (2000) Demand fulfillment and ATP. In: Stadtler H, Kilger C (eds) Supply
chain management and advanced planning: concepts, models, software, and case studies. Springer,
Heidelberg, pp 179–196
Lee YM (2006) Analyzing the effectiveness of availability management process. In: Jung H, Chen FF, Jeong
B (eds) Trends in supply chain design and management: technologies and methodologies. Springer,
Heidelberg
Moses S, Grant H, Gruenwald L, Pulat S (2004) Real-time due-date promising by build-to-order environ-
ments. Int J Prod Res 42(20):4353–4375
O’Marah K, Souza J (2004) DDSN: 21st century supply on demand. AMR Research
Swaminathan JM, Tayur SR (1998) Managing broader product lines through delayed differentiation using
vanilla boxes. Manage Sci 44(12):S161–S172
Yunes TH, Napolitano D, Scheller-Wolf A, Tayur S (2007) Building efficient product portfolios at John
Deere and Company. Oper Res 55(4):615–629
Inventory reservation and real-time order promising
in a Make-to-Stock system
Richard Pibernik · Prashant Yadav

OR Spectrum (2009) 31:281–307
DOI 10.1007/s00291-007-0121-4
Abstract In this paper we consider a Make-to-Stock order fulfillment system facing

random demand with random due date preferences from two classes of customers. We
develop an integrated approach for reserving inventory in anticipation of future order
arrivals from high priority customers and for order promising in real-time. Our research
exhibits three distinct features: (1) we explicitly model uncertain due date preferences
of the customers; (2) we consider multiple receipts in the planning horizon that can be
utilized to fulfill customer orders; and (3) we choose to utilize a service level measure
for reserving inventory rather than estimating short- and long-term implications of
order promising with a penalty cost function. We propose an algorithm that exploits
the time structure in order arrivals and time-phased material receipts to determine
inventory reservations for high priority orders. Numerical experiments are conducted
to investigate the performance and the benefits of the inventory reservation and order
promising approach under varying system parameters.
Keywords Make-to-Stock · Inventory reservation · Order fulfillment ·

Order promising · Inventory rationing
R. Pibernik (B)
European Business School, International University Schloss Reichartshausen,
Supply Management Institute, Wiesbaden, Germany
R. Pibernik · P. Yadav
MIT-Zaragoza International Logistics Program,
Zaragoza Logistics Center, Zaragoza, Spain

170 R. Pibernik, P. Yadav
1 Introduction
Today’s Enterprise Resource and Advanced Planning Systems include order fulfill-
ment modules that play a critical role in efficiently matching a company’s supply and
demand in the short term. Based on inventory on-hand, planned receipts and prior order
commitments, these systems calculate the available-to-promise (ATP) quantities and
allocate them to incoming customer orders. By processing orders, promising due dates
and (pre-) allocating them to available inventory and capacity, they not only impact the
customer service but also have significant influence on scheduling and execution of
manufacturing and logistics activities. Because companies are increasingly realizing
the importance of earmarking a portion of the ATP quantities for future (uncertain)
demand of important customers, software vendors are enhancing their order fulfill-
ment modules by what is termed as “allocation planning” (Meyr 2007). We extend the
scope of allocation planning in order fulfillment by (1) developing a novel approach
to inventory reservation for future uncertain demand of important high priority cus-
tomers, and (2) by integrating this approach with real-time order promising. Consider
the following industry example that motivates our analysis:
An Original Equipment Manufacturer (OEM) of DSL telecommunications equip-
ment sources all of its products from Electronic Manufacturing Service (EMS) pro-
viders with manufacturing facilities located in low cost production regions. The OEM
uses a three month rolling forecast from its customers to determine replenishment
orders. Once the orders are placed, the EMS provides firm delivery dates with a lead
time of eight weeks. In many instances, there is a significant deviation between the
forecast provided by the customers to the OEM and the actual orders received. This
often results in situations wherein the OEM anticipates that the available inventory
will not be sufficient to fulfill all customer demand on time. Due to the long replen-
ishment lead times, however, he has no way to remedy this anticipated shortage by
re-ordering. In order to maintain the on-time delivery performance promised to some
of the key customers, the OEM resorts to reserving a portion of the available inventory
only for these high priority customers. Determining these reservations is, however, a
difficult task, given that multiple inventory receipts are scheduled and future demand
is uncertain with respect to order quantities and the required due dates. Also, the OEM
faces the dilemma that if he reserves too much inventory, he risks not being able to
meet orders of other customers with inventory that was in fact available in the ware-
house. Alternatively, if he reserves too little, he may not be able to satisfy the targeted
on-time delivery performance for the high priority customers.
The problem described in our example is faced by many companies dealing with
long product replenishment lead times and stringent requirements on due date delivery
performance from some of their customers. The analysis in this paper attempts to lend
insight into this “inventory allocation/reservation” problem. We develop an approach
to inventory reservation with the objective of enhancing a manufacturer’s ability to ful-
fill uncertain future demand from high priority customers while keeping the fulfillment
of low priority orders at a reasonable level. We assume that long term consequences of
not meeting customer due date requirements cannot be accurately estimated through
cost based measures and, therefore, suggest that inventory is reserved with the objec-
tive of attaining a specific service level for high priority customers. The reservation
Inventory reservation and real-time order promising in a Make-to-Stock system 171
quantities become the input to order promising. We assume that order promising is
performed in real-time, i.e. order acceptance/rejection decisions are made instantly at
the time an order arrives. The order promising engine (that decides upon acceptance or
rejection of incoming orders and determines order due dates) can utilize the reserved
inventory only for orders placed by high priority customers. The unreserved portion
can either be dedicated exclusively to low priority orders (non-nested reservations) or,
more realistically, can also be made available to any high priority demand exceeding
the reserved quantities (nested reservations). Our approach is distinct with respect to
the time structure in the uncertain order stream and the inventory receipts across time.
As described in our motivating example, we consider the case when multiple inventory
receipts are scheduled to occur during the planning horizon. The time interval between
any two inventory receipts constitutes a cycle for which reservation quantities have to
be determined based on aggregate uncertain demand due in this time interval. Due to
the cumulative nature of inventory, these reservations are not independent across time.
We develop a simple algorithm that utilizes aggregate demand forecasts and scheduled
inventory receipts to compute interrelated inventory reservation quantities and show
how these can be used for promising orders in real-time.
Through numerical analyses we identify factors determining inventory reservation
quantities and provide insights and recommendations on how to set relevant system
parameters of an order fulfillment system. Our analysis provides valuable insight into
a fundamental problem described in our initial example: ensuring a target service level
for future high priority orders comes at the expense of a decrease in the system’s over-
all fulfillment performance. We analyze the non-linear trade-off between the service
level for uncertain future high priority demand and the expected loss in the overall
system’s fulfillment caused by inventory reservation. We also show how this relation-
ship is impacted by exogenous parameters such as the number of scheduled receipts
in the planning horizon and the customer due date requirements.
The remainder of this paper is organized as follows. In Sect. 2 we review the litera-
ture related to our research. In Sect. 3 we present our inventory reservation and order
promising approach. Key results and findings of our numerical analysis are presented
in Sect. 4. In Sect. 5 we provide conclusions and managerial insights.
2 Literature review
The research in this paper relates to two different streams of literature: order fulfill-
ment and inventory rationing for multiple demand classes. In this section we briefly
review the previous work in these areas and contrast the contributions of our work.
Many researchers examine the impact of order allocation on profit, cost or service
level measures in a Make-to-Order (MTO) or Assemble-to-Order (ATO) environment.
Ball et al. (2004) provided an excellent review of work in this area. They present mixed-
integer programming models for allocating orders arriving within a pre-determined
time interval to available components and assembly capacity in an ATO setting. The
objective of such models is to determine an allocation which maximizes (short-term)
profit calculated on the basis of per unit revenue of different demand classes, pro-
duction cost, inventory holding cost and penalties for order rejection and capacity
under-utilization (see also Chen et al. 2001, 2002). They propose a “push–pull-frame-
work” for available to promise, which exhibits similarities to the structure of the model
we develop in this paper. Push-based available to promise allocates available resources
to forecasted demand of different customer classes. Pull-based available to promise
allocates received orders to available resources. For push-based available to prom-
ise, the authors propose a deterministic optimization model for allocating available
capacity and material to forecasted demand of different customer classes. The authors
also suggest employing stochastic models for determining booking levels for different
customer classes and relate push-based available to promise to revenue management.
Fleischmann and Meyr (2003) developed linear and mixed integer programming
models for allocating orders to finished goods inventory as well as available resources
of an ATO system. Allocation decisions are based on a penalty cost parameter that
simultaneously captures the costs for backlogging, early allocation and order denial.
The proposed approaches are myopic in nature: orders received in a certain time period
are allocated without considering the impact on the fulfillment of orders arriving in
future. The authors acknowledge that resource reservation should be addressed in more
detail: “Further research effort has to be put on the allocation mechanisms that assign
the projected production quantities [. . .] to different order classes [. . .]”. Meyr (2007)
provides deterministic model formulations for the allocation of available to promise
inventory to a given number of different customer classes (allocation planning) and
also explores approaches to determine the optimal number of customer classes. Kilger
and Meyr (2008) provide an overview of the order allocation functions provided by
commercial order fulfillment systems. They point out that order allocation is typically
performed on the basis of a set of heuristic allocation rules that can be customized to
accommodate company specific requirements. Pibernik (2006) reviewed several basic
mechanisms for allocating scarce inventory to customer orders and evaluates their
performance in a situation where inventory availability is severely constrained. With
a case study from the pharmaceutical industry he illustrates that the logic currently
employed in order fulfillment systems does not enable a company to adequately con-
sider customer priorities when deciding upon order acceptance/rejection and quoting
due dates.
Whereas order fulfillment research mainly focuses on a deterministic setting in
which a set of orders has to be allocated to available resources, inventory rationing
research addresses the allocation of available inventory to uncertain future demand and
usually considers the replenishment quantities to be endogenous. Research on inven-
tory rationing between customer classes is motivated by the following conception: if
finished goods inventory is scarce (and production capacity is constrained), it may
be reasonable to reject demand from less valuable classes in anticipation of demand
from higher value classes. In inventory literature, the early work of Topkis (1968)
considered how inventory should be allocated between demand classes. Each demand
class is characterized by a different shortage cost and the trade-off involves comparing
the benefit of filling demand for low class items in the current period vs. reserving
the available inventory to fill potential demand for higher class items in subsequent
periods. Nahmias and Demmy (1981) evaluated fill rates for given rationing and reor-
der levels in a (Q, r) inventory system with Poisson demand and two demand classes.
More recent examples of work in characterizing the optimal inventory rationing policy
are Ha (1997a,b) and Vericourt et al. (2000). Optimal control policies and rationing
levels are determined on the basis of holding and backorder cost/lost sales cost in these
papers. Benjaafar et al. (2004) extended these approaches to account for multiple prod-
ucts and facilities. For a single product inventory system, Hariharan and Zipkin (1995)
considered uncertainty in the due date preferences of the customers. They term the
time from a customer’s order arrival to the desired due date as demand lead time. For
a simple setting, they show that the effect of an increase in the uncertain demand lead
time is precisely the same as an equivalent reduction in the uncertain supply lead time.
The research presented in this paper is closely related to “push-based available to
promise”, introduced by Ball et al. (2004). We extend the approach of allocating avail-
able to promise quantities of finished goods to future demand. As a major distinction,
we consider demand to be stochastic with respect to the time of arrival, the due date
preferences and the customer class. We consider multiple receipts across the planning
horizon and combine elements of inventory rationing and class booking limits known
from revenue management. In contrast to more recent inventory rationing research, we
assume exogenous replenishment quantities. Also, we consider inventory reservation
as a measure for maintaining responsiveness of an order fulfillment system to prevent
the negative long-term consequences resulting from not meeting delivery date prefer-
ences of high priority customers. As long-term implications of not meeting customer
preferences are extremely difficult to capture, we derive inventory reservations on the
basis of a disaggregated service level for high-priority customer demand. Therefore,
our approach is not based on maximizing short term profits or revenues generated
from different demand classes but on a service level as measure for the responsiveness
of an order promising system in regard to future demand of high priority customers.
3 The model
We consider a facility of a MTS manufacturer that carries inventory of a single prod-

uct. We assume a planning horizon spanning over T periods. Typically, the periods
t ∈ {1, . . . , T } represent working days during which orders are being received and for
which delivery dates of individual orders are quoted. In any period t the manufacturer
faces uncertain order streams for a single product from two customer classes.1 Any
order j, arriving in period t can be characterized by a required due date d j ∈ {1, . . . , T }
(specified by the customer) and the customer class (high or low priority).
We assume that the manufacturer employs an automated order fulfillment system
which handles orders and promises due dates in real-time, i.e. in the sequence of their
arrival. The system keeps track of the uncommitted, cumulative inventory quantities
in the planning horizon, the “available to promise” (ATP) quantities. We assume that
the replenishment schedule is frozen throughout the planning horizon (i.e., inventory
receipts are fixed) and thus the ATP quantities are only dependent on the number of
customer orders promised. When an order j with required due date d j arrives, the
1 We assume that demand occurs one unit at a time and orders for more than one unit can be split into
single unit orders. Hereafter, we use the term “order” to refer to a single unit of demand.
Forecast Demand Distribution Determine Available Inventory
Determine Inventory Reservations
Update Update
& Order Promising &
Roll-over Roll-over
Fig. 1 Inventory reservation and order promising framework
system computes the ATP quantities in the planning horizon. If it finds a positive ATP
quantity in period d j , the order is accepted; else the system rejects the order.
To ensure fulfillment of high priority orders in case of inventory scarcity, the manu-
facturer wants to protect some portion of the ATP quantities from being consumed by
low priority orders. As described previously, we assume that the manufacturer wants to
reserve ATP quantities to ensure that a pre-determined target service level is achieved
for high priority orders. Many variations of service level measures continue to be used
in inventory theory reflecting the different types of costs that arise from not being able
to meet demand in time (for a more involved discussion see for example Ronen 1982,
1983, as well as Silver et al. 1998). Although it is difficult to quantify these costs,
knowing their nature helps determine an appropriate service level measure. In our
case we assume the cost associated with not being able to meet high priority customer
demand to be a fixed cost; hence we choose to work with an α-service level. This also
corresponds to our observations for the case example of the DSL OEM described in
Sect. 1. Note, however, that our approach is not limited to an α-service level; reser-
vation quantities could, for example, also be based on a pre-defined fill rate target.
Admittedly, the choice of the target service level is an important decision problem by
itself; in this section we assume an exogenously set α. In Sect. 4.2 we provide insights
into the problem of setting α.
Figure 1 gives an overview of our modeling framework and the individual activities
in our inventory and order promising approach.
At the beginning of period t = 1, the manufacturer forecasts the demand and
due date distribution for periods t = 1, . . . , T . Based on the demand forecast, calcu-
lated available inventory and the pre-defined α-service level, reservation quantities are
computed for high priority orders throughout the planning horizon. The reservation
quantities are used as an input to order promising. We assume that the reservations
for the whole planning horizon are frozen during a period which is determined by the
roll-over interval. During this frozen time period, orders arrive and are either accepted
and quoted their required due date, or rejected. Orders from low priority customers
can only utilize the remaining, unreserved portion of the available inventory. After the
roll-over interval has elapsed, the inventory reservations are unfrozen, available inven-
tory is updated based on orders committed and any new information about arrivals of
inventory receipts are incorporated into the subsequent calculation of available inven-
tory. Also, the demand forecast may be updated to accommodate any new demand
information received. Finally, the planning horizon is rolled-over and the procedure
repeated. The length of the roll-over interval is determined by the time until the next
inventory receipt occurs. In Sect. 3.3 we show that interrelated reservation quantities
have to be determined for individual inventory cycles. These cycles are given by the
points in time for which subsequent inventory receipts are scheduled.
In the following sub-section we first describe how we model the order arrival pro-
cess for future demand (forecast of the demand distribution). In Sect. 3.2, we show
how the available inventory is calculated and outline the real-time order promising
mechanism. Thereupon we introduce our approach to reserving inventory for high
priority orders.
3.1 Demand and due date distribution
The manufacturer receives orders for a single product from two classes of customers
(high and low) that are uncertain with respect to the time of arrival within period t, and
the due date preference. To model the uncertainty in due date preferences, we assume
that each order has a random demand lead time of (independent of the arrival time)
that follows a discrete probability distribution g( = τ ) with cdf G. We assume that
the distribution of has a bounded support (τb , τe ) such that for some reasonable
τe , g(τe ) > 0 and G(τe ) = 1. Thus, an order arriving in time period t with a demand
lead time of τ has a due date preference d = t + τ . d = t + τb is the earliest due date a
customer can request for any order placed in period t. τb is typically governed by the
lead time for transportation and order picking. By (Q kt,d (x) : x ≥ 0) we denote
the count process that represents the demand of class k ∈ {H, L} orders arriving in
the interval (0, x] in time period t with a desired due date of d. We assume that the
count process has finite moments and also has stationary and independent increments
(Lévy process). The number of orders that arrive in any period t with a due date
k (·).
d can then be denoted by Q kt,d with its corresponding distribution function Ft,d
The knowledge of Ft,d (·) alone is sufficient for the modeling approach presented in
k
Sect. 3.3. However, we present here the underlying Lévy process (Q kt,d (x) : x ≥ 0)
for reasons that will become clear in Sect. 3.4.
3.2 Real time order promising
At the beginning of the planning horizon, the inventory on hand, denoted by inv 0 ,
as well as planned receipts in periods t ∈ {1, . . . , T }, denoted by st , are known. We
assume that receipts occur at the beginning of the period so that these additional quan-
tities can be fully utilized in the period of their arrival. Without loss of generality, a
delivery lead time of zero is assumed. By inv t we denote the available inventory in
period t. It can be calculated as

t t
inv t = inv 0 + sω − cd , t = 1, . . . , T, (1)
ω=1 d=1
where cd represents inventory quantities committed to orders accepted previously
with due date d ∈ {1, . . . , T }. We assume that committed orders are binding; their
promised due date cannot be changed in favor of new incoming orders. Given inv t
we can determine the ATP quantities, denoted by at p d , by setting at p d = inv T for
d = T and recursively computing

at p d = min inv d , at p d+1 . for d = T − 1, . . . , 1. (2)
Immediately after order j arrives, at p d j is calculated. If at p d j > 0, the order is

accepted and due date d j is quoted; cd j is updated and inv t is recalculated for the
remaining planning horizon. We assume that the order is rejected if at p d j = 0.
3.3 Inventory reservation without nesting
In this section we show how inventory reservations for future high priority demand
can be determined for the non-nesting case, i.e. when high priority orders can only be
fulfilled from reserved inventory and any high priority demand exceeding the reserved
portion will be rejected.
When determining reservation quantities the time structure of the order arrivals,
due date preferences and the time-phased inventory availability have to be taken into
account. We begin to outline our reservation approach for the simplest case where no
inventory receipts occur during the planning horizon. We further assume that the next
receipt will occur in period T + 1, i.e. in the first period after our planning horizon.
Without loss of generality we assume that at the beginning of the planning horizon
cd = 0 for all d ∈ {1, . . . , T }.
Because no receipts are scheduled to occur during the planning horizon, the avail-
able inventory can be considered as a single resource for which high and low priority
orders compete. Therefore, we do not need to distinguish different time periods in the
planning horizon and can define inv = inv 0 as the inventory available throughout
the planning horizon. By r we denote the portion of the available inventory the man-
ufacturer wants to reserve for high priority
demand with a required due date d ≤ T .
T T
For the sake of notational ease let Q H = t=1 H
d=t Q t,d . Given the manufacturer’s
service level objective, r has to be chosen to satisfy constraint (3).

Pr Q H ≤ r ≥ α. (3)
Clearly, determining r is equivalent to determining a safety stock level that ensures a

probability α of not stocking out in the lead time. Constraint (3) can be written as
r ≥ FQ−1H (α).2 (4)
Given the service level objective, the manufacturer wants to choose the minimum
reservation quantity that satisfies (4). The remaining inventory can then be utilized by
low priority orders. Considering that the available inventory may not suffice to satisfy
(4), the minimal (and inventory feasible) reservation quantity
2 For notational simplicity we use F −1 (α) to denote min F(X ) ≥ α.

X x

r ∗ = min FQ−1H (α), inv .
The order promising mechanism can easily be modified to account for inventory
reservations. The available inventory inv is split into inv H = r ∗ −c H for high priority
orders and inv L = inv 0 − r ∗ − c L for low priority orders. By c H and c L we denote
inventory committed to high priority and low priority orders respectively. The com-
mitted quantity is updated after an order has been accepted. Any high (low) priority
orders will be accepted as long as inv H > 0 (inv L > 0).
Now consider the case where an additional material receipt is scheduled for period
t (1 < t ≤ T ). At the beginning of the planning horizon the available inventory is
given by
⎧
⎪
t
⎪
⎨ inv 0 − cd for t = 1, . . . , t − 1,
inv t = d=1
(5)
⎪
⎪
t
⎩ inv 0 + st − cd for t = t , . . . , T.
d=1
In this case, the manufacturer has to determine two (interrelated) reservation quanti-
ties. Although in effect only the inventory reservation for the first cycle is implemented,
it is still important to determine the inventory reservation for the second cycle. This
prevents low priority orders arriving in the first cycle from consuming inventory which
would still be required to achieve the target service level for high priority orders in the
second cycle. Also, as some of the inventory reserved in the first cycle may remain
unused, it needs to be accounted for in computing the inventory reservation for the
second cycle. By r1 and r2 we denote the quantity of available inventory reserved
for high priority demand in periods t = 1, . . . , t − 1 and periods t = t , . . . , T ,
respectively. To simplify our notation we define the demand due before period t as
t −1 t −1 H
Q 1H = t=1 d=t Q t,d and the demand with a required due date later than t as
T T
Q 2H = t=t H
d=t Q t,d . Following our approach of choosing a reservation quantity
based on the probability of fulfilling all high priority demand in the lead time, r1 has
to be set to satisfy the service level constraint (6):

Pr Q 1H ≤ r1 ≥ α. (6)
r2 cannot be determined independent of r1 . When determining r2 we have to take into

+
account that some random quantity r1 − Q 1H may remain unused by demand with
+
a due date requirement before t and that both r1 − Q 1H and r2 can be utilized to
fulfill demand Q 2H (r2 however cannot be utilized by Q 1H ). Considering this, we can
write service level constraint (7) for r2 :
+
Pr Q 2H ≤ r2 + r1 − Q 1H ≥α (7)
Ignoring inventory availability constraints, we can determine r1 and r2 by first setting

r1 = FQ−1H (α) independently and then solving
1
+
arg min Pr Q 2 H
− r1 − Q 1
H
≤ r2 ≥ α (8)
r2
to determine r2 . Because the expression involves a convolution of a random variable

with another truncated random variable, we cannot derive a closed form solution to
(8). We employ a two-moment approximation to a normal distribution for the distri-
+
bution of Q 2H − r1 − Q 1H which is commonly used in inventory literature (see, for
example, Zipkin 2000, pp. 304–305).
From (5) we know that r2 has an upper bound st . If the material receipt st is
less than the required reservation quantity r2 , determined from (8), some additional
portion of inventory inv 0 has to be reserved to ensure that the probability of fulfilling
high priority demand Q 2H is at least α. This implies that r1 has to be increased in case
r2 > st . The choice of r1 then has to be based on random demand Q 1H and the random
+
shortfall Q 2H − st . Thus, if r2 > st we first set r2 = st and then solve
+
arg min Pr Q 1H + Q 2H − st ≤ r1 ≥ α (9)
r1
to determine r1 .
From the previous analysis we can derive the following procedure to determine
reservation quantities r1∗ and r2∗ for the case of one receipt in the planning horizon:
1. r̄1 = min(inv 0 ; FQ H (α))

1 +
2. r̄2 = arg min Pr Q 2H − r̄1 − Q 1H ≤ r2 ≥ α
r2
3. If r̄2 ≤ st then r1∗ = r̄1 and
r2∗ = r̄2 ; End
+
4. Else: r2∗ =s t and r1∗ = min inv 0 ; arg min Pr Q 1H + Q 2H −s t ≤ r1 ≥ α
r1
Knowing r1∗ and r2∗ it is straightforward to compute the available inventory for high
and low priority orders:
⎧
⎪
t
⎪ ∗
⎨ r1 − cdH for t = 1, . . . , t − 1
inv tH = d=1
(10)
⎪
⎪
t
⎩ r1∗ + r2∗ − cdH for t = t , . . . , T
d=1
⎧
⎪
t
⎪ ∗
⎨ inv 0 − r1 − cdL for t = 1, . . . , t − 1
inv tL = d=1
(11)
⎪
⎪
t
⎩ inv 0 + st − r1∗ − r2∗ − cdL for t = t , . . . , T
d=1
Given inv tH and inv tL we can determine separate available to promise quantities at p dH
and at p dL for any due date d as shown in conjunction with Eq. (2) and perform real-time
order promising as described in Sect. 3.2.
We now generalize this approach for n > 1 scheduled receipts. Let
n = |{st > 0 |t ∈ {2, . . . , T } }| denote the number of receipts in the planning hori-
zon. With n receipts,
reservation
quantities
have to be determined for n + 1 cycles
1, t 1 − 1 , t 1 , t 2 − 1 , . . . , t n , T , with t i denoting the period in which the ith
receipt occurs. We denote by ri the reservation quantity for cycle i (i = 1, . . . , n + 1)
and by Q iH the high priority demand with a required due date in cycle i.
Following our approach for the one receipt case, we start by calculating the res-
ervation quantity for the first cycle and set r̄1 = min(inv 0 ; FQ H (α)). We define by
+ 1
L 1 = r̄1 − Q 1H the random reservation quantity of cycle i = 1 that remains
unused, i.e. the quantity that can be utilized by high priority demand with a due date
requirement in subsequent cycles. Given r̄1 and L 1 we can employ a simple recursive
procedure to calculate the reservation quantities for cycles i = 2, . . . , n + 1:

r̄i = arg min Pr Q iH − L i−1 ≤ ri ≥ α, (12)
ri
+
where L i = L i−1 + ri − Q iH . (13)
From the one receipt case we know that r̄2 , . . . , r̄n+1 are bounded from above by
st 1 , . . . , st n . Therefore, the reservation quantities calculated from (12) are only feasible
iff r̄i ≤ st i−1 for i = 2, . . . , n + 1.
As before, if r̄i > st i−1 then r̄i ← st i−1 , and r̄i−1 has to be recalculated. From (9)
we obtain
+
r̄i−1 ← arg min Pr Q i−1H
+ Q iH − st i ≤ ri−1 ≥ α. (14)
ri−1
However, the recalculated r̄i−1 is only feasible iff r̄i−1 ≤ st i−2 . Otherwise, r̄i−1 ←
st i−2 and
+ +
r̄i−2 ← arg min Pr H
Q i−2 + H
Q i−1 − st i−1 + Q iH − st i ≤ ri−2 ≥ α. (15)
ri−2
This procedure is continued for r̄i−2 , . . . , r̄1 . Note that through this procedure, any
“shortfalls” are transferred into earlier periods. Consequently, inventory feasibility
can be determined based on the reservation quantity r̄1 computed for the first cycle.
We know that the reservation quantities determined from recursively solving (14) and
(15) are only inventory feasible iff inv 0 ≥ r̄1 . In cases where r̄1 ≥ inv 0 , the reserva-
tion quantities are (r1∗ , r2∗ , . . . , rn+1
∗ ) = (inv , s , . . . , s ). Based on the developed
0 1 n
expressions we can derive the following algorithm for computing inventory feasible
reservation quantities for the n-receipt case:
1. Set r̄1 = min(inv 0 ; FQ H (α)); determine L 1

1
2. For i = 2, . . . , n + 1,

a. r̄i = arg min Pr Q iH − L i−1 ≤ ri ≥ α (from (12))
ri
+
b. L i = L i−1 + ri − Q iH (from (13))
c. For κ = i, . . . , 1 (from (14) and (15))

If r̄κ > st κ−1
r̄κ ← st κ−1
i

+
r̄κ−1 ← arg min Pr H
Q κ−1 + Q ωH − st ω ≤ rκ−1 ≥α
rκ−1 ω=κ
Else : Next i
Next κ
Next i
3. Set r1∗ = min (inv 0 , r̄1 ); ri∗ = r̄i for i = 2, . . . , n + 1; End

Given r1∗ , . . . , rn+1
∗ we can calculate the available inventory quantities for high and
low priority orders for all periods of the planning horizon.

i
t
inv tH = rω∗ − cdH ∀t ∈ t i−1 , t i ;
ω=1 d=1
i = 1, . . . , n + 1; t 0 = 1; t n+1 = T (16)

t
i t
inv tL = inv 0 + sω − rω∗ − cdL
ω=1 ω=1 d=1

∀t ∈ t i−1 , t i ; i = 1, . . . , n + 1; t 0 = 1; t n+1 = T (17)
Based on inv tH and inv tL we can again determine available to promise quantities at p dH
and at p dL for any due date d based on Eq. (2). at p dH and at p dL can be used for real-time
order promising as described in Sect. 3.2.
3.4 Inventory reservation with nesting
The analysis in the previous section shows how reservations can be determined by
clearly partitioning a portion of the inventory for high priority orders. This helps our
understanding of the general approach to reservation in the setting considered here.
However, the more practically interesting problem is how to determine inventory reser-
vations when the orders from high priority customers, besides the exclusively reserved
inventory, can also utilize the unreserved portion. This is similar to nested protection
levels used in airline and other yield management applications.
We again begin by first considering the simpler case with no receipts occurring
in the planning horizon. As before, we reserve r only for high priority orders and
once (and if) high priority orders have exhausted this quantity, they compete equally
with the low priority orders for the remaining inventory. Computing the expression
for the probability that all high priority orders are fulfilled is now more involved as it
requires explicitly considering the time structure of the order arrival process from the
two customer classes.
The probability that the reserved inventory itself is sufficient to fulfill all high
priority orders arriving in the planning horizon is Pr{Q H ≤ r }. When the reserved
inventory r is not sufficient to meet all high priority orders, any excess demand can
also be met from the remaining unreserved inventory. By (Q i (x) : x ≥ 0) we de-
note a count process to represent the cumulative demand observed in the continuous
time interval [0, x] where x ∈ [0, lT ]. Suppose that the reserved inventory is ex-
hausted at time V ∈ [0, lT ]. The inventory available to meet joint demand occurring
in the time interval (V, lT ], i.e., after the reserved inventory has been exhausted, is
+
inv 0 − r − Q L (V ) , where the last term Q L (V ) accounts for the inventory con-
sumed by the low priority orders up to time V . Suppose the last high priority customer
order occurs at time W ∈ (V, lT ]. Then all high priority customer orders will be
fulfilled iff
inv 0 − r − Q L (V ) − Q H (W − V ) − Q L (W − V ) ≥ 0. (18)
Combining the two cases, the conditional probability of high priority fulfillment is
given by the following function:

ζ (inv 0 , r |V, W ) =
1 i f V ≥ lT (19)
Pr Q L (W ) + Q H (W − V ) ≤ inv 0 − r i f V < lT
(19) can be simplified and written as

ζ (inv 0 , r |V, W ) = Pr Q H (lT ) ≤ r + Pr Q H (lT ) > r

× Pr Q L (W ) + Q H (W − V ) ≤ inv 0 − r . (20)
By the independent increment property of the assumed stochastic process for order
arrival we can rewrite Q H (W − V ) as Q H (W ) − Q H (V ). Also, Q H (V ) = r by
definition of V . Substituting and rearranging terms we can write the unconditional
probability of fulfilling all high priority orders as3
lT
α = Pr Q H
≤r +Pr Q H
>r Pr Q L (w)+ Q H (w) ≤ inv 0 f (W = w)dw
w=0
(21)

3 To avoid trivial issues we assume α ≥ Pr Q L (W ) + Q H (W ) ≤ inv .
0
The first part of the above expression represents the probability of high priority fulfill-
ment without nesting. For brevity, we denote this by α N N . The second term represents
the contribution of nesting to the overall service level α. We can express this as
lT
α−α NN
= Pr Q H
>r Pr Q L (w) + Q H (w) ≤ inv 0 f (W = w)dw (22)
w=0
Next we analyze the case where an additional material receipt is scheduled for period
t (1 < t ≤ T ). As in the previous section, the manufacturer has to determine two
cycle specific reservation quantities r1 and r2 . For the first cycle, the probability α1 of
fulfilling all high priority orders can be calculated as for the case of no receipts. For
brevity we assume that r1 has been determined based on α1 . The inventory available

L + . All high
in the second cycle is inv 0 + st + inv 0 −Q 1 − min inv 0 − r1 , Q 1
H
priority orders with due date d ∈ t , . . . , T will be fulfilled iff

+
inv 0 + st + inv 0 − Q 1H − min inv 0 − r1 , Q 1L − r2 − Q 2L (V2 )
−Q 2H (W2 − V2 ) − Q 2L (W2 − V2 ) ≥ 0. (23)
where V2 corresponds to the time required to observe a cumulated demand of r2 in the

second cycle and W2 denotes the time at which the last high priority customer order in
the second cycle arrives. With Q 2H (W2 − V2 ) being equivalent to Q 2H (W2 ) − Q 2H (V2 )
and Q 2H (V2 ) = r2 we can write the conditional probability of high priority fulfillment
in the second cycle as
ζ2 (inv 0 , r1 , r2 , st |V2 , W2 )
⎧
⎨1 i f V2 ≥ lT

= Q 2H (W2 ) + Q 2L (W2 )−
⎩ Pr + i f V2 < lT
inv 0 − Q 1H − min inv 0 − r1 , Q 1L ≤ inv 0 + st
(24)
From (24) we can derive the unconditional probability α2 of fulfilling all high priority
orders in the second cycle:

α2 = Pr Q 2H ≤ r2 + Pr Q 2H > r2
lT +
· Pr Q 2L (w2 ) + Q 2H (w2 ) − inv 0 − Q 1H − min inv 0 − r1 , Q 1L
w2 =lt

≤ inv 0 + st f (W2 = w2 )dw2 (25)
Clearly, (21) and (25) do not lend themselves for a closed form solution for r1 and r2
and the integration over the random variable W makes it difficult to employ simple
approximation methods as described in Sect. 3.2. For a larger number of inventory

receipts n > 2, the expressions for αi become even more complex to approximate.
Owing to these difficulties, we choose to conduct numerical experiments to quantify
the effect of nesting on the high priority service level for multiple receipts. The results
of these experiments are presented in Sect. 4.1.
4 Numerical experiments
In this section we present results and insights from numerical analyses we conducted
to investigate the impact and performance of our integrated inventory reservation and
order promising approach. For the purposes of the numerical analysis we assume that
the order arrival process is a Poisson process and hence in each time period of equal
length l, orders from high and low priority customers arrive to the system following
Poisson processes with rates λ H and λ L respectively.
The tool we employ for numerical analysis is implemented in Microsoft Excel and
Visual Basic for Applications (VBA) and includes modules for order generation based
on the Poisson arrival process with the general stochastic demand lead time structure
described in Sect. 3.1; inventory reservation as described in Sect. 3.3; and real-time
order promising (with nesting) based on the orders’ time of arrival (as in Sect. 3.2).
All analyses are conducted for a planning horizon spanning over T = 15 peri-
ods. The demand parameters chosen for the simulation experiments are λ H = 20
and λ L = 80. One of the objectives of the numerical analysis is to investigate how
the performance measures of interest behave as we change the preset service level α
and the degree of inventory tightness. For the preset service level (denoted by α N N )
we chose four commonly used factor levels {0.75, 0.90, 0.95, 0.99}. For the over-
all inventory availability in the planning horizon, we consider T eight
T factor levels:
{0.2c, 0.4c, 0.6c, 0.8c, 0.9c, 1.0c, 1.1c, 1.2c}, where c = t=1 d=t λtd denotes
the total mean demand with a required due date in the planning horizon. With respect
to the number of inventory receipts occurring in the planning horizon we consider four
different cases: (1) one inventory receipt in period t = 1(n = 1), i.e., all inventory is
available at the beginning of period t = 1;4 (2) n = 3 receipts; (3) n = 5 receipts;
and (4) n = 15 receipts occurring in every period. For a given n, we assume equal
cycle lengths T /n and equal sizes of material receipts of st i = 0.2c
n , n , . . . , n (i =
0.4c 1.2c
1, . . . , n).
Since our reservation approach is based on an α-service level for high priority
demand due in individual inventory cycles, we measure the relative frequency of
100% high priority fill rate for demand due within an inventory cycle and term it as
the realized service level α̂. This is used as an estimator of the probability of fulfilling
all high priority demand due within an inventory cycle.
To determine the statistical robustness of the sample size for our experiments we
calculated the confidence band for the case of five receipts at a capacity equal to the
mean demand (1.0c) for 500 sample runs. The standard deviation observed from the
test runs was 0.0049. We inferred that with this standard deviation we can obtain good
4 Note that this is identical to the case of no receipt described in Sect. 3.2.
Table 1 Reservation quantities for different levels of α N N
αN N
0.75 0.9 0.95 0.99
i 1 2 3 1 2 3 1 2 3 1 2 3
ri 107 100 99 113 104 102 117 105 105 124 108 107
estimates of our service levels within a bound of ±0.008 with a confidence level of
95% and conducted 500 sample runs for all experiments reported in this chapter.
In Sect. 3.3 we were not able to derive closed form expressions for the probability
of fulfilling all high priority demand within a cycle for the case of nested reserva-
tions. In our first set of analyses (presented in Sect. 4.1) we conduct experiments to
quantify the effect of nesting and to determine whether α N N (the service level target
without nesting) can instead be used as an adequate system parameter for determin-
ing reservation quantities. Thereupon, in Sect. 4.2, we perform analyses to quantify
the impact of inventory reservation on the overall performance of the system. These
experiments specifically address an important trade-off the manufacturer faces when
choosing a target service level for high priority orders: high target service levels for
high priority orders may come at the expense of a significant decrease in the system
performance. Our results provide insights into the relationship between these two per-
formance measures. In Sects. 4.1 and 4.2, we also highlight how the number of receipts
in the planning horizon impacts the effects of nesting and the overall fill rate of the
system. In Sect. 4.3 we explore the effects of different due date distributions on the
performance of our reservation approach. Finally, in Sect. 4.4 we compare the perfor-
mance of our approach to a simple myopic reservation policy in which reservations
are made individually for every cycle without taking carry-over effects into account.
4.1 Effect of nested inventory reservation on the high priority service level
In this section we analyze the impact of reserving inventory for different levels of
inventory availability and explore the effects of nesting. All simulation experiments
presented in this section are conducted for a customer lead time distribution of τe = 0,
i.e., all customers require instant delivery.
We first show results for the case of three receipts in the planning horizon. Thereafter
we point out the effects of a varying number of inventory receipts. In Table 1 we pro-
vide the reservation quantities which result from the reservation approach developed
in Sect. 3.3.
Figure 2 shows the realized service level α̂ vs. preset α N N for different levels of
inventory availability. The realized service level α̂ depends not only on the reservation
quantity (determined by α N N ) but also on the extent of inventory availability in the
system. In a severely constrained scenario, it may constitute an upper bound on the
reservation quantities. Also, the effects of nesting are dependent on the availability
of inventory. At very low inventory availability (0.2c), setting a higher target service
1,0
0,9
0,8
0,7
0,6
0,5
αˆ
0,4
0,3
0,2
0.2c 0.4c 0.6c 0.8c 0.9c 1.0c 1.1c 1.2c
0,1
0,0
0,0 0,75 0,9 0,95 0,99
NN
α
Fig. 2 Realized service level α̂ for different α N N dependent on inventory availability
level for high priority orders does not guarantee the desired service level. The required
reservation quantities exceed the available inventory required in the individual inven-
tory cycles. As more inventory becomes available, it is sufficient to meet the required
reservation quantities and there may also be an unreserved portion of the inventory
available for fulfilling low priority and any excess high priority orders.
However, this unreserved quantity may still not lead to a significant increase of the
high priority service level above α N N ; at the hitting time V , low priority orders have
already consumed most of the unreserved portionof inventory. In such instances there
are little or no benefits from nested reservations (e.g., at inventory levels of 0.6c, 0.8c
and 0.9c). Beyond a threshold level of 0.9c, the unreserved portion is large enough
to also fulfill some of the excess high priority demand. Therefore, at levels of inven-
tory availability beyond this threshold we realize the benefits of nested reservations.
Figure 3 illustrates the effects of nesting dependent on the capacity level for a preset
α N N = 0.75. Beyond the threshold level of 0.9c we start realizing benefits of nesting
(see highlighted area in Fig. 3); they are increasing in inventory availability. It should
be noted that at high levels of inventory availability (1.1c, 1.2c) the differences in α̂
with reservation and without reservation approach zero; the benefits of nesting become
irrelevant as we realize (almost) the same service levels with nested reservations as
without reserving inventory.
The benefits of nested reservations depend not only on the available inventory as
depicted in Fig. 3 but also on the manufacturer’s choice of α N N and the corresponding
reservation quantities. In Fig. 4 we illustrate this effect by varying α N N at the inven-
tory level 1.0c. We can see that at a value of α N N = 0.99 there are almost no effects
of nesting. Below α N N = 0.99, the realized high priority service levels α̂ are higher
1,0
With Nesting Without Nesting
0,9
0,8
0,7
0,6
α̂ 0,5
0,4
0,3
0,2
0,1
0,0
0,1c 0,2c 0,4c 0,6c 0,8c 0,9c 1,0c 1,1c 1,2c
Inventory Availability
Fig. 3 High priority service levels for nested and non-nested reservations (α N N = 0.75; n = 3)
than α N N due to the additional benefits of nesting. The benefits increase as α N N is

decreased.
These results can guide the decision maker in choosing an appropriate α N N . More
specifically, they provide insight into the service level the manufacturer can expect
from setting an α N N , given a certain inventory availability. The results also identify
the levels of inventory and preset α N N at which the manufacturer will experience
positive effects of nesting on the realized service level. In our experiments we observe
that significant effects of nesting only occur if inventory is not very constrained and
if the manufacturer does not aim for very high service levels. From a practical point
of view, however, tight inventory and high target service levels (i.e. α ≥ 0.95)are the
more interesting and relevant scenarios for inventory reservation. For such instances,
α N N alone is an adequate parameter for determining inventory reservation quantities
(even with nesting). Further analyses suggest that this conclusion holds irrespective of
the number of inventory receipts scheduled for the planning horizon. We do observe,
however, that the benefits of nesting increase in the number of receipts in the planning
horizon. In Fig. 5 we give an overview of the benefits of nesting experienced with
alternative numbers of receipts at a relevant inventory availability level of 1.0c and
target service levels of α N N = 0.75 and α N N = 0.9.
In the case of just one receipt, orders are accepted until all remaining inventory
is consumed. From this point on, all orders have to be rejected and there are no fur-
ther effects of nesting. In cases with multiple receipts, however, high and low priority
orders will already be rejected if there is not sufficient inventory available within the
1,00
With Nesting Without Nesting
0,95
0,90
αˆ
0,85
0,80
0,75
0,75 0,9 0,95 0,99
α NN
Fig. 4 Effects of nesting dependent on α N N
0,20
0,15
∆ α̂ 0,10
0,05
0,00
1 3 5 15
Number of Receipts (n)
Fig. 5 Nesting effects dependent on the number of receipts in the planning horizon (α̂ represents the
difference in the realized service level with and without nesting)
specific cycle in which they are due. Also, any remaining inventory gets carried over
to the next cycle. The (Poisson) “race” for the available inventory is re-initiated in
every cycle, providing multiple opportunities to realize positive effects of nesting.
2,0%
0,75 0,90 0,95 0,99
1,8%
1,6%
1,4%
1,2%
fr
1,0%
0,8%
0,6%
0,4%
0,2%
0,0%
0,2c 0,4c 0,6c 0,8c 0,9c 1c 1,1c 1,2
Fig. 6 Impact of inventory reservations on overall fill rate ( f r represents the difference in expected
overall fill rate with and without inventory reservation)
4.2 Effect of inventory reservation on the overall system performance
In the previous section we analyzed the impact of inventory reservations on the high
priority service level and characterized the effect of nesting. When setting a ser-
vice level target for high priority orders, ex-ante the manufacturer needs to consider
the expected impact on the fulfillment of low priority orders. In general, reserving
inventory for high priority orders will decrease the (expected) number of low priority
orders fulfilled. It is reasonable to assume that the manufacturer will always accept
this decrease as long as it is offset by an equal increase in the expected number of high
priority orders fulfilled, i.e. for every rejected low priority order one additional high
priority is fulfilled. However, as the high priority service level is increased through
higher reservation quantities, the probability that some portion of the reserved quantity
remains unused also increases. This unused portion of the reserved inventory could
have been utilized to fulfill low priority orders that are currently getting rejected. The
manufacturer therefore has to consider the trade-off between ensuring a certain service
level for high priority orders and the expected negative impact on overall order ful-
fillment. To measure this negative effect, we utilize the decrease in the overall system
fill rate attributed to reservation. It shows how many additional low priority orders
have to be sacrificed on average to ensure a certain α-service level. Any decrease in
the overall fill rate can be attributed to unused reservation quantities which are not
available to fulfill low priority orders. Note that the two performance measures (high
priority service level and decrease in overall system fill rate) allow the manufacturer
to evaluate the trade-off between the decrease of the “risk” of missing a high priority
order and the associated decrease in the overall system performance. In Fig. 6 we plot
the difference in the overall system fill rate with and without inventory reservation.
We observe that at very low levels of inventory availability (e.g. 0.2c), reservation
causes only a small decrease in the overall fill rate. At these low levels inventory
is not sufficient to reserve the required quantity. As a consequence, only in very few
instances the realized demand is lower than the quantity reserved. Severely constrained
inventory prevents the system from achieving the target high priority service level (see
Fig. 2) but also decreases the negative impact on the overall fill rate. We observe a
sudden increase in f r as sufficient inventory becomes available to achieve the target
service level for high priority orders (0.4c). At this level of inventory availability the
manufacturer has to sacrifice approximately 1.8% of the overall fill rate to ensure a
service level of 0.99 for high priority orders. Note that at these low levels of inven-
tory availability the manufacturer does not benefit from nesting (see Sect. 4.1). As
more inventory becomes available, the negative impact on the overall fill rate slightly
decreases. This can be explained by very few instances in which benefits of nesting
can be utilized to slightly increase the overall fill rate. Up to an inventory availability
of up to 0.9c these effects are, however, only marginal. At the nesting threshold of 1.0c
two effects lead to a significant decrease in f r . In cases of elevated high priority
demand, nesting has a positive impact on the overall fill rate. As seen in the previ-
ous section (see Figs. 3 and 4), notable nesting effects are only realized beyond the
threshold level of 1.0c. Furthermore, when high priority demand is low and some of
the reserved inventory remains unused, the unreserved portion may still be sufficient
to fulfill all low priority demand. For the same reason the negative effect of inventory
reservation approaches to zero at high levels of inventory availability (e.g. 1.1c, 1.2c).
The results of our analysis support the manufacturer in setting an appropriate service
level target for high priority orders. Knowing the “cost” of reservation (measured in
terms of the loss in system fill rate) allows him to evaluate the (non-linear) trade-off
between different target service levels for high priority fulfillment and overall system
performance.
Further experiments with a varying number of inventory receipts indicate that the
negative effect of inventory reservation on the overall fill rate increase in the number
of receipts. In Fig. 7 we plot the loss in fill rate caused by inventory reservation ( f r )
for n = 1,3,5,15 receipts across different levels of inventory availability and a target
service level of 0.95.
Given that the benefits of nesting increase in the number of inventory receipts (see
Fig. 5), it seems counterintuitive that the loss in overall system fill rate decreases in the
number of receipts. In the case of one receipt, all inventory is available at the begin-
ning of the planning horizon and can be used to fulfill high priority orders and all low
priority orders up to the protection level. Pooling effects are realized across time up to
the point at which inventory is completely exhausted. In the case of n > 1 receipts the
available inventory in the planning horizon is allocated to the n + 1 inventory cycles.
In every cycle orders are rejected if the cycle-specific inventory is not sufficient to
fulfill all orders due in this cycle. Through this “demarcation” of inventory quanti-
ties less demand pooling is realized across time. Although the nesting effects have
a positive impact on high priority fulfillment, a larger number of low priority orders
is rejected. This effect becomes most evident when considering the case of n = 15
receipts. In every single period high priority orders can consume the reserved portion
and any remaining unreserved inventory. Low priority orders can only consume the
unreserved portion of inventory available in the respective period. In any period, low
priority demand exceeding the unreserved portion is lost and cannot be compensated
by lower demand realizations in subsequent periods.
2,0%
n=1 n=3 n=5 n=15
1,8%
1,6%
1,4%
1,2%
fr
1,0%
0,8%
0,6%
0,4%
0,2%
0,0%
0,2c 0,4c 0,6c 0,8c 0,9c 1c 1,1c 1,2c
Fig. 7 Impact of the number of receipts on overall fill rate ( f r represents the difference in expected
overall fill rate with and without inventory reservation)
4.3 Effect of the demand lead time distribution
In our previous analyses we assumed that all orders have to be fulfilled immediately
in the period of their arrival. Our reservation approach as described in Sect. 3, how-
ever, explicitly accounts for uncertain due date requirements of the customers. In this
section we explore the impact of different demand lead time distributions on overall
fill rates of the system. We conducted experiments with three alternative demand lead
time distributions: (1) a uniform distribution gu ∼ U (0, 4); (2) a left-skewed distri-
bution gl ∼ (0.4, 0.3, 0.1, 0.1, 0.1); and (3) a right-skewed distribution gr ∼ (0.1,
0.1, 0.1, 0.3, 0.4). We choose to work with n = 5 receipts so that the length of any
inventory cycle (T /n) < τe , and a target service level of α N N = 0.95. In general, we
observe that demand lead time distribution does not have a very strong impact on the
performance of our reservation approach. This result is rather intuitive since different
demand lead time distributions only lead to a different allocation of overall demand
to the periods of the planning horizon. In the Fig. 8 we exemplarily plot the values of
f r for gu , gl , gr and g0 = g(τe = 0) = 1.
Overall we see an insignificant impact of the due date distributions on the overall
fill rate. Only for gr we observe a higher effect on f r . In this case mean demand
is lower in the initial periods of the planning horizon, leading to a lower reservation
quantity for the first cycle. As a consequence, the expected remaining quantity carried
over to the subsequent cycle is lower, requiring higher reservation quantities in later
periods. Only negligible effects on the realized high priority service levels and nesting
are caused by different demand lead time distributions.
Although the demand lead time distribution does not have a significant impact on
order fulfillment performance, it plays an important role in the correct calculation
2,0%
1,8%
1,6%
1,4%
1,2%
fr 1,0%
0,8%
0,6%
0,4%
0,2%
g0 gu gl gr
0,0%
0,2c 0,4c 0,6c 0,8c 0,9c 1,0c 1,1c 1,2c
Fig. 8 Impact of the demand lead time distribution on the overall fill rate ( f r represents the difference
in expected overall fill rate with and without inventory reservation)
of the reservation quantities. If the demand lead time distribution is not adequately
accounted for (as described in Sect. 3.3), the mean demand quantities due in the indi-
vidual inventory cycles are not calculated correctly and significant deviations in the
reservation quantities may occur across the inventory cycles. Although this will not
cause a negative effect on high priority service level, it may harm the overall fill rate
of the system. To exemplify this, we conducted an experiment in which reservation
quantities were based on a (wrong) due date distributions g(τe = 0); the true distri-
bution which determines the due dates of orders arriving in the planning horizon was,
however, assumed to be gr . In Fig. 9 we display the impact on f r compared to the
results obtained from the previous experiment in which both reservations and actual
due dates were based on gr . The significant additional loss in the overall fill rate is
caused by two effects: (1) Mean demand in the first cycle is severely overestimated,
resulting in a very high reservation quantity. A significant amount of reserved inven-
tory is not utilized by high priority orders and a large number of low priority orders are
rejected although inventory would have been available. (2) The overall mean demand
in the planning horizon is overestimated, leading to higher reservation quantities than
required for achieving the target service level of 0.95.
4.4 Comparison with a myopic reservation approach
With the final set of analyses we give evidence on how our reservation approach per-
forms in comparison to a myopic reservation policy. As explained in Sect. 3, we base
our reservation approach on the concept of a cycle-based service level. A distinct
feature of our approach is, however, that we consider the dynamic carry-over effects
across a multiple number of inventory cycles, dependent on the number of receipts
scheduled. Reserving inventory in a myopic fashion would imply that these carry over
8%
g0 and gr
gr and gr
7%
6%
5%
fr
4%
3%
2%
1%
0%
0,2c 0,4c 0,6c 0,8c 0,9c 1,0c 1,1c 1,2c
Fig. 9 Impact of basing inventory reservations on a “wrong” demand lead time distribution (n = 5 and
α N N = 0.95)
effects are ignored and that reservation quantities are only based on the demand dis-
tribution for an individual inventory cycle. Formally, this would imply that for any
cycle i, the reservation quantity ri would be set to FQ H (α), where Q iH represents the
i
high priority demand due in cycle i. By accounting for the carry-over effect (through
+
random quantity ri − Q iH available in the subsequent inventory cycle i + 1), the
overall reservation quantities are lower as compared to a myopic reservation policy,
while the target service levels are still achieved. This clearly leads to a positive impact
on the overall fill rate of the system. Lower reservation quantities will inevitably result
in a lower f r for a given target service level α N N . In Fig. 10 we plot the differ-
ences between the f r incurred with our inventory reservation approach and the f r
incurred with myopic inventory reservations across multiple receipts and α N N = 0.95.
In the case of one receipt, our approach and the myopic approach will lead to the
same results. From Fig. 10 we observe, however, that significant additional negative
consequences are incurred in the case of multiple receipts under the myopic approach.
Also, these negative consequences increase as the number of receipts increases. This
effect can be attributed to demand pooling across time. In the case of n = 3 receipts,
for example, ri = FQ H (α) for i = 1,2,3. The myopic approach does not account for
i
+ +
the fact that r1 − Q 1H and r2 − Q 2H will be available for subsequent cycles.
As the number of receipts increases to n = 5 the reservation quantities increase in
relative terms. Shorter inventory cycles lead to lower mean demand quantities per
cycle and to higher reservation quantities across the whole planning horizon. The
+
consequences of not accounting for carry-over of the unused quantities ri − Q iH
increase as the number of receipts increases. These results show that a “traditional”
cycle-based approach leads to an unnecessary negative impact of inventory reservation
7%
n=3 n=5 n=15
6%
fr
5%
Differences in
4%
3%
2%
1%
0%
0,2c 0,4c 0,6c 0,8c 0,9c 1,0c 1,1c 1,2c
Fig. 10 Differences in overall fill rate, proposed vs. myopic reservation approach
on the system performance. With the approach presented in this paper we are able to
achieve target high-priority service levels with significantly lower negative impact on
the system’s performance.
5 Conclusions
In this paper we extend the scope of traditional order fulfillment systems by integrating
inventory reservation for high priority customers and order promising. We develop an
integrated model to assist manufacturers in reserving inventory for future high priority
demand and determining order acceptance/rejection for incoming orders. Realizing
the practical difficulties associated with assigning each order a penalty cost to capture
the long term effects of not being able to fulfill an order according to the due date
requirements, we utilize a target service level as the basis for determining inventory
reservations. Our study is among the very few studies that utilize target service level as
an analogue for capturing the long term effects of not being able to quote the desired
due dates in order promising and inventory rationing. Most previous studies have
assumed that the short term and long term costs associated with due date quoting are
explicitly known.
We develop a model that captures uncertainty in the time of arrival, the desired due
date and the type of customer orders. A first contribution of this paper is the develop-
ment of an algorithm to calculate the amount of inventory to reserve in a MTS system
with multiple inventory receipts within the planning horizon. In addition, expressions
to characterize the probability of high priority fulfillment with nested inventory res-
ervations are presented. Unfortunately, these expressions do not yield closed-form
solutions and we performed simulation-based numerical analysis to trace the effects
of nesting.
Our numerical study also shows other interesting findings with strong managerial
relevance. We see that the effect of nested inventory reservations on the high priority
service level increases as the number of inventory receipts increases. However, the
loss in overall fill rate due to inventory reservation also increases in the number of
receipts. We also demonstrate that not knowing the true due date distribution or using
a myopic method to determine reservations can have strong detrimental effects on the
overall system performance.
In summary, apart from presenting an approach for integrated inventory reserva-
tion and order promising, our analysis generates various insights for managers on the
impact of inventory reservation under different system settings.
We have assumed that the time of arrival of scheduled inventory receipts is deter-
ministic to avoid an over-parameterized model with limited tractability. Future exten-
sions could consider modeling the uncertainty in inventory receipts and its impact
on inventory reservation and order fulfillment performance. Also, the demand within
the planning horizon is assumed to be stationary and identically distributed in each
period. This assumption, although not unreasonable in certain settings, may not be
applicable for others. Finally, we assumed that orders are rejected if the customer due
date requirements cannot be met. In many practical settings, customers may accept
some reasonable delay beyond their required due date. Analyzing the performance of
an inventory reservation approach with different customer behavior may be of theo-
retical and practical interest.
Acknowledgment The authors express their gratitude to the Spanish Ministry of Education and Science
(Project Reference: PSE-370500-2006-1) for financial support.
References
Ball MO, Chen C-Y, Zhao Z-Y (2004) Available to promise. In: Simchi-Levi D, Wu SD, Shen ZM (eds)
Handbook of quantitative supply chain analysis: modeling in the e-business era. Kluwer Academic
Publishers, Norwell, pp 447–484
Benjaafar S, ElHafsi M, de Vericourt F (2004) Demand allocation in multiple-product, multiple-facility,
make-to-stock systems. Manage Sci 43:1431–1448
Chen C-Y, Zhao Z-Y, Ball MO (2002) A model for batch advanced available-to-promise. Product Oper
Manage 11(4):424
Chen C-Y, Zhao Z-Y, Ball MO (2001) Quantity-and-due-date-quoting available-to-promise. Inform Syst
Front 3(4):477–488
Fleischmann B, Meyr H (2003) Customer orientation in advanced planning systems. In: Dyckhoff H et al
(ed) Supply chain management and reverse logistics. Springer, Berlin, pp 297–321
Ha AY (1997a) Inventory rationing in a make-to-stock production systems with several demand classes and
lost sales. Manage Sci 43:1093–1103
Ha AY (1997b) Stock rationing policy for a make-to-stock production system with two priority classes and
backordering. Naval Res Logist 43:458–472
Hariharan R, Zipkin P (1995) Customer-order information, lead times and inventories. Manage Sci
41:1599–1607
Kilger C, Meyr H (2008) Demand fulfilment and ATP. In: Stadtler H, Kilger C (eds) Supply chain man-
agement and advanced planning, 4th edn. Springer, Berlin, pp 181–198
Meyr H (2007) Customer segmentation, allocation planning and order promising in make-to-stock produc-
tion. OR Spect (forthcoming)
Nahmias S, Demmy WS (1981) Operating characteristics of an inventory system with rationing. Manage
Sci 27:1236–1245
Pibernik R (2006) Managing stock-outs effectively with order fulfilment systems. J Manufact Technol
Manage 17(6):721–736
Ronen D (1982) Measures of product availability. J Bus Logist 3(1):45–58
Ronen D (1983) Inventory service levels–comparison of measures. Int J Oper Product Manage 3(2):37–45
Silver EA, Pyke DF, Peterson R (1998) Inventory management and production planning and scheduling,
3rd edn. Wiley, New York
Topkis DM (1968) Optimal ordering and rationing policies in a non-stationary dynamic inventory model
with n demand classes. Manage Sci 15:160–176
Vericourt FD, Karaesmen F, Dallery Y (2000) Dynamic scheduling in a make-to-stock system: a partial
characterization of optimal policies. Oper Res 48:811–819
Zipkin P (2000) Foundations of inventory management. McGraw-Hill/Irwin, Boston
Setting safety stocks in multi-stage inventory systems
under rolling horizon mathematical programming
models
Youssef Boulaksil · Jan C. Fransoo ·

Ernico N. G. van Halm

OR Spectrum (2009) 31:121–140
DOI 10.1007/s00291-007-0086-3
Abstract This paper considers the problem of determining safety stocks in multi-item
multi-stage inventory systems that face demand uncertainties. Safety stocks are nec-
essary to make the supply chain, which is driven by forecasts of customer orders,
responsive to (demand) uncertainties and to achieve predefined target service levels.
Although there exists a large body of literature on determining safety stock levels,
this literature does not provide an effective methodology that can address complex
multi-constrained supply chains. In this paper, the problem of determining safety
stocks is addressed by a simulation based approach, where the simulation studies are
based on solving the supply chain planning problem (formulated as a mathematical
programming model) in a rolling horizon setting. To demonstrate the utility of the
proposed approach, an application of the approach at Organon, a worldwide operating
biopharmaceutical company, will be discussed.
Keywords Safety stocks · Advanced planning and scheduling · Simulation · Supply

chain planning · Organon
1 Introduction
Supply chains are exposed to different types of uncertainties that stem from random
yields, processing times or forecast errors. These uncertainties can be covered to a
Y. Boulaksil (B) · J. C. Fransoo

Department of Technology Management, Technische Universiteit Eindhoven,
Eindhoven, The Netherlands
J. C. Fransoo
E. N. G. van Halm
Supply Chain Management Department, Organon N.V., Oss, The Netherlands

200 Y. Boulaksil et al.
large extent by mechanisms like safety time, safety stocks or combinations of these
(Whybark and Williams 1976; Wijngaard and Wortmann 1985). This paper focuses on
the determination of safety stocks in multi-item multi-stage inventory systems that face
demand uncertainties. We assume that the inventory system is planned and controlled
by a central decision authority that plans the supply chain based on deterministic math-
ematical programming models. However, demand uncertainty is an important factor
to be considered in supply chain planning. Planning systems based on mathematical
programming models are widely implemented in so-called Advanced Planning and
Scheduling systems (APS) (Stadtler and Kilger 2005).
When a particular supply chain is facing demand uncertainties, stock outs can occur
at all stages in the supply chain. A stock out may cause lost sales, emergency ship-
ments, or loss of goodwill. Therefore, safety stocks should be kept to increase the
service levels. Traditionally, safety stocks are determined in advance based on models
from inventory theory (Silver et al. 1998). However, it is not obvious how to determine
safety stock levels that cover demand uncertainties in complex supply chains that face
several constraints such as batch sizes, capacity constraints, non-stationary demand
process or forecast errors.
The approach proposed in this paper enables the determination of safety stocks
in multi-item multi-stage inventory systems that face demand uncertainties. This
approach considers all kinds of constraints that are also considered in supply chain
planning practice such as batch sizes and capacity and materials constraints. The
approach is based on a simulation of the supply chain planning model in a rolling
horizon setting. Based on target service levels, safety stocks are determined after per-
forming simulations, assuming that the demand process and replenishment decisions
are independent of the safety stock levels.
The core of the approach is solving the supply chain planning problem very fre-
quently, where the safety stocks are excluded from the supply chain planning model
or by setting them equal to zero. Since we assume that all unsatisfied demand is back-
ordered at all stages in the supply chain, backorder quantities are recorded after each
solving round. The safety stock level is an increasing function of the target service
level, which we measure by the fill rate, i.e. the long-run fraction of demand satisfied
routinely from the shelf (Silver et al. 1998). Based on the stored backorder quantities
and the target service levels, safety stock levels can be determined.
The proposed approach is suitable for companies that have implemented an APS.
APS systems are planning systems that are based on cost minimization models that
ensure that, given the resource and material availability constraints of the produc-
tion system and given certain service level constraints, the best possible quantity of a
certain item is released at the lowest value of the objective function. These planning
systems are based on mathematical programming models that are solved in a rolling
horizon setting (Spitter et al. 2005).
The proposed approach has been applied successfully at Organon, a worldwide
operating biopharmaceutical company with production sites, warehouses, and distri-
bution centers spread all over the world.
The remainder of this paper is organized as follows. Section 2 discusses a literature
review on this topic. Next, Sect. 3 discusses the problem definition and thereafter,
Sect. 4 discusses the proposed approach. The approach has been applied in a real-life
Setting safety stocks in multi-stage inventory systems 201
situation, which is discussed in Sect. 5. Finally, Sect. 6 draws some conclusions about
the approach.
2 Literature review
There is an extensive amount of literature available on inventory control models in

multi-stage or multi-echelon inventory systems incorporating uncertainties. We refer
the reader to survey articles by Van Houtum et al. (1996) and Diks et al. (1996).
Our research is within the field of Supply Chain Operations Planning (De Kok and
Fransoo 2003). The objective of Supply Chain Operations Planning is to coordinate
the release of materials and resources in a supply chain network such that customer
service constraints are met at minimal costs (De Kok and Fransoo 2003). Two different
approaches exist for modelling the Supply Chain Operations Planning problem.
One approach is based on multi-echelon stochastic inventory theory. In this ap-
proach, demand that is faced by the supply chain is modelled as a stochastic variable.
The key decisions of this approach are the inventory positioning at the various stock-
points in the supply chain, the allocation of quantities at inventory points where the
product flow diverges, and the determination of safety stock levels at the several stock-
points. Therefore, the determination of safety stocks is defined as part of the problem.
Lead times are (deterministic) input variables to the model and capacity is assumed to
be controlled through a combination of order acceptance in the demand management
function and a workload control function in the production department. The logic is
based on a line of research that has been initiated by Clark and Scarf (1960).
The alternative approach is based on mathematical programming principles. In this
approach, demand is inserted into the model as forecasts for every period in the plan-
ning horizon. Safety stocks are input parameters to the model and the key decisions
are the allocation of inventory quantities at the stockpoints in the supply chain. Lead
times are either modelled as deterministic input variables (e.g., Spitter et al. 2005)
or are observed as output variables of the model (e.g., Stadtler 2003). Capacity con-
straints are modelled explicitly as aggregate constraints. The principles are based on
research stemming from advanced MRP modelling (Billington et al. 1983) or from
multi-period lot sizing problems (Tempelmeier and Derstroff 1996; Stadtler 2003).
The principles have been implemented in commercial software, mostly using CPLEX
solving logic (See also Stadtler and Kilger 2005).
The two approaches differ also from a safety stock perspective. In this first approach,
safety stocks are defined as part of the problem, whereas in the second approach
safety stocks are input parameters to the planning model, which have to be determined
externally.
This paper focuses on the determination of safety stocks for the latter type of plan-
ning approaches. A lot of papers appeared in the last decades on determining safety
stocks in multi-stage or multi-echelon inventory models for covering demand uncer-
tainties. We mention a set of papers that are related to our work.
Inderfurth and Minner (1998) propose a dynamic programming approach to treat
the problem of determining safety stocks in multi-stage inventory systems, assum-
ing normally distributed demand and periodic review base stock control policies.
Furthermore, they also assume that no internal delays occur and that each stockpoint
is satisfying a service level constraint. More approaches that are based on dynamic
programming algorithms can be found in Inderfurth (1991) and Minner (1997).
Graves and Willems (2000) discuss the so-called guaranteed-service model for set-
ting safety stocks in a multi-stage setting to cover demand uncertainties. They develop
a model for positioning safety stocks in a supply chain where each stage is controlled
by a base-stock policy, assuming an upper bound for the (customer) demand level.
Therefore, the safety stocks set by their approach cover demand realizations below
the upper bounds. This assumption is necessary to model guaranteed service times
between each stage in the supply chain and its customers.
There are also papers on determining safety stocks in multi-stage inventory systems
where the approach is based on simulation studies. Optimization methodologies based
on simulation of inventory systems are discussed in Kleijnen and Wan (2006).
Eilon and Elmaleh (1968) perform simulation studies to compare the performance of
five alternative inventory control policies given wide fluctuating and seasonal demand
patterns. The results of the simulations are several non-linear curves showing the rela-
tion between the fill rate and mean stock level. Three of these five control policies
include safety stocks, but the authors do not discuss how they determined the safety
stock parameters.
Wemmerlöv and Whybark (1984) also perform simulation experiments to evaluate
several single-stage lot sizing procedures under demand uncertainty. Cost compari-
sons of the procedures are made with a service level of at least 99.999%. The safety
stocks needed to achieve these service levels are determined by a search routine, i.e.
repeating the simulations until the target service levels are reached.
De Bodt and Van Wassenhove (1983) present a case study at a company, which uses
MRP in a dynamic environment, i.e. the company faces substantial demand uncertain-
ties. The safety stock setting is analysed by a simulation study. Several strategies were
defined (combinations of safety stock and safety time) and analysed which resulted
in graphs relating average inventory level to service levels. They provide managerial
insight by showing that considerable savings can be made at this company, but do not
discuss how the safety stocks should be determined.
In the studies of Callarman and Hamrin (1984) the performance of three lot sizing
rules in MRP systems is compared, given an uncertain demand process. The cost com-
parisons have been made by introducing safety stocks at each run to keep the service
levels at 95 and 98%. The required safety stocks are determined by using the so-called
Service Level Decision Rule (SLDR), which has been developed by Callarman and
Mabert (1978). The SLDR is based on linear regression analysis on simulated values
of the following set of factors: forecast errors, coefficient of variation of demand, and
the expected time between orders. In order to achieve the target service level, the
SLDR is used with a search routine.
Our work is closely related to Kohler-Gudum and De Kok (2001) who propose a
so-called Safety Stock Adjustment Procedure (SSAP) to obtain target service levels
in simulation models. The technique is based on the assumption that a Time Phased
Order Point (TPOP) policy is applied. Their simulation study aims to determine the
discrete probability density function of the net stock process. Based on this probability
distribution, the safety stock is adjusted to ensure the specified target service level.
Our approach differs from Kohler-Gudum and De Kok (2001) on two aspects.
First, our approach determines the empirical distribution of the backorder quantities
for setting safety stocks instead of determining the probability density function of the
inventory process. Based on a set of assumptions (independence of the demand and
replenishment process of safety stock levels), the necessary amount of safety stock is
determined by adding the adjustment quantity to the initial safety stock that can be an
arbitrary value. In our approach, the initial safety stock is set equal to zero and after-
wards, the safety stock level is determined based on backordered quantities. Second,
the model that is used in our approach is a planning model that is solved in rolling
horizon setting, and therefore, the planning process is imitated as much as possible.
Kohler-Gudum and De Kok (2001) do not discuss the simulation model extensively.
Although there exists a large body of literature on determining safety stock levels,
to our knowledge, this literature does not provide an effective methodology that can
address supply chains that face several constraints like capacity constraints, production
in batch sizes, and non-stationary forecast process. Most approaches make restrictive
assumptions about the demand process (Inderfurth and Minner 1998; Graves and
Willems 2000) or do not explicitly discuss how they set the safety stock levels (Eilon
and Elmaleh 1968; De Bodt and Van Wassenhove 1983). Our approach is closely
related to Kohler-Gudum and De Kok (2001), but we extend the approach by using
an empirical supply chain planning model in the simulation study and that makes the
results of the approach more reflecting the (planning) practice.
3 Problem definition
We consider a supply chain that consists of an arbitrary number of stages and stock-
points in which a product passes through multiple production sites before it is finally
delivered to outside customers. This supply chain is planned and controlled by a cen-
tral decision authority that has access to all relevant status information (like inventory
levels and work-in-process quantities) at all production sites and makes release deci-
sions for the entire supply chain. The release decisions result from a deterministic
mathematical programming model that is solved in a rolling horizon setting (Stadtler
and Kilger 2005), which has been implemented in an Advanced Planning and Sched-
uling system. For these kinds of planning models, safety stocks are input parameters
that have to be determined externally.
Formulation of the planning problem by a mathematical programming model
assumes a deterministic view of supply chain planning by considering all model param-
eters, as demand, lead times, production rates to be known with complete certainty.
This assumption of complete and deterministic information is desirable from a model
complexity point of view, but given the dynamic and uncertain nature of most supply
chains, this assumption is violating reality. Demand uncertainty is an important factor
to be considered in supply chain planning, and therefore, safety stocks are kept to
cover part of the demand uncertainties.
The core function of supply chain planning models is to coordinate material and
resource release decisions in the supply chain such that predefined customer ser-
vice levels are achieved with minimal costs. Safety stocks are kept to deal with
demand uncertainties and consequently to increase service levels. The service level
is an increasing function of the safety stock level. Therefore, more safety stocks are
needed to increase the service level, which results in increased inventory holding costs.
From the other side, demand uncertainties can cause stock outs that result in lost
sales, emergency shipments, or loss of goodwill. Since we assume that all excess
demand that is not directly satisfied from inventory is backordered, costs that are
related with a backorder are backorder costs, which are harder to quantify than inven-
tory holding costs. The problem of setting safety stocks is mainly a trade-off between
inventory holding costs and backorder costs. Section 4 discusses the modelling of
these costs and discusses also the considered supply chain planning model in detail.
4 The approach
We consider a supply chain that is planned and controlled by a central decision author-
ity, which may be supported by an Advanced Planning and Scheduling system. We
assume that the supply chain planning model is based on a mathematical programming
model that is solved in a rolling horizon setting, where the forecasts may be updated
when the planning horizon is shifted. We also assume that the demand process and
replenishment decisions are independent of the safety stock level. Furthermore, we
assume that all excess demand at all stages in the supply chain is backordered. We
do not make any assumption about the demand and forecast process, which makes
this approach less restrictive to a certain probability density function of the demand
process.
Based on the discussed assumptions, a simulation experiment is performed in the
following way. The planning horizon is divided into a fixed number of time buckets,
which are filled by forecasts of the demand generated by a demand generator, which
generates a series of forecasts based on historical demand and forecasts data. Then,
the planning model with demand forecasts is solved given all kinds of materials and
resources constraints. The planning model may be based on linear programming mod-
els or mixed-integer programming models if some decisions require integer variables.
Such discrete decisions can, among others, regard lotsizing in production or transpor-
tation. At the end of the first time bucket (planning cycle), the state of the system (e.g.
the inventory levels and forecasts) is updated and the planning cycle is repeated with
the horizon shifted by one period.
Figure 1a illustrates the inventory development of a certain product and the result-
ing backorder process that is output of 100 simulation runs. After the simulation runs,
the horizontal axis is shifted (see Fig. 1b) such that the number of backorders is lim-
ited, i.e. the customer service level is increased to a certain predefined level. Figure 1b
shows that increasing the safety stock level decreases the number of backorders, and
therefore, increases the service level.
Thus, by solving the supply chain planning problem very frequently where each
time the forecasts are updated, long-run backorder quantities indicate the amount of
safety stocks that was needed to prevent the backorders partially, i.e. to achieve a
certain customer service level. Note that the customer service level is externally deter-
mined for all products at all stages in the supply chain. The customer service level has
Inventory A Inventory Inventory B

level level
0 0
Backorders Safety stock
0‘
time
Fig. 1 The inventory development of a certain product; a shows the results of simulation runs and b shows
how the horizontal axis is shifted to limit the number of backorders, i.e. to achieve a certain customer service
level
f(d), µ d,i, d,i
Demand generator
dˆ i (t), t 1 = ,...,T
SC Planning model horizon

shift
Optimal solution
Backorder quantities
Service level setting
Safety stock levels
Explanation of used symbols:

f(d) probability density function of the demand process
µ d,i Expected (exogenous) demand of end-item i
d,i Standard deviation of forecast errors of end-iem i
T Planning horizon
dˆi (t ) Forecast of demand of end-item i in period t (t=1,...,T)
Fig. 2 Several steps of the approach
been defined as the long-run fraction of demand satisfied directly from shelf (fill rate
measure). Having discussed the theoretical idea behind the approach, the steps of the
approach (see Fig. 2) will be discussed in the following sections in more detail.
4.1 Demand generator
The first step of the approach is the generation of a series of forecasts which are
input to the supply chain planning model. We do not make any assumption about the
distribution of the demand process. Historical data about demand and forecasts may
be (statistically) fitted into the best fitting probability distribution function. Having
chosen the most suitable probability density function for the demand distribution, the
first two moments of the distribution can be derived to determine the parameters of
the demand distribution. Figure 2 shows that these parameters (µd,i , σd,i ) are input
for the demand generator where µd,i is the expected (exogenous) demand of end-item
i and σd,i is the standard deviation of forecast errors of end-item i.
Suppose that we have historical sales data of n time periods, then µd,i can be
calculated by
−1
1
µd,i = di (t + s) (1)
n s=−n
where di (t) is the demand for item i in period t. Forecast errors can be determined
by several measures (Silver et al. 1998). The Mean Absolute Deviation (MAD) is
recommended for its computational simplicity. The MAD for item i as function of the
forecast horizon h can be calculated by
1
−1
MADi (h) = di (t + s) − d̂i (t + s − h, t + s) (2)
T
s=−T
where T is the length of the planning horizon, di (t) the demand for item i in period
t, and d̂i (t − h, t) the forecast made in period t − h for the demand in period t. It is
reasonable to assume that the MAD is an increasing function of the forecast horizon
h (Heath and Jackson 1994). The conversion of MADi (h) to σi (h) is extensively dis-
cussed in Silver et al. (1998). Having determined the parameters of the demand distri-
bution, the random generator can generate a series of forecasts of the demand d̂i (t) for
t = 1, . . . , T . The generated forecasts are input for the supply chain planning model.
Several extensions are possible. For example, the demand process may not be sta-
tionary which makes the µd,i a function of time (demand process follows a trend or
has a seasonal effect). Furthermore, the standard deviation of the forecast errors may
be not a function of the forecast horizon h. Several kinds of adaptations are possible
in order to imitate the demand and forecast process as much as possible.
4.2 Supply chain planning model
One approach to supply chain planning models is based on deterministic mathemat-

ical programming principles (De Kok and Fransoo 2003). The advantage of using
the supply chain planning model (implemented in an APS system) for the simulation
study is that it already contains the network structure(s), the item list, bill-of-mate-
rials structure, batch sizes, and the routings. Using the supply chain planning model
for this approach is highly recommended, as these models are reflecting the planning
practice. Furthermore, for those companies that have implemented an APS system,
little modelling effort is required for this approach.
The mathematical programming model that is used to determine the safety stock
levels is a stand-alone model, but derived from the supply chain planning model. The
supply chain planning model may have to be adapted, as safety stocks have to be set
Fig. 3 A three-stage supply chain considered in the supply chain planning model
equal to zero in the supply chain planning models or replenishment decisions should
not consider safety stock levels. Then, the planning problem should be solved without
considering safety stocks, i.e. backorders are planned when demand exceeds available
inventories. Depending on the supply chain planning model, the planning problem has
to be solved such that all relevant cost factors have to be considered except costs asso-
ciated with consumption of safety stocks. The solution of the supply chain planning
model contains order releases for the production system and planned inventory levels.
The order releases within the length of the lead time of a certain item at a certain stage
(frozen horizon) are stored, as they are not allowed to be changed in the next solving
round. Below, we discuss the considered supply chain planning model in detail, which
is used to determine the safety stock levels. Figure 3 shows a rough outline of the
three-stage supply chain that is considered in the supply chain planning model.
4.2.1 Objective function
Equation (3) is the objective functions of the supply chain planning model. The objec-
tive function minimizes the total costs (TC), which consist of several cost factors that
are assigned to several stages in the supply chain. We consider a (pharmaceutical) sup-
ply chain with three stages. Stage 3 is the most upstream stage where the raw materials
(active ingredients) are stored. The active ingredients are processed to tablets, which
are stored at the second stage. Thereafter, the tablets are packaged and stored at the
most downstream stage (stage 1). N j is the total number of items at stage j with
j ∈ {1, 2, 3}, n j is a certain item that belongs to stage j, t is a certain (discrete) time
period, and T is the planning horizon.

T
T
Min TC = c1 · UDn 1 (t) + c2 · OPn 1 · BMn 1 (t)
n1 n1
t=1 t=1

T
T
+ c3 · EIn 1 (t) + c4 · UDn 2 (t)
n1 n2
t=1 t=1

T
T
+ c5 · UDDn 2 (t) + c6 · BCn 2 (t)
n2 n2
t=1 t=1

T
T
+ c7 · EIn 2 (t) + c8 · UDn 3 (t)
n2 n3
t=1 t=1

T
T
+ c9 · UDDn 3 (t)+ c10 · EIn 3 (t) (3)
n3 n3
t=1 t=1
The first three terms of the objective function are related to the first stage in the
supply chain. For this stage, we consider three cost factors that have to be minimized.
The first terms are costs associated with unsatisfied demand (backorders) c1 · UDn 1 (t)
where UDn 1 (t) is the backorder quantity for item n 1 in period t. The second term
deals with costs for replenishing a quantity that deviates from the (minimum) replen-
ishment quantity c2 · OPn 1 · BMn 1 (t). OPn 1 is the period order quantity for item n 1
and BMn 1 (t) the deviation from the minimum replenishment quantity for item n 1 in
period t. Campaign sizes are determined based on a trade-off between ordering costs
and inventory holding costs, whereas batch sizes are quantities that are determined by
legislative authorities. Therefore, producing in fixed batch sizes is required, whereas
deviating from the campaign size is undesired. The third term is the total inventory
holding cost at this stage c3 · EIn 1 (t) where EIn 1 (t) is the inventory level of item n 1 at
the end of period t.
For the second stage of the supply chain, four cost factors are considered. The first
terms sum backorders that result from exogenous demand at this stage. So, c4 ·UDn 2 (t)
is unsatisfied demand (backorder) costs for item n 2 in time period t, whereas the sec-
ond term c5 · UDDn 2 (t) considers unsatisfied demand (backorders) that result from
endogenous (derived) demand from the first stage of the supply chain. Since the sec-
ond stage in the supply chain considers the production of tablets in campaigns (a fixed
multiple of batch sizes), the third term c6 · BCn 2 (t) considers costs associated with
deviating from the fixed campaign size BCn 2 (t) for item n 2 in period t. The fourth
term considers the total inventory holding costs for item n 2 .
The third stage in the supply chain considers three cost factors: costs associated with
unsatisfied demand (backorders) of exogenous demand c8 · UDn 3 (t), costs associated
with unsatisfied demand (backorders) that result from endogenous demand (from the
second stage of the supply chain) c9 · UDDn 3 (t), and total inventory holding costs
c10 · E In 3 (t) of item n 3 . For confidentially reasons, we cannot show the values of the
cost parameters, except that c1 > c2 > · · · > c10 . The determination of these cost
parameters was not part of this study, as they can be taken over from the objective
function of the supply chain planning model.
4.2.2 Stage 1 model
The objective function (3) is minimized subject to several constraints, which are dis-
cussed below per stage in the supply chain. Equations (4) are materials balance equa-
tions with EIn 1 (t) is the inventory level of item n 1 at the end of period t, T Rn 1 (t) the
replenishment quantity of item n 1 in period t, and TDn 1 (t) the (exogenous) demand
for item n 1 in period t. The latter parameter contains data that are input to the planning
model. Further, EIn 1 (0) is the initial inventory level.
EIn 1 (t) = EIn 1 (t − 1)+TRn 1 (t)−TDn 1 (t), n 1 = 1, . . . , N1 , t = 1, . . . , T (4)
Equation (5) determine the minimum replenishment quantity for item n 1 in period
t, as the replenishments are based on periodic order quantity (OP). I Dn 1 (t) is the
(forecast of) independent demand for item n 1 in period t.

OP
MRn 1 (t) = I Dn 1 (t + i), n 1 = 1, . . . , N1 , t = 1, . . . , T (5)
i=1
Having determined the minimum replenishment quantity, Eq. (6) determine the real
replenishment quantities SMn 1 (t) for item n 1 in period t. BMn 1 (t) is then the deviation
from minimum replenishment quantity for item n 1 in period t that is considered in the
objective function.
SMn 1 (t) = MR n 1 (t) − BMn 1 (t), n 1 = 1, . . . , N1 , t = 1, . . . , T (6)
The total replenishment quantity for item n 1 for the entire planning horizon (TR n 1
(t)) is determined by two parts: SMn 1 (t) which we have just discussed and FPn 1 (t)
which are fixed replenishment quantities of item n 1 in period t determined in previ-
ous solving rounds. The binary parameter α regulates that within the lead time of the
planning horizon no new decisions are taken.
TRn 1 (t) = α · S Mn 1 (t) + (1 − α) · FPn 1 (t), n 1 = 1, . . . , N1 ,

0 if t L
t = 1, . . . , T, α = (7)
1 if L < t T
Equation (8) determine which part of the exogenous demand IDn 1 (t) for item n 1 in
period t is satisfied (SDn 1 (t)). The unsatisfied demand quantity UDn 1 (t) for item n 1
in period t is punished in the objective function.
SDn 1 (t) = IDn 1 (t) − UDn 1 (t), n 1 = 1, . . . , N1 , t = 1, . . . , T (8)
Equation (9) are equations for TDn 1 (t) which is determined by SDn 1 (t) that result
from Eq. (8) plus UDn 1 (t − 1) which is the unsatisfied demand in t − 1, i.e. backorder
quantity for item n 1 from period t.
TDn 1 (t) = SDn 1 (t) + UDn 1 (t − 1), n 1 = 1, . . . , N1 , t = 1, . . . , T (9)
4.2.3 Stage 2 model
Several constraints apply to stage 2 which will be discussed now. Like in stage 1,
Eq. (10) are the balance equations for the materials flow. The symbols have the
same meaning as in stage 1, except that indices show that the equations apply to this
particular stage.
EIn 2 (t) = EIn 2 (t −1)+TRn 2 (t)−TDn 2 (t), n 2 = 1, . . . , N2 , t = 1, . . . , T (10)
Equation (11) determine the total replenishment quantity for item n 2 in period t
where α is the same binary parameter that is used in Eq. (7). PPn 2 (t) is the production
quantity to be produced of item n 2 in t and FPn 2 (t) are firmed production quantities
that are determined in previous solving rounds.
T Rn 2 (t) = α · PPn 2 (t)+(1−α) · F Pn 2 (t), n 2 = 1, . . . , N2 , t = 1, . . . , T (11)
Equation (12) require that the production quantity of item n 2 to be produced in

period t must be an integer multiple of Q n 2 , the batch size of item n 2 multiplied by
yn 2 , the yield factor of the production process that produces item n 2 .
PPn 2 (t) = Q n 2 · yn 2 · NBn 2 (t), n 2 = 1, ..., N2 , t = 1, . . . , T, (12)
with N Bn 2 (t) ∈ N0 .
Equation (13) determine the derived (endogenous) demand at stage 2. This is the
multiplication of the (with lead time L shifted) replenishment quantities of items n 1
with the BOM factor.

DDn 2 (t) = BOMn 2 ,n 1 · SMn 1 (t − L), n 2 = 1, . . . , N2 , t = 1, . . . , T (13)
n1
Equation (14) determine the costs associated with going below the campaign size
BCn 2 (t), which is punished in the objective function. CSn 2 is the campaign size (a
certain number of batches of n 2 ) of item n 2 .
BCn 2 (t) = CSn 2 − NBn 2 (t), n 2 = 1, . . . , N2 , t = 1, . . . , T, (14)
with BCn 2 (t) ∈ N0 .

Unsatisfied demand from t−1 (resulting from either exogenous demand UDn 2 (t−1)
or endogenous demand UDDn 2 (t − 1) determine the backorder quantity BOn 2 (t) of
item n 2 in period t.
BOn 2 (t) = UDDn 2 (t − 1) + UDn 2 (t − 1), n 2 = 1, . . . , N2 , t = 1, . . . , T (15)
Equation (16) show that the satisfied part of demand for item n 2 in period t SDn 2 (t)
is equal to the exogenous demand IDn 2 (t) for item n 2 in period t minus unsatisfied
demand quantity UDn 2 (t) for item n 2 in period t, which is punished in the objective
function.
Equation (17) are the application of the same idea (as Eq. 16) to the dependent
(endogenous) demand for item n 2 in period t.
SDDn 2 (t) = DDn 2 (t) − UDDn 2 (t), n 2 = 1, . . . , N2 , t = 1, . . . , T (17)

The sum of SDn 2 (t), SDDn 2 (t), and the backorders for item n 2 in period t BOn 2 (t)
are equal to TDn 2 (t), total demand for item n 2 in period t.
TDn 2 (t) = SDDn 2 (t)+SDn 2 (t)+BOn 2 (t), n 2 = 1, . . . , N2 , t = 1, . . . , T (18)
4.2.4 Stage 3 model
Constraints (19) till (26) apply to the third stage of the supply chain. Equation (19) are
the balance equations for this stage. EIn 3 (t) is the inventory level of item n 3 at the end
of period t, TRn 3 (t) is the replenishment quantity of item n 3 in period t, and TDn 3 (t)
is the total demand of item n 3 in period t.
EIn 3 (t) = EIn 3 (t − 1)+TRn 3 (t)−TDn 3 (t), n 3 = 1, . . . , N3 , t = 1, . . . , T (19)
The replenishment quantity TRn 3 (t) is partly determined in the previous solving
rounds (FPn 3 (t), firm planned replenishment orders for item n 3 in period t) and new
released orders On 3 (t) to be determined for item n 3 in period t. The orders are sent to
(external) supplier(s).
TRn 3 (t) = (1 − α) · FPn 3 (t) + α · On 3 (t), n 3 = 1, . . . , N3 , t = 1, . . . , T (20)
Furthermore, constraints (21) require that the ordered items are (a) integer multi-
ple(s) of Q n 3 , batch sizes for item n 3 .
On 3 (t) = NBn 3 (t) · Q n 3 , n 3 = 1, . . . , N3 , t = 1, . . . , T, (21)
with NBn 3 (t) ∈ N0 .

The total demand for item n 3 is determined by adding the satisfied parts of the
dependent (endogenous), independent (exogenous) demand plus the backorders for
item n 3 in period t.
TDn 3 (t) = SDDn 3 (t)+SDn 3 (t)+BOn 3 (t), n 3 = 1, . . . , N3 , t = 1, . . . , T (22)
Equations (23) and (24) show how the satisfied parts of the dependent SDDn 3 (t)
and independent demand SDn 3 (t) for item n 3 in period t are determined. IDn 3 (t) is
the independent demand for item n 3 in period t and DDn 3 (t) is the dependent demand
for item n 3 in period t.

SDDn 3 (t) = DDn 3 (t) − UDDn 3 (t), n 3 = 1, . . . , N3 , t = 1, . . . , T (24)
The dependent demand DDn 3 (t) is determined by multiplying the BOM-factor with
TPn 2 (t − L) with L is the lead time of second stage of the supply chain.

DDn 3 (t) = BOMn 3 ,n 2 TPn 2 (t − L), n 3 = 1, . . . , N3 , t = 1, . . . , T (25)
n2
The backorder quantity for item n 3 in period t BOn 3 (t) is the summation of the
unsatisfied dependent and independent demand for item n 3 in period t − 1.
BOn 3 (t) = UDDn 3 (t − 1) + UDn 3 (t − 1), n 3 = 1, . . . , N3 , t = 1, . . . , T (26)
Finally, non-negativity constraints have to be considered.
EIn 1 (t), TDn 1 (t), TRn 1 (t), SMn 1 (t), BMn 1 (t), SDn 1 (t), UDn 1 (t), EIn 2 (t),
TDn 2 (t), TRn 2 (t), PPn 2 (t), UDn 2 (t), UDDn 2 (t), SDn 2 (t), SDDn 2 (t), EIn 3 (t),
TDn 2 (t), TRn 2 (t), On 2 (t), UDn 2 (t), UDDn 2 (t), SDn 2 (t), SDDn 2 (t) 0 (27)
4.3 Backorders and safety stocks
After solving the mathematical programming model that we discussed in the previous
section, the planning horizon is shifted by one period after which the demand gen-
erator generates a new series of forecasts for the shifted horizon. The order releases
within the frozen horizon determined in the previous solving round are not allowed to
be changed, as these orders are assumed to be scheduled in a more detailed planning
level or already taken in process. The supply chain planning model is solved again,
but since a frozen horizon, fixed order releases and an update of the forecasts are taken
into consideration, backorders may occur if the available inventories are no longer
sufficient to satisfy the updated required quantities.
The planned backorder quantities after each solving round are stored. A large num-
ber of replications is necessary to draw valid conclusions on the empirical distribution
of the backorders. Furthermore, the results of the first couple of runs have to be ignored,
as the system has to reach a state that is independent of the initial conditions. The rela-
tion between the backorder quantities and the determination of safety stock levels will
be explained in the following. Suppose that the safety stock levels were set before-
hand equal to the maximum measured backorder quantities at all stages in the supply
chain, a service level of 100% would have been achieved in the supply chain, given the
generated forecasts of the demand process. Therefore, the last step of the approach is
to set a target customer service level for the several items at the several stages. Based
on these target customer service levels, safety stock levels can be determined for the
items.
The service level for the most downstream stage (stage 1) in the supply chain is
determined by Eq. (28) where βn 1 is the fill rate for item n 1 , i.e. the long-run fraction
of independent demand IDn 1 (t) satisfied directly from shelf (without backordering).
UDn 1 (t)

βn 1 = 1− , n 1 = 1, . . . , N1 (28)
t
IDn 1 (t)
For stages 2 and 3 of the supply chain, Eq. (29) determine the service level, as these
stages face exogenous (independent) demand and endogenous (dependent) demand
from the next (downstream) stage of the supply chain.

UDn j (t) + UDDn j (t)
βn j = 1− , j = {2, 3}, n j = 1, . . . , N j (29)
t
IDn j (t) + DDn j (t)
5 Application of the approach
The discussed approach has been implemented at Organon, a worldwide operating

biopharmaceutical company with an annual turnover of more than 2.4 billion Euros.
The company consists of more than 10 production sites and about 60 national distri-
bution centres spread all over the world. Organon has more than 30 branded products
in its portfolio and markets only prescription medicines for improving both the health
and quality of human life.
Figure 4 shows a rough outline of one of Organon’s tablet supply chains with the
main production processes and stockpoints. Active ingredients form input to the tab-
lets production process. Some additional materials may be needed for this production
process. The packaging process blisters the tablets, packs the blistered tablets in car-
tons and instructions for use are added. Next, the finished products are shipped to
more than 60 national warehouses (which are owned by Organon) spread all over the
world. From these national warehouses, finished products are sold and distributed to
customers like hospitals, pharmacists, and wholesalers.
This supply chain is planned and controlled by an APS, which was implemented
a couple of years ago. The APS is a planning system that controls the supply chain
by calculating high-level production plans for the several stages in the supply chain.
The forecasts which are input to the planning problem are provided by the forecasting
system which calculates statistical forecasts of the expected demand on SKU level
based on historical demand information. Having implemented the Advanced Plan-
ning and Scheduling system, Organon was facing the question how to determine the
safety stock parameters (which are input to the planning models) that cover (partially)
demand uncertainties such that the entire supply chain is considered and total inventory
holding costs are minimized given certain customer service levels.
In the following sections, we discuss the results of a project performed within
Organon to determine the safety stock levels using the discussed approach. For con-
fidentially reasons, the product names for which this approach has been implemented
will not be mentioned. Further, the numbers do not reflect the real numbers, as they
are divided by an arbitrary factor. With respect to the simulation experiment, the run
length was set equal to 100 periods and the number of replications was five, fol-
lowing the approach proposed by Law and Kelton (2000). The average CPU time is
about 2 mins. The supply chain planning model has been implemented in a standard
Advanced Planning System that uses CPLEX as solver.
Fig. 4 A rough outline of the supply chain of Organon

5.1 The supply chain of product A
Figure 5 shows the supply chain of product A. The most upstream stage in this supply
chain is the active ingredient (AI 1). The tablet production is performed at two produc-
tion sites of Organon, and therefore, active ingredients 1 (AI 1) are shipped to another
production site (AI 2). After tablets production 2 (TB 2), the tablets are packaged and
shipped to warehouses 1 till 11, which supply Organon’s end customers. Production
site 1 is supplying warehouses 12 till 26. Warehouses 27 till 34 are supplied by local
subcontractors who get the active ingredients from Organon. Therefore, a direct link
has been made between the stockpoints AI 1 and warehouses 27 till 34.
We applied the proposed approach to this supply chain to determine the safety
stocks levels of each item at each stockpoint. We note that safety stocks cannot be
pooled, since the products in each warehouse are different due to the fact that they are
country-specific. The lead times, batch sizes, bill-of materials structure and all other
characteristics of this supply chain have been taken over from the supply chain plan-
ning model. The mathematical formulation of the mixed-integer programming model
that is solved in a rolling horizon setting is the one discussed in Sect. 4.2.
Based on stored historical demand and forecasts data, we found that the normal
distribution is statistically fitting the forecasts and sales data the best. A goodness-of-
fit test has been used to find the suitable distribution that fits the best to the data. The
result was that the average demand is time-independent, but forecast errors showed a
strong correlation with the forecast age, i.e. the number of periods between the moment
the forecast was made and the moment the demand is realized. The demand genera-
tor randomly generates a series of forecasts based on the parameters of the normally
distributed demand (µd,i , σd,i (h)).
Fig. 5 The supply chain of

product A
100
90
80
Backorder quantity
70
60
50
40
30
20
10
0
Time
Fig. 6 An example of the backorders that result from a simulation study for a certain item
Table 1 Safety stock levels

Current situation Model suggestion
based on the proposed approach
and the current safety stock
National warehouses 100 89
levels
Tablets stockpoint 51 47
Active ingredients 49 21
As discussed in the previous section, the outputs of the approach are series of
planned backorders that are stored. Figure 6 shows the development of backorders of
a certain item. As this figure shows, the backorder quantities are mostly equal to zero,
which means that there is mostly enough inventories available to satisfy the required
quantities. Whenever the required quantity (either dependent or independent demand)
is not (fully) satisfied, a backorder is planned for the next period.
It is not possible to show here all results that we obtained from the implementation
of the discussed approach. However, the results of our approach for this particular
supply chain are presented in Table 1. The second column of Table 1 shows the cur-
rent safety stock levels and the third column shows the safety stock levels that result
from our approach based on a target service level of 99%. It was not our intention to
decrease current safety stock quantities, but Table 1 shows that in this case, substantial
savings may be achieved by implementing the approach. However, the comparison
is not completely fair, as in the current situation also other types of uncertainties are
covered.
5.2 The supply chain of product B
The supply chain of product B has also been used for the validation of the proposed
approach. Figure 7 shows the supply chain of product B. Contrary to product A; two
active ingredients (AI 1 and AI 2) are required for the tablet production. After the
production process, the tablets are shipped to two packaging sites where the tablets
are also stored (TB2 and TB3). After packaging of TB2, the products are shipped to 19
national warehouses all over the world, whereas TB3 is shipped to only one national
Fig. 7 The supply chain of 1

product B
AI 1 TB 2
TB 1 19
AI 2 TB 3 20
Table 2 Safety stock levels

Current situation Model suggestion
based on the proposed approach
and the current safety stock
National warehouses 162 136
levels
Tablets stockpoint 71 46
Active ingredients 29 22
warehouse. The same approach has been applied for this supply chain to determine the
safety stock levels. As we mentioned in the case of supply chain of product A, safety
stocks can not be pooled due to the fact that each product is country-specific. Based
on historical demand and forecasts data, the demand generator generates a series of
forecasts that are input to the mathematical programming model that is presented in
Sect. 4.2.
Table 2 presents the results of the proposed approach (to obtain a service level of
99%) and the current safety stock levels. The results show that substantial savings
can be made, but even more important, the approach turns out to give satisfying and
reasonable results.
6 Conclusions
In this paper, we introduced an approach to determine safety stock levels in multi-item

multi-stage inventory systems that face demand uncertainties. The problem of deter-
mining safety stock levels in a supply chain to meet certain predefined target customer
service levels is based on a simulation study where the supply chain planning problem
is solved in a rolling horizon setting. We assume that the supply chain is planned and
controlled by a central authority that sets releases to the production system based on
mathematical programming models. Combining the long run backorder quantities that
result from the simulation study with predefined target customer service levels, the
approach allows for determining safety stock levels in the supply chain.
The approach does not make any assumption about the demand process. Further-
more, all kinds of constraints can be included that are also considered in the supply
chain planning model. The approach is based on two main assumptions. The first
assumption is that the requirement process and replenishment decisions are completely
independent from the safety stock levels. The second assumption is that all unsatisfied
demand is backordered. As a form of validation, we discussed an application of the
approach to two supply chains at Organon, a worldwide operating pharmaceutical
company. The approach helped the company to determine the safety stocks that cover
demand uncertainties.
A shortcoming of our approach is that we assume that any upstream unavailability
of stock leads to an order delay at the next stage, which affects the performance of the
inventory system. This is not necessarily what happens in practice. A short study that
we performed showed that usually protection against a shortage is not only achieved
through the use of safety stocks, but also by using the slack in the lead times or by
reprioritizing the orders such that a higher customer service level is achieved than ini-
tially planned. This effect can be compensated by setting the target customer service
level lower than the ‘real’ target customer service level and this could be an object of
further study.
Acknowledgments The authors would like to extend their word of thanks to Organon, especially to John
Koelink and Joop Wijdeven for initiating and supporting this project. Further, we would like to thank two
anonymous referees for their valuable and constructive comments that helped to improve the clarity of this
paper.
References
Billington PJ, McClain JO, Thomas LJ (1983) Mathematical programming approaches to capacity-
constrained MRP systems: review, formulation and problem reduction. Manage Sci 29:1126–1141
Callarman TE, Mabert VA (1978) Using material requirements planning with demand uncertainty. In:
Proceedings of the 9th annual midwest AIDS conference. pp 151–155
Callarman TE, Hamrin RS (1984) A comparison of dynamic lot sizing rules for use in a single stage MRP
system with demand uncertainty. Int J Oper Prod Manage 4(2):39–49
Clark AJ, and Scarf H (1960) Optimal policies for a multi-echelon inventory problem. Manage Sci 6:
475–490
De Bodt MA, Van Wassenhove LN (1983) Lot sizes and safety stocks in MRP: a case study. Prod Inventory
Manage 24(1):1–16
De Kok AG, Fransoo JC (2003) Planning supply chain operations: definition and comparison of planning
concepts. In: De Kok AG, Graves SC (eds) Design and analysis of supply chains: design, coordina-
tion and operation (Handbooks in Operations Research and Management Science, Volume 11). North
Holland, Amsterdam, pp 597–675
Diks EB, De Kok AG, Lagodimos AG (1996) Multi-echelon systems: a service measure perspective. Euro
J Oper Res 95:241–263
Eilon S, Elmaleh J (1968) An evaluation of alternative inventory control policies. Int J Prod Res 7(1):3–14
Graves SC, Willems SP (2000) Optimizing strategic safety stock placement in supply chains. Manu Serv
Oper Manage 2(1):68–83
Heath DC, Jackson PL (1994) Modeling the evolution of demand forecasts with application to safety stock
analysis in production/distribution systems. IIE Trans 26(3):17–30
Inderfurth K (1991) Safety stock optimization in multi-stage inventory systems. Int J Prod Econ 24:103–113
Inderfurth K, Minner S (1998) Safety stocks in multi-stage inventory systems under different service
measures. Eur J Oper Res 106:57–73
Kleijnen JPC, Wan J (2006) Optimization of simulated inventory systems: OpQuest and alternatives. Tilburg
University: CentER Discussion Paper, no. 2006-75
Kohler-Gudum CK, De Kok AG (2002) A safety stock adjustment procedure to enable target service lev-
els in simulation of generic inventory systems. Technische Universiteit Eindhoven: BETA Working
Paper 71
Law AM, Kelton WD (2000) Simulation modelling and analysis, 3rd edn. McGraw-Hill, New York
Minner S (1997) Dynamic programming algorithms for multi-stage safety stock optimization. OR Spektrum
19:261–271
Silver EA, Pyke DF, Peterson R (1998) Inventory management and production scheduling. Wiley,
New York
Spitter JM, Hurkens CAJ, De Kok AG, Negenman EG, Lenstra JK (2005) Linear programming models
with planned lead times. Eur J Oper Res 163:706–720
Stadtler H (2003) Multi-level lot sizing with setup times and multiple constrained resources: internally
rolling schedules with lot-sizing windows. Oper Res 51(3):487–502
Stadtler H, Kilger C (2005) Supply chain management and advanced planning, 3rd edn. Springer, Berlin
Tempelmeier H, Derstroff M (1996) A Lagrangean-based heuristic for dynamic multi-level multi-item
constrained lotsizing with setup times. Manage Sci 42(5):738–757
Van Houtum GJ, Inderfurth K, Zijm WHJ (1996) Materials coordination in stochastic multi-echelon system.
Eur J Oper Res 95:1–23
Wemmerlöv U, Whybark DC (1984) Lot-sizing under uncertainty in a rolling schedule environment. Int
J Prod Res 22(3):467–484
Whybark DC, Williams JG (1976) Material Requirements Planning under uncertainty. Decis Sci 8(4)
Wijngaard J, Wortmann JC (1985) MRP and inventories. Eur J Oper Res 20:281–293
Supplier managed inventory in the OEM supply chain:
the impact of relationship types on total costs and cost
distribution
P. L. M. Van Nyen · J. W. M. Bertrand ·

H. P. G. Van Ooijen · N. J. Vandaele

OR Spectrum (2009) 31:167–194
DOI 10.1007/s00291-007-0105-4
Abstract We investigate the impact of four variants of supplier managed inventory

on total costs and cost distribution in a capital goods supply chain consisting of a parts
supplier who delivers parts to an original equipment manufacturer’s assembly plant.
The four supplier managed inventory variants differ in the components of inventory
costs that the supplier has to carry. The performance of the supplier managed inventory
relationships is benchmarked with the situation where the assembly plant manages the
inventories. Interesting managerial insights follow from this comparison.
Keywords Supply chain relationships · Vendor managed inventory · Coordination ·

Production–inventory system · Lotsizing
P. L. M. Van Nyen (B)

OM Partners, Koralenhoeve 23, 2160 Wommelgem, Belgium
J. W. M. Bertrand
Technische Universiteit Eindhoven, Faculty of Technology Management,
Den Dolech 2, P.O. Box 513, Paviljoen F08, 5600 MB Eindhoven, The Netherlands
H. P. G. Van Ooijen
Technische Universiteit Eindhoven, Faculty of Technology Management,
Den Dolech 2, P.O. Box 513, Paviljoen F01, 5600 MB Eindhoven, The Netherlands
N. J. Vandaele
Katholieke Universiteit Leuven, Campus Kortrijk,
E. Sabbelaan 53, 8500 Kortrijk, Belgium

220 P. L. M. Van Nyen et al.
1 Introduction
Original equipment manufacturers (OEMs) often subcontract production of parts to

specialized firms that, by working for various customers, can operate at a scale that
allows for the economic operation of their technologies. Such parts manufacturing
shops generally use a number of different technologies grouped into work centers to
manufacture a wide range of parts, with varying routings in the shop and each part
having low to medium demand. These shops are characterized in the literature as job
shops. In this paper we study whether it can be advantageous for a parts manufacturing
shop to engage in a supplier managed inventory (SMI) relationship with its customers.
In such a relationship, it is the supplier who manages the inventory of customer-specific
parts that it produces for the OEMs. For the supplier, a main advantage of engaging
in the SMI relationship is the possibility to optimize production batch sizes, resulting
in lower system-wide costs.
Substantial savings can be obtained from optimizing production batch sizes, both
for the parts supplier from reduced work-in-process and setup costs and for the OEMs
from reduced delivery times and lower safety stocks. However, this requires optimiz-
ing batch sizes from a system-wide perspective, which in turn requires centralized
decision making and implies specific organizational arrangements. An organizational
arrangement that would enable centralized decision making about batch sizes is SMI.
Under SMI, the parts manufacturer, acting as the supplier, would be responsible for
availability of parts at the OEMs assembly plants, and would be free to set a production
and delivery batch size for each of the parts in order to optimize system-wide costs
under a service constraint. In this paper we investigate for which situations it can be
advantageous for a parts supplier who runs a job shop like production system, to engage
in a SMI relationship with his OEM customers. In particular we focus on the advan-
tages that can be obtained from being able to optimize batch sizes. We thus neglect
other advantages that might result from applying SMI, such as reduced transportation
costs or mitigation of the bull whip-effect or improved shop floor scheduling. SMI
also allows for improving the coordination of inventory and transportation decisions,
which may result in considerable cost savings. We will not take into account this effect
in our research. This problem is called the inventory routing problem and is discussed
in a.o. Bell et al. (1983), Campbell et al. (1998) and Bertazzi et al. (2002). SMI can
also contribute to mitigating the bullwhip effect in supply chains. Lee et al. (1997)
analyze the relationship between supplier and customer and identify four mechanisms
that contribute to the amplification of demand variations. One of these mechanisms
occurs if a customer is uncertain about the supplier’s lead time and adapts its reorder
point in response to realized lead times. This mechanism is eliminated if the supplier
manages the inventory. Finally, SMI can also lead to improved shop floor scheduling.
Research of Zheng and Zipkin (1990) has shown that substantial inventory cost reduc-
tions can be obtained from having the production priorities depend on the inventory
position. In this research, we do not consider this effect and assume FCFS sequencing
of production orders.
Our research approach is as follows. We first conceptually analyze the performance
of OEM-managed inventory and four variants of SMI. The conceptual analysis gives
insights into the advantages and disadvantages of the proposed SMI relationship types,
Supplier managed inventory in the OEM supply chain 221
for both the supplier and the OEMs. After the conceptual analysis, a numerical study
is performed. The numerical study is designed such that it allows us to determine the
characteristics of the situations for which it would be beneficial for the supplier to
engage in a SMI relationship, and what relationship type is needed.
The analysis and numerical study is carried out as follows. We have defined five
factors that we expect to influence the advantages to be obtained from system-wide
batch size optimization, and have selected low and high values for each of these fac-
tors. Then we have randomly selected five instances of a 10-products–5-machine job
shop production system that we use as a research tool. For each instance and each
setting of the factors, we have first calculated the batch sizes that would result if each
OEM would optimize its own inventory and ordering costs, and then calculated the
system-wide costs and the distribution of these costs over supplier and OEMs that
would result from the application of these batch sizes. This non-SMI case serves as a
reference for calculating the benefits obtained from SMI. Next we have defined four
variants of SMI, each variant differing in the inventory costs elements incurred by the
supplier. For each of these variants we have calculated optimal supplier determined
batch sizes, and the corresponding system-wide costs and their distribution over the
supplier and OEMs. In this set of experiments, we assume that ordering costs and
inventory holding costs are identical for the OEM and supplier. Optimization of batch
sizes is achieved with a heuristic that incorporates an approximate queuing model
that has been shown to give accurate results (Van Nyen et al. 2005). The numerical
results obtained are analyzed to identify the conditions for which system batch size
optimization is highly advantageous, and the type of supply relationships for which it
is attractive for the partners to engage in such a relationship.
The rest of this paper is organized as follows. In Sect. 2 we review literature on batch
size coordination and on Supplier Managed Inventory. Section 3 presents the supply
chain studied in this paper. Section 4 presents the relationship types and cost models
per relationship type. Section 5 gives a conceptual analysis of the generic effects of
each of the four SMI relationships on the distribution of costs. A numerical study of
the effects of the relationship types on total costs and their distribution is given in
Sect. 6, followed by an analysis of the data. Conclusions are given in Sect. 7.
2 Literature review
Supplier managed inventory is quite common in real life. Especially in the retail
industry this management policy is frequently applied. Most often it is referred to
as vendor managed inventory (VMI). Several case studies on VMI are described in
the literature. Well known VMI implementations include Campbell Soup (Clark and
McKenney 1994), Barilla SpA (Hammond 1994; Simchi-Levi et al. 2000) and the
agreement between Wal-Mart and Procter & Gamble (Cottrill 1997). Other cases
can be found in the automotive industry (Valentini and Zavanella 2003), the food
industry (Tyan and Wee 2003), the retail industry and the health care sector (Gerber
1991). More examples can be found in Andel (1996), Burke (1996), Cottrill (1997),
Holmström (1998) and Waller et al. (1999). Many authors mention the coordination
of production and inventory decisions as a main advantage of VMI. However, to the
best of our knowledge, there is little research on how to coordinate production and
inventory decisions in VMI relationships. Two contributions can be mentioned. Fry
et al. (2001) study the savings due to better coordination of production, delivery and
inventory facilitated by a VMI contract. They focus on a supply chain that consists
of a single retailer and a single supplier, which have a (z, Z ) type VMI agreement.
The (z, Z ) levels correspond to minimum and maximum allowed inventory levels at
the retailer. Bertazzi et al. (2005) study a production–distribution system in which one
item is produced. Two different types of VMI policies are investigated. Both types aim
to determine the production policy, retailer replenishment policies and transportation
policy so as to minimize total system costs. The computational results show that the
VMI policies significantly reduce the average costs compared to the traditional retailer
managed inventory policy. The cost reduction mainly results from improvements in
the transportation costs.
Similar to the work mentioned above, we model supplier–OEM relationships in
order to determine the benefits to be obtained from better coordination of production
and inventory. Specifically we consider the batching decisions. However, our research
is different from the previous work since we focus on a multi-product situation in
which production orders for different products compete for limited production capac-
ity in a production system consisting of multiple work centers. Moreover, we explicitly
consider setup times and setup costs at the work centers. Unlike Fry et al., we do not
allow the supplier to outsource some of its production. Similar to Fry et al., but unlike
Bertazzi et al., we allow the decision maker to act in its own interest, i.e., we model a
decentralized decision maker. Unlike Bertazzi et al., we focus on a production–inven-
tory system, and we do not include transportation issues. Finally, we introduce a wider
spectrum of VMI relationship types than those studied in previous work.
Other research on VMI uses analytical models to analyze the gain of using VMI
in a two-echelon inventory system and to support the vendor with the tactical inven-
tory control decisions. Unlike our approach, these models do not explicitly include
manufacturing operations. In this category, we mention the contributions of Aviv and
Federgruen (1998), Achabal et al. (2000) and Kaipia et al. (2002). Also, analytical
models have been developed to analyze the synchronization of inventory and trans-
portation decisions, see Bell et al. (1983), Campbell et al. (1998), Çetinkaya and Lee
(2000), Axsäter (2001), Cachon (2001), Bertazzi et al. (2002), Cheung and Lee (2002)
and Disney et al. (2003). Furthermore, some analytical models have been developed
to understand the role of VMI in a supply chain channel, see Dong and Xu (2002) and
Disney and Towill (2002). Mishra and Raghunathan (2004) investigate the impact of
VMI on brand competition. Their research shows that VMI intensifies the competition
between competing brands because of brand substitution. Finally, the research on VMI
and other supply chain relationships is strongly related to the research on information
sharing and supply chain coordination. An overview of research on information shar-
ing can be found in Chen (2003) and Huang et al. (2003). For a review of the research
on supply chain coordination, we refer the reader to Thomas and Griffin (1996). Sahin
and Robinson (2002) cover both information sharing and supply chain coordination.
In this paper we study the benefits of determining ordering and production batch
sizes under a SMI relationship, where the suppliers run a job shop like production
system. There is quite some literature on the selection of batch sizes in stochastic job
shop production systems. Contributions in this area are Bertrand (1985), Karmarkar
et al. (1985) and Lambrecht et al. (1998). Vandaele et al. (2000, 2007) present real-life
applications of this kind of batch size optimization models. However, all contributions
focus on the performance of the job shop and ignore inventory holding costs. Van Nyen
et al. (2005) extend the method developed in Lambrecht et al. (1998) to incorporate
both production and inventory costs. In this paper we will use this extended method
to numerically investigate the different relationship types.
3 The supply chain model
We study a supply chain that consists of a parts supplier and OEMs that order parts at
the supplier. The supplier runs a multi work center job shop that produces the parts for
the OEMs. After finishing the production process at the supplier, the parts are kept in
stock at the OEMs. The OEMs take the parts out of the stock when they need them in
the assembly processes. This kind of supply chain can be modeled as a multi-product,
multi-machine production-inventory (PI) system with a single inventory echelon, see
Fig. 1.
We consider K parts (k = 1, . . . , K ) that are kept on stock. The demand for each
product is a stationary renewal process. The demand interarrival times Ak are stochas-
tic variables with a known expectation E[Ak ] and squared coefficient of variation (scv)
c2 [Ak ]. The demand size (number of product units requested per demand) is equal to
one. Demand that cannot be satisfied directly from stock is backordered. The stock
management has to ensure that a target fill rate βk is attained. This type of service
level agreement is common in the parts supply business. The target fill rate βk is set
during contract negotiations between the supplier and the OEM. We assume that this
target fill rate is not renegotiated when changing the relationship type.
The inventory management generates replenishment orders for the different prod-
ucts in order to satisfy the demand according to a (bk , Q k ) continuous review reorder
point inventory policy. Each time the inventory position drops below the reorder point
bk , a replenishment order with batch size Q k is generated. The reorder points bk and
Fig. 1 Supplier–OEMs relationship modeled as a multi-product, multi-machine production–inventory

system
the batch sizes Q k for k = 1, . . . , K are the decision variables. When an order is
generated, a fixed cost ok is incurred. For the items of product k in stock a carrying
charge is incurred. The carrying charge consists of a financial inventory holding cost,
p
h fk and a physical inventory holding cost, h k . The replenishment orders are made-to-
order by the production system. Therefore, the replenishment orders are equivalent to
production orders and production batch sizes are equal to replenishment batch sizes.
The production orders are manufactured in a production system that consists of M
functionally organized work centers. We assume there is ample supply of raw material.
Each of the products requires a specific serial sequence of production steps, which
results in a job shop routing structure. The production orders for different products
compete for capacity at the different work centers. Before the production of an order
for product k at a work center j can start, a machine setup has to be performed. This
machine setup takes a certain time L jk and cost s jk . The setup costs over the entire
routing of product k are denoted as sk . The setup times and costs are sequence and
batch size independent. After the setup, the processing of the production order starts.
The production time for one unit of product k on work center j is given by P jk . The
manufacturing process is subject to variability: setup times and processing times are
stochastic variables with a known expectation and scv. When the production of the
entire batch is completed, the batch is transferred to the next work center on the routing
of the product. These transfer batch sizes are equal to the production batch sizes. A
wip
carrying charge h k is incurred for the work-in-process inventory for each item of
product k that is in process per unit of time. After finishing the last production step, the
entire batch is transferred to the inventory at the OEMs. We assume that the transfer
times are negligible.
4 Relationship types
In this section we present the five supply relationship types, the availability of infor-
mation at the parties as a function of the supply relationships and we introduce the
different cost elements.
4.1 Relationship types
The choice of relationship is a strategic decision since it influences the responsibilities

of the parties, the cost division over the parties and the access to information by the
parties. We consider one non-SMI and four SMI relationship types, based on three
responsibilities that are related to the management of inventories. Each relationship
type defines which party (OEMs or supplier) (1) controls the inventory by setting the
reorder points bk and batch sizes Q k and incurs the ordering costs; (2) incurs the phys-
ical inventory costs; (3) incurs the financial inventory costs. The physical inventory
costs consist of all costs related to the inventory storage, e.g., expenses incurred in
running a warehouse, handling and counting costs, etc. The financial holding costs
consist of all the costs that come with the ownership of inventory, e.g., the opportu-
nity cost of capital, insurance, obsolescence and damage (Silver et al. 1998 Chap. 3;
Table 1 Characterization of supplier–OEM relationships
Non-SMI SMI-NC SMI-S SMI-C SMI-F
Inventory control O S S S S
Inventory storage O O S O S
Inventory ownership O O O S S
O original equipment manufacturer (OEM), S supplier
Valentini and Zavanella 2003). This leads to the definition of the following relationship
types:
1. Non-SMI. The OEMs are responsible for ordering and storing their own parts.
2. SMI–Non consignment (SMI-NC). The supplier is responsible for the inventory
management at the OEM’s premises, but does not incur any inventory costs.
3. SMI–storage (SMI-S). The supplier is responsible for the management and storage
of the OEM’s parts.
4. SMI–consignment (SMI-C). The supplier is responsible for the inventory man-
agement at the OEM’s premises. Moreover, he owns the products until they are
taken from the stock.
5. SMI–full (SMI-F). The supplier is responsible for the management and the storage
of the OEM’s parts. Additionally, he owns the products until they are taken from
the stock.
Table 1 summarizes the five relationship types and indicates the responsibilities of
each party. In the literature we encountered the SMI-NC and SMI-C relationship
(Simchi-Levi et al. 2000). We introduce SMI-S and SMI-F because they correspond
to actual practices in industry and because they have interesting analytical properties.
4.2 Information availability
The access to information is dependent on the relationship type. We make two assump-
tions. Firstly, the party that is responsible for the inventory control has information
about the demand process and the ordering and inventory holding costs. The litera-
ture on information sharing describes that the availability of demand information is
a major benefit for the supplier, who can use the additional information to improve
his production and inventory control decisions. See, e.g. Chen (2003) and references
therein. Secondly, only the supplier has access to production-related information (pro-
cessing times, setup times, routing structure, work-in-process costs and setup costs).
However, we observed that in numerous real-life cases the setup costs are signaled (or
even transferred) to the OEM during supply contract negotiations.
4.3 Cost definition
In this subsection we define the relevant costs. We introduce and model the different
cost elements and we present cost objective functions for the different relationship
types.
4.3.1 Cost elements
We consider five cost elements: ordering costs (OC), financial inventory holding costs
(FIC), physical inventory holding costs (PIC), production setup costs (SC) and work-
in-process holding costs (WIPC). Table 2 presents the allocation of these cost elements
over the parties under the various relationships. The supplier always carries the setup
costs and work-in-process holding costs. In the case of NON-SMI, the OEMs carry
the ordering costs and the physical and financial inventory holding costs. A shift from
NON-SMI to other relationship types implies that the supplier carries more and more
cost components, so that under SMI-F the supplier carries all the costs considered in
this model.
4.3.2 Cost models
We consider five cost components. Table 3 introduces some additional notation. In

p
this paper, we assume that ok , h fk and h k are identical for the OEMs and supplier.
1. Ordering costs.
Under a continuous review, fixed order quantity (bk , Q k ) policy, a new replenish-
ment order for product k is generated when Q k demands have arrived after the previous
replenishment order was generated. The average time between two successive demand
arrivals is E[Ak ]. Therefore, in a (bk , Q k ) policy, a new replenishment order is gen-
erated every Q k E[Ak ] time units, on average. Since each time when a new order is
Table 2 Allocation of costs over OEMs and supplier
Cost element NON-SMI SMI-NC SMI-S SMI-C SMI-F
OC O S S S S
FIC O O O S S
PIC O O S O S
SC S S S S S
WIPC S S S S S
O OEM, S supplier
Table 3 Some additional notation
(Q 1 , . . . , Q k , . . . , Q K )
ssk () safety stock for product k
Tk () throughput time of production orders for product k (stochastic variable)
generated a fixed cost ok is incurred, the ordering costs per unit of time are:
ok
OCk (Q k ) =
Q k E[Ak ]
2. Financial inventory holding costs.

The average amount of inventory that is available at the stock points is given by
Qk
2 + ss(), see e.g., Silver et al. (1998, Chap. 7, p. 258). Therefore, the financial
inventory holding costs are given by:

f Qk
FICk () = hk + ssk ()
2
3. Physical inventory holding costs.

Similarly to the financial holding costs, the physical inventory holding costs are:

p Qk
PICk () = h k + ssk ()
2
4. Setup costs.
In a make-to-order model every replenishment order generates a production order.
Therefore the derivation of the setup costs is similar to the derivation of the ordering
costs. Then, the setup costs are given by:
sk
SCk (Q k ) =
Q k E[Ak ]
5. Work-in-process inventory holding costs.

Using Little’s law, we compute the expected amount of work-in-process inventory
in the production system as the average order throughput time multiplied by the order
1
arrival rate. In our supply chain, the order arrival rate equals Q k E[A k]
and the average
time spent in the production system is E[Tk ()]. Then, the work-in-process holding
costs can be written as:
wi p E[Tk ()]
WIPCk () = h k
Q k E[Ak ]
Above, expressions for the relevant cost components of the supply chain are pre-
sented. In our numerical experiments, we would like to compute these costs for dif-
ferent problem instances. It can be seen from the above formulae that the supply chain
costs depend strongly on the selection of batch sizes = Q 1 , . . . , Q k , . . . , Q K . This
dependence can be either directly, in case of setup and ordering costs, or indirectly,
through the dependence of safety stocks ssk () and throughput times E[Tk ()] on
the batch sizes . For the numerical analysis, we use an approximate analytical model
developed by Van Nyen et al. (2005) to obtain estimates for ssk () and E[Tk ()].
The analytical model is summarized in Appendix 1. Using these estimates, the vector
of batch sizes can be optimized and the corresponding costs can be computed for
specific problem instances.
4.3.3 Cost objectives
The party that is responsible for making the inventory and production control deci-
sions tries to minimize its own costs. Therefore, the cost objective that is optimized
depends on the relationship type. Below, we define a cost objective function for each
relationship type.
NON-SMI In the NON-SMI relationship, the OEMs determine the batch sizes so that
their own costs are minimized, without taking into account the impact on the supplier.
Typically, the batch sizes and reorder points are set sequentially; see e.g. Silver et al.
(1998, Chap. 7, p. 254). In general, we expect the OEMs to be unaware of the cost
parameters at the supplier’s side. However, during contract negotiations the supplier
may signal setup costs (e.g., via batch size dependent prices). Since we do not want to
include pricing mechanisms into the model because this would unnecessarily compli-
cate the analysis, we mimic this effect by assuming that the OEMs determine economic
batch sizes based on inventory carrying costs and the sum of his own ordering costs
and set-up costs at the supplier. In formula:

2(ok + sk )
NON-SMI
Qk =
f p
E[Ak ] h k + h k
Supplier managed inventory Under the SMI relationships, the OEMs transfer their
power for managing the inventories to the supplier. The supplier has access to detailed
demand and process information so that he can coordinate batch sizes with the objec-
tive of minimizing his costs under a service level constraint. The cost components that
the supplier carries under the different SMI variants are listed in Table 2. We modeled
each cost component in Sect. 4.3.2. This leads to the following cost objective functions
for the different SMI relationship types.
1. SMI-NC

K
ok + sk wip E[Tk ()]
+ hk
Q k E[Ak ] E[Ak ]
k=1

Q +
p k
+ hk + hk
f
+ ss() − FICk
NSMI
− PICk
NSMI
2
2. SMI-S

K
ok + sk wip E[Tk ()] p Qk
+ hk + hk + ss()
Q k E[Ak ] E[Ak ] 2
k=1

+
Q k
+ h fk + ss() − FICNSMIk
2
3. SMI-C

K
ok + sk wip E[Tk ()] f Qk
+ hk + hk + ss()
Q k E[Ak ] E[Ak ] 2
k=1

+
p Qk
+ hk + ss() − PICk NSMI
2
4. SMI-F
K

Q
ok + sk wip E[Tk ()] p k
+ hk + h fk + h k + ss()
Q k E[Ak ] E[Ak ] 2
k=1
These cost functions can be interpreted as follows. The sum is taken of all relevant
cost components over all products. The first term within the summation represents
the ordering and setup costs. Under all SMI relationships, the supplier controls the
inventories and incurs the ordering costs. Thus ordering costs are always transferred to
the supplier. The second term is the work-in-process cost. Both terms follow directly
from the cost definition above. The terms related to the financial and physical inven-
tory costs depend on the relationship type and need some more explanation. Under the
SMI relationships, the supplier’s batch size decisions fully determine the inventory
costs. Under the SMI-NC, -S, and -C relationship, the OEMs still carry all or some
of the inventory costs (see Table 2). As a result, the OEM’s inventory costs are deter-
mined by the decisions of the supplier. It seems natural to assume that the OEMs do
not allow increases in their inventory costs after changing from NON-SMI to a SMI
relationship. Therefore, the supplier must ensure that the inventory costs of the OEMs
do not increase in the SMI relationship. We have encountered this kind of arrangement
in real life cases (Kikkert 2006), where an OEM allowed its supplier to control the
inventories at the site of the OEM under a service level constraint and a constraint
on the maximum physical inventory per item. However, imposing a hard constraint
may unnecessarily restrict the search space for the optimization of the batch sizes.
Therefore, we propose to use a soft constraint by assigning a cost when the average
inventory under NON-SMI is exceeded. A similar arrangement was studied in Fry et al.
(2001), where a penalty cost b+ is charged when the inventory position of product k
exceeds the maximum inventory level Z k . The proposed policy can be interpreted as a
compensation mechanism: if the inventory costs under SMI are higher than the inven-
tory costs under NON-SMI, the supplier compensates the OEMs for the cost increase.
This policy gives the supplier sufficient flexibility to optimize his costs by setting the
batch sizes, while the OEMs do not face any increase in their inventory holding costs.
The compensation mechanism is modeled in the objective functions for the SMI-NC,
-S and -C relationship with the terms that contain the operator [a]+ , which denotes
that the maximum of 0 and a is taken. Under SMI-F the supplier carries all inventory
costs, so no compensation mechanism is needed.
5 Conceptual analysis
This section presents a conceptual analysis of the OEM–supplier relationships, under

p
the assumption that ok , h fk and h k are identical for the OEM and supplier. First we
discuss the relationship that leads to the lowest system-wide costs. After this we ana-
lyze the division of costs over the different parties. We introduce two cost effects, the
transfer effect and the coordination effect.
5.1 Supply chain’s optimal cost
In the case of SMI-F the supplier carries all the relevant costs. Therefore, the supplier
can optimize all the costs simultaneously (system-wide optimization), while in the
other relationships only a subset of the costs are optimized (partial optimization). As
a consequence, the total supply chain costs of the SMI-F relationship are lower than
or equal to those of the other relationships and the cost of SMI-F is the supply chain’s
optimal cost:
TRCSMI−F
SC = TRC∗SC
This observation is useful in situations where the achievement of the lowest system-
wide costs is more important than the division of costs over the different parties. This
may occur, for example, when the OEMs and supplier are subdivisions of the same
company. In this case, it is more important to obtain the lowest cost for the whole
company, than to locally optimize the costs of the subdivisions.
5.2 Distribution of costs over the parties: transfer effect and coordination effect
Suppose that initially the supplier and OEMs have a NON-SMI relationship. Fur-
ther suppose that the supplier now wants to change the relationship into a SMI type
relationship. The impact of this change is twofold. Firstly, cost components are trans-
ferred from the OEMs to the supplier. This cost transfer is called here the transfer
effect. Together with the costs, however, also the power for controlling the inventory
is transferred from the OEMs to the supplier. This allows the supplier to determine
the batch sizes, which may results in cost reductions that we refer to as the coordina-
tion effect. In the remainder of this section, the coordination effect and transfer effect
are investigated in more detail. First, we introduce some additional notation. TRCxy
denotes the total relevant costs of a SMI relationship x for a certain party y that result
from minimizing the objective functions in Sect. 4.3.3. OCNSMI

k denotes the ordering
cost for product k in the NON-SMI relationship.
5.2.1 Transfer effect
x can be computed for the different relationships x:

The OEM’s transfer effect TEO
K
TESMI
O
- NC = − OCNSMI
k
k=1
K

TESMI
O
-S =− OCNSMI
k + PICNSMI
k
k=1
K
SMI - C = −
TEO OCNSMI + FICNSMI
k k
k=1
K

SMI - F = −
TEO OCNSMI + FICNSMI + PICNSMI
k k k
k=1
Note that TE xO ≤ 0, since it is a cost reduction. On the other hand, the supplier’s
transfer effect TE xS is a cost increase that consists of all the costs that are transferred
from the OEMs to the supplier:
TESx = −TEO
x
≥0
x is the sum of the transfer effect of the
The transfer effect of the supply chain TESC
OEMs and the supplier, so that the transfer effect for the supply chain is zero:
x
TESC = TESx + TEO
x
=0
5.2.2 Coordination effect
The supplier can reduce his costs in a SMI relationship by coordinating batch sizes.
His coordination effect is characterized by:
CESx ≤ 0
The OEMs are compensated for increases in the inventory holding costs (see Sect. 4.3.3).
Therefore, the OEMs can only benefit from the integration of production and inventory
control decisions, so also their coordination effect is smaller than or equal to zero:
x
CEO ≤0
The coordination of production and inventory decisions results in cost decreases, both
for the OEMs and the supplier. Therefore, the coordination effect always results in
lower supply chain costs:
x
CESC = CESx + CEO
x
≤0
5.2.3 Total costs
The OEMs’ cost in the SMI relationship x equals their cost under NON-SMI plus the
transfer effect and the coordination effect index. This implies that the OEMs have a
guaranteed benefit of a change in the relationship, which is at least as high as the trans-
fer effect TE xO . Therefore, the OEMs have a clear incentive to change the relationship.
x
TRCO = TRCNSMI
O + TEO
x
+ CEO
x
≤ TRCNSMI
O + TEO
x
The supplier’s cost can be computed in a similar way. However, for the supplier
the transfer effect and the coordination effect are opposing and their magnitude is
unknown. Therefore, the total cost effect of a change in the relationship is uncertain.
>
TRCSx = TRCNSMI + TESx + CESx TRCNSMI
S < S
Finally, we discuss the effect of a change in relationship on the supply chain costs.
Since the transfer effect is zero and the coordination effect is always non-positive, a
change from NON-SMI to any SMI relationship will always result in equal or lower
supply chain costs.
x
TRCSC = TRCNSMI
SC + TESC
x
+ CESC
x
≤ TRCNSMI
SC
From the point of view of the OEMs and the total supply chain, a transition from
NON-SMI to SMI always results in lower costs. Therefore, the only uncertainty is the
cost impact of a change on the supplier’s costs. This cost impact will be quantified for
specific instances in the next section.
6 Numerical analysis
A crucial element in the conceptual analysis in the previous section is the magnitude of
the coordination effect. In this section we use numerical analysis of a selected number
of case situations to gain insights into the magnitude of this effect as a function the
case characteristics. This allows us to generate insights into the situations where the
use of SMI type relationships is beneficial.
All cases in the numerical study are variants of the supply chain with 10 products
that are produced to stock in a job shop consisting of 5 work centers with one machine
each. We consider this 10 products, 5 machine supply chain to be sufficiently complex
to capture all effects that occur in larger scale real life systems, while still being suffi-
ciently small to allow for numerical optimization of the batch sizes and safety stocks.
The analysis of integrated production and inventory systems is not straightforward;
see e.g. Zipkin (1986) for an overview of the complexities. We use an approximate
analytical model to estimate the expected order throughput times, the work in process
and the safety stock as a function of the batch sizes for the different products . This
model is summarized in Appendix 1, and extensively described in Van Nyen (2005)
and Van Nyen et al. (2005). We refer the reader to these contributions for technical
details.
For the NON-SMI relationship, the approximate analytical model is used to esti-
mate the replenishment order lead time distribution that emerges when the OEMs set
their batch sizes according to Sect. 4.3.3. Next, safety stocks are set to achieve the
target service level, given this lead time distribution. For the SMI relationships, the
approximate analytical model is used in combination with a heuristic search procedure
to find the vector of batch sizes that minimizes the supplier’s cost objective functions
(as defined in Sect. 4.3.3) and to estimate the corresponding throughput times, safety
stocks and costs.
For the numerical analysis we have generated five instances of our 10 products, 5
machines job shop system. The input data are based on real-life data obtained from
two component suppliers in the OEM market and on the experience of the authors in
the OEM component supply market. Each product always requires processing once
on each machine and the routing of each product is generated randomly. The expected
processing time per item on a machine is randomly generated from the set of values
5, 10, 15, 20 and 25 min. Actual setup and processing times per item on a machine are
exponentially distributed. The products are identical in terms of expectation and scv of
interarrival times and setup times. For each of the five randomly generated instances
we have numerically investigated the supply chain costs for 32 different scenarios
regarding factors that we expect to affect cost effects as we move from NON-SMI to
fully supplier managed inventory. These factors are: the net utilization of the machines,
the variability of demand, the ordering and setup costs, the setup time and the target fill
rate. Each of these factors has been varied over two values: high and low. Specifically,
the following values have been used:
Low High
A. Net utilization of machines (without setup times) 0.70 0.85
B. Scv of interarrival times of demand 0.5 2.0
C. Sum of ordering costs and setup costs ok + sk 100 300
D. Expectation of setup times 150 450
E. Target fill rate 0.90 0.99
This leads to 5 × 32 = 160 different case situations. For all cases, the inventory
p wip
holding costs are set as follows: h fk = 1.67, h k = 0.83 and h k = 1.67 e/item per
year. The sum of orderings costs and setup costs, ok + sk is a factor in our design
p
(factor C), but their split is fixed: oskk = 21 . The cost parameters ok , h fk and h k are
identical for the OEM and supplier.
6.1 Total supply chain cost
For each of the five relationships, we have computed the total supply chain costs for
each of the 32 different scenarios and for each of the 5 randomly generated routing
structures. From the conceptual analysis we know that SMI-F results in the lowest sup-
ply chain costs among the five relationship types. We have calculated for each SMI
Table 4 Percentage cost savings of SMI over NON-SMI
SMI-NC SMI-S SMI-C SMI-F
min 1 (%) 17.7 23.2 25.2 25.7

avg 1 (%) 7.4 10.9 12.6 13.0
max 1 (%) 1.4 3.7 5.1 5.4
Table 5 Results of ANOVA
Source Sum of squares df Mean square F-ratio P value
A. Utilization 0.559152 1 0.559152 2185.39 0.0000∗

B. Scv of interarrival times 4.6092E − 06 1 4.6092E − 06 0.02 0.8934
C. Fixed costs 0.00388724 1 0.00388724 15.19 0.0001∗
D. Expectation of setup times 0.00115679 1 0.00115679 4.52 0.0352∗
E. Target fill rate 0.241044 1 0.241044 942.1 0.0000∗
AB 3.10978E−05 1 3.10978E−05 0.12 0.7279
AC 0.00262453 1 0.00262453 10.26 0.0017∗
AD 0.000315344 1 0.000315344 1.23 0.2688
AE 0.0127712 1 0.0127712 49.92 0.0000∗
BC 0.000598789 1 0.000598789 2.34 0.1283
BD 0.000299138 1 0.000299138 1.17 0.2814
BE 0.000403185 1 0.000403185 1.58 0.2115
CD 0.00194649 1 0.00194649 7.61 0.0066∗
CE 0.000667867 1 0.000667867 2.61 0.1084
DE 0.000678567 1 0.000678567 2.65 0.1057
Blocks 0.000805844 4 0.000201461 0.79 0.5352
Total error 0.0358203 140 0.000255859
Total (corr.) 0.862207 159
relationship the percentage cost savings 1 of SMI over NON-SMI for each of the 160
cases. These percentages are summarized in Table 4, which shows per relationship the
minimum, the average and the maximum relative cost savings.
The data in Table 4 suggest that the benefits resulting from each SMI relationship
type can be quite substantial. The cost savings are the highest for the SMI-F relation-
ship. For this relationship type, we observe average cost savings over the total cost
under NON-SMI of 13%, but the savings can be as high as 25.7%. Moreover, we see
that the simplest SMI-type SMI-NC already results in half of the cost savings obtained
with the full-blown variant SMI-F. The SMI-C relationship achieves almost minimal
costs: on average only 0.5% higher than SM-F. As mentioned before, SMI-F gives the
minimal costs.
Using the randomly generated process configurations as randomizing factor, we
statistically analyzed the effect of the five factors on the cost difference between NON-
SMI and SMI-F. The ANOVA reveals that seven factors have P values less than 0.05,
Fig. 2 Standardized Pareto chart for factors in the ANOVA
indicating that they are significantly different from zero at the 95.0% confidence level
(see Table 5). However, two factors account for a major part of the variance in results,
as can be seen from Fig. 2. These are the net utilization of the machines and the target
fill rate of the products. These results imply that a large reduction in total costs can be
obtained from engaging in a SMI relationship if utilizations are high and/or target fill
rates are high. This can be explained as follows. It is well known that the sensitivity of
throughput times for batch size decisions is high when the utilization of the production
system is high. This effect is illustrated in the much cited paper on lot sizing and lead
times by Karmarkar (1987). Therefore it makes sense that SMI performs much better
than NON-SMI in situations with high-capacity utilization. Moreover, the throughput
time reductions that result from the SMI-F, lead to stronger decreases in the safety
stock costs when the target fill rates are high. In real life, parts suppliers need to operate
at high levels of capacity utilization in order to remain profitable in a market in which
sales prices are continuously under pressure. Moreover, the OEMs are imposing ever
increasing service levels on their parts suppliers. Based on the numerical results, we
expect that supply chains operating in such competitive markets may greatly benefit
from engaging in SMI type relationships.
6.2 Distribution of costs over the parties: transfer effects and coordination effects
The numerical data show that for high machine utilizations and/or high target fill rates,
the supply chain cost savings from implementing SMI can be substantial. However,
when going from a NON-SMI relationship to a SMI relationship, one or more cost
components are transferred from the OEMs to the supplier. The question therefore is
whether it is financially attractive for the supplier to engage in a SMI relationship.
This will be the case if the coordination effect leads to a reduction in the supplier’s
costs that is larger than the costs transferred to him. To check this, we have computed
for each of the 32 different scenarios the decrease in costs relative to NON-SMI under
each of the four SMI relationships.
First, the magnitude of the transfer effect TE xy and the coordination effect CE xy
are calculated. We present the transfer and coordination effect relative to the costs
in the NON-SMI setting. The relative transfer and coordination effect for party y for
relationship x is defined as:
TE xy CE xy
RTE xy = × 100% RCE xy = × 100%.
TRCNSMI
y TRCNSMI
y
We also computed the total relative cost effect of each SMI relationship compared to
the NON-SMI relationship:
TE xy + CE xy
xy = × 100%
TRCNSMI
y
Table 6 summarizes the minimum, average and maximum over the 160 instances of
RTE xy , RCE xy and xy for the OEMs, supplier and the total supply chain in the different
SMI relationship types.
The numerical results in Table 6 contradict a common belief in the literature, namely
that the availability of additional (demand) information and the power to integrate pro-
duction and inventory decisions are sufficient incentives for a supplier to engage in
SMI (or VMI) type relationship. In this set of experiments, it appears that in all cases
for the supplier the transfer effect dominates the benefits of the coordination effect.
This implies that the supplier always faces a cost increase going from NON-SMI to
SMI on this set of experiments. Additional incentives for the supplier that could result
from applying SMI, such as reduced transportation costs or mitigation of the bull
whip-effect or improved shop floor scheduling, are not studied here. The inclusion of
these effects would lead to a better performance of SMI for the supplier. We also see
that under SMI-NC, the simplest type of SMI, the coordination effect for the OEM’s
is already quite high (on average 11.3%), whereas for the supplier the coordination
effect is still quite low (on average 3.7%). For the supplier the coordination effect
strongly increases under more advanced type of SMI, up to on average 26.1% for full
SMI. However, for the supplier also the transfer costs increase, resulting in a strongly
negative total cost effect (up to an average 70.6% increase in costs). This suggests
that the OEMs are the main benefactors of engaging in SMI relationships. In our set
of experiments, the supplier does not have a direct cost benefit of entering into a SMI
type relationship and additional incentives are therefore necessary for making such
a relationship financially attractive. In the next sub-section we will investigate the
coordination effect in more detail.
6.3 Mechanisms behind the coordination effect
In this sub-section, we investigate the mechanisms through which the supplier achieves
the coordination effect on his own costs and on the costs of the OEMs. Both work-in-
process and safety stock are for a large part determined by the throughput times of the
Table 6 Relative transfer effect, coordination effect and total cost effect
x→ SMI-NC SMI-S SMI-C SMI-F
RTE xO
Min −25.0 −50.0 −75.0 −100.0
Avg −18.1 −45.4 −72.7 −100.0
Max −11.7 −41.1 −70.6 −100.0
RCE xO
Min −22.1 −22.8 −13.7 0.0
Avg −11.3 −13.1 −8.8 0.0
Max −2.9 −5.8 −4.9 0.0
xO
Min −39.8 −64.1 −84.4 −100.0
Avg −29.4 −58.5 −81.5 −100.0
Max −19.2 −53.1 −79.4 −100.0
RTESx
Min 13.6 32.8 52.0 71.2
Avg 16.7 43.4 70.0 96.7
Max 19.7 53.1 86.7 120.3
RCE xS
Min −12.5 −23.7 −38.6 −55.7
Avg −3.7 −8.9 −16.6 −26.1
Max −0.2 −1.7 −5.1 −9.6
Sx
Min 1.7 17.8 30.8 42.4
Avg 12.9 34.5 53.5 70.6
Max 19.3 48.9 74.3 97.1
x
RCESC
Min 0.0 0.0 0.0 0.0
Avg 0.0 0.0 0.0 0.0
Max 0.0 0.0 0.0 0.0
x
RCESC
Min −17.7 −23.2 −25.2 −25.7
Avg −7.4 −10.9 −12.6 −13.0
Max −1.4 −3.7 −5.1 −5.4
x
SC
Min −17.7 −23.2 −25.2 −25.7
Avg −7.4 −10.9 −12.6 −13.0
Max −1.4 −3.7 −5.1 −5.4
production orders in the shop. The throughput times in turn are to a large extent deter-
mined by the batch sizes. The relation between the throughput times and batch sizes
in the multi-product production–inventory system can be modeled by the approximate
queueing model presented in Appendix 1. From elementary results from queueing

theory, we learn that there are four main elements that affect the throughput times in
the shop: (1) machine utilization ρ; (2) variation of the interarrival times ca2 ; (3) the
expectation of the processing time E[P]; and (4) the variation of processing times cp2 .
This is represented in the Kingman approximation (1961) for the expectation of the
throughput times in a GI/G/1 queue:

ca2 + cp2 ρ
E[T ] = E[P] + E[P]
2 1−ρ
Batch size decisions affect all these four factors simultaneously. To minimize work-
in-process costs, the supplier should select batch sizes that minimize the average
throughput time. However, under most SMI relationships the supplier also (partly)
incurs costs of keeping finished goods inventory. Since safety stocks are strongly
dependent on the variability of replenishment lead times, the supplier should also take
into account the standard deviation of the throughput times when optimizing the batch
sizes.
Now we present numerical results that illustrate how the batch sizes depend on
the supplier–OEM relationship type. Table 7 gives for each of the five relationships
the batch sizes and the expectation and standard deviation of the throughput times
for one specific problem instance. E[Tk ] and σ [Tk ] are expressed in days. In the case
of NON-SMI, all products have the same batch size because they are symmetrical in
terms of their costs and arrival rates and because the differences between products in
terms of processing requirements are ignored. Going from NON-SMI to SMI-F, we
observe that the average of the batch sizes and the average and standard deviation in the
throughput time decline systematically. Moreover, for all SMI relationships the batch
sizes differ between products in order to take into account the differences in processing
requirements. As a result of the optimization of the batch sizes, the supplier incurs
more ordering and setup costs. This cost increase is compensated by decreases in the
throughput time-related costs (work-in-process and safety stock costs), that follow
from the substantial reductions in the average and variation of the throughput times.
The data in Table 7 illustrate how SMI relationships achieve reductions in sys-
tem-wide costs. It is the supplier who is in a position to realize these cost reductions
by cleverly setting the batch sizes. However, our numerical results indicate that under
each SMI relationship type, the supplier’s cost increase makes it unattractive for him to
engage in such a relationship. It is in the interest of the OEMs to let the supplier make
the batching decisions, since the numerical data in Table 6 show that this substantially
reduces their costs. Therefore, it is in the proper interest of the OEMs to engage in sup-
ply contracts in which the supplier is made responsible for setting the batch sizes. In
order to make such a contract attractive for both parties, the supplier should probably
be compensated for the costs that are transferred. Costs that can be identified easily
by the supplier and agreed upon are ordering costs and financial inventory costs. The
physical inventory cost per part at the OEMs is much more difficult to establish and to
be compensated for. Our numerical results in Table 4 show that the performance under
SMI-C, where ordering costs and financial inventory costs are transferred, is very close
Table 7 Relation between batch sizes and average and standard deviation of throughput times for one
problem instance
k 1 2 3 4 5 6 7 8 9 10 Avg
NON-SMI
Qk 442 442 442 442 442 442 442 442 442 442 442

E Tk 49 55 49 53 61 58 64 60 57 60 57

σ Tk 19 19 19 19 19 19 19 19 19 19 19
SMI-NC
Qk 490 440 490 475 390 430 330 390 430 370 425

E Tk 48 52 48 52 55 55 54 54 53 53 53

σ Tk 17 17 17 17 17 17 17 17 17 17 17
SMI-S
Qk 485 370 510 390 325 360 280 330 360 325 375

E Tk 44 46 45 45 48 48 47 47 46 47 46

σ Tk 15 15 15 15 15 15 15 15 15 15 15
SMI-C
Qk 410 315 430 330 280 310 245 285 310 280 320

E Tk 39 40 40 40 42 42 42 42 41 41 41

σ Tk 13 13 13 13 13 13 13 13 13 13 13
SMI-F
Qk 365 280 380 295 250 275 215 255 275 250 285

E Tk 36 37 36 36 39 38 38 38 37 38 37

σ Tk 12 12 12 12 12 12 12 12 12 12 12
to the performance under SMI-F. Thus SMI-C seems to be a realistic candidate for
modification by including a compensation for taking over ordering cost and financial
inventory costs.
A supplier who considers offering SMI-C to an OEM can, given the demand levels
per part Di indicated by the OEM, use the queueing model in Appendix 1 to calculate
optimal batch sizes Q i and estimate the magnitude of the transfer effect and coordi-
nation effect that would result from applying these optimal batch sizes. Sophisticated
models are available nowadays to perform these calculations for real life production–
inventory systems (see, e.g., Lambrecht et al. 1998 or Van Nyen et al. 2005). Let the
transfer effect TESMI -C and coordination effect CESMI-C for party y be defined as in
y y
Sect. 5.2. Further let pi be the price increase per part demanded by the supplier
under SMI-C. Our numerical results indicate that for the supplier the cost impact
of going from NON-SMI to the SMI-C relationship is always disadvantageous, i.e.
TESSMI-C + CESSMI-C > 0. Then the SMI-C offer would be financially attractive for
TESMI-C +CESMI-C
the supplier only if there is a price compensation: pi ≥ S Di S .
The OEM is the benefactor
of changing the relationship: his costs decrease by
TESMI-C + CESMI-C . In order to convince the supplier to enter into the SMI-C
O O
relationship, the OEM may share some of his benefits with the supplier by paying
a higher price. The OEM would never want to pay more than what he expects to
TESMI-C +CESMI-C
gain from entering in the SMI-C relationship, i.e., pi ≤ O
Di
O
. In this
expression, the benefits of the coordination effect for the OEM are shared between the
supplier and the OEM. However, during real-life negotiations the magnitude of the
coordination effect is not yet known by the OEM, so we may expect that the OEM is
not willing to share these benefits. Therefore, the price increase is typically bounded
TESMI-C
by: pi ≤ O
Di .
It follows that this negotiation game has a solution if there exists a price increase:
SMI-C
TESMI -C + CESMI-C TE
S S O
≤ pi ≤
Di Di
Since TESMI -C = TESMI-C and CESMI-C ≤ 0, this is always the case and we may
S O S
conclude that the supplier can always find a pi that is attractive both for the OEM
and for himself.
7 Conclusions
In this paper we have modeled and numerically analyzed a supply chain consisting of a
parts manufacturer and his OEM customers in order to study the conditions for which
it can be advantageous for the supplier to manage the inventories of his customers. We
consider setup costs, setup time and work-in-process costs at the supplier, and ordering
costs, inventory costs and a service constraint at the OEMs. We proposed five types
of relationships between the supplier and OEMs. In the first relationship the OEMs
manage their inventories and place replenishment orders at the supplier. The other four
relationships are variants of supplier managed inventory (SMI). In these relationships
the supplier manages the inventories at the OEMs and carries one or more compo-
nents of the ordering and inventory costs. If the supplier manages inventories, he can
coordinate batch sizes so as to minimize his own costs. Costs reductions, both at the
supplier and at the OEMs, that result from coordinated batch sizes are referred to as
the coordination effect. We have investigated four SMI variants, one in which only the
ordering costs are transferred to the supplier, one in which either financial or physical
inventory costs are transferred and one full-SMI, in which all costs are transferred to
the supplier.
Numerical analysis of a set of problem instances revealed that substantial system-
wide cost saving can be achieved under all SMI variants, in particular if the shop
operates under high-capacity utilization and/or the OEMs require high service levels.
We have shown that these savings are due to the strongly reduced order through-
put times that are possible if batch sizes can be coordinated. As a result, inventory
costs always decrease under SMI, making SMI attractive for the OEMs. However, for
all problem instances studied, the supplier costs increased under SMI, because SMI
implies the transfer of one or more cost components from the OEM to the supplier.
The OEMs always are better off under SMI. This suggests that a supplier should not
offer SMI to the OEMs unless there is some compensation for his net increase in costs.
We have shown that, since system-wide costs always decrease under SMI, there exists
under each SMI relationship a range of product price increases that at least compensate
the supplier for his increased costs and still make it financially attractive for the OEMs
to engage in the SMI relationship.
In this paper, we assumed that inventory holding costs are identical for supplier
and OEMs. In real life supply chains, this is often not the case. Cost differences may
be caused by differences in the cost of capital (interest rate, required return on invest-
ments, etc.) or differences in the costs of labor and space (labor contracts, location of
warehouses, etc.). In future research, it may be worthwhile to investigate how these
cost differences have an impact on the choice for a certain supply chain relationship
type.
Acknowledgements The authors would like to thank the editors and the referees for their detailed and
very helpful comments.
Appendix I: Modeling the physical supply chain
In this appendix, we present a model to compute the characteristics of the physical

supply chain for a given vector of batch sizes for all products . We determine the
following characteristics of the physical supply chain: (1) the expectation and scv of
the batch interarrival times and batch production times; (2) the expectation and vari-
ance of order throughput times in the production system; (3) the reorder points for the
stock points.
Characteristics of production batches
In the production–inventory system studied in this paper, the generation of a replenish-

ment order results in a production batch. Therefore, we can derive the characteristics
of the production batches by analyzing the characteristics of the replenishment orders.
In a (bk , Q k ) policy, a replenishment order of size Q k is placed every time the inven-
tory position hits the reorder level bk . Consequently, the expected time between two
production batches for product k arriving to the shop is given by: E A0k B = Q E[A ].
k k
Demand interarrival times are assumed B to be i.i.d., so the variance of the interarrival
times of production batches is: σ 2 A0k = Q k σ 2 [Ak ].
Production batches are of fixed size Q k . Consequently, the expected processing

time of a production batch of product k at work center j is given by: E P jk B =

E P jk Q k + E L jk . Since processing times of single units are assumed to be i.i.d.
and independent
from the setup times, the variance of the processing time of a batch

is: σ P jk = σ P jk Q k + σ 2 L jk .
2 B 2
Throughput times in job shop
In this section, we compute the expectation and variance of the throughput times
through the job shop. In general, the arrival and production processes are non-
Markovian processes. This implies that standard queueing theory, e.g., on product
form networks, cannot be used to find performance measures. Instead, we use approx-
imative techniques that were developed by Whitt (1983) to analyze general open
queueing networks.
Interaction between different work centers
In this first step, we analyze the interaction between the different work centers in the
job shop production system. Due to the conservation of flow property, the expected
interarrival time of production batches to each work center j in the routing of product
k is given by:

E A Bjk = E A0k
B
Next, the expected aggregate batch interarrival time of production orders to work
−1 K −1
center j can be computed from: E A Bj = E A Bjk .
k=1
The expected aggregate batch production time at a work center is given by:

K E AB
j
E P jB = E P jk
B
B
k=1 E A jk
The scv of the aggregate production time is given by:

⎡ ⎤
E A Bj K B
E 2 P jk
c2 P jB = ⎣ c2 P jk
B
+1 ⎦−1
2 B
E P j k=1 E A jkB
The utilization of work center j is:

E P jB
ρj =
E A Bj
In the queueing network, the arrival process to a work center is constituted by arrivals
of new batches and by the departure process of batches leaving the previous work cen-
ter in the routing of a product. Therefore, the scv of the interarrival times of batches
of product k at machine i is given by:

M
c2 Aik
B
= c2 AkB rk (0, i) + c2 D Bjk rk ( j, i)
j=1

In this expression, c2 D Bjk is the scv of the departure process of batches of product
k leaving work center j, rk (0, i) is a 0/1 variable that equals 1 if work center i is the
first work center in the routing of product k and rk ( j, i) is a 0/1 variable that equals
1 if work center i is the successor of work center j in the routing of product k. Whitt
(1994) presents an approximation:

c2 D Bjk ≈ ρ 2jk c2 P jk
B
+ f jk ρ 2jl f jl−1 c2 A Bjl + c2 P jlB + 1 − 2ρ jk ρ j + ρ 2jk c2 A Bjk
l=k
where
B
E Aj E P jk
f jk = ρ jk =
E A jk E A Bjk
By combining these expressions, we get a system of linear equations. By solving this

linear system, we obtain the scv of interarrival times for all products and work centers.
After this, an approximation for the scv of the aggregate arrival process at work center
j can be obtained with:

K
c2 A Bj ≈ w j f jk c2 A Bjk +1−w j . In this formula, w j is a weighting function:
k=1
−1
2 −1
K
wj = 1 + 4 1 − ρj n ∗j −1 with n ∗j = 2
f jk .
k=1
Performance measures for individual work centers
Now, the network of interrelated work centers is decomposed into individual

work
centers. We obtain approximations for the expected waiting times E W j , using an
adaptation of the Kraemer and Langenbach-Belz formula proposed by Whitt (1983):

c2 A Bj + c2 P jB ρj
E Wj ≈ E P jB g j
2 1 − ρj
where ⎧ 2
⎪
⎪
⎨ 2(1−ρ j ) 1−c A j
2 B
exp − 3ρ j , c2 A Bj < 1
gj = c2 A Bj +c2 P jB

⎪
⎪
⎩ 1, c2 A Bj ≥ 1

An approximate expression for σ 2 W j , the variance of the waiting times, is due to
Whitt (1983). This approximation is omitted for reasons of brevity. The expectation
and variance of the throughput time of a production batch of product k at work center
j can be approximately computed as:

E T jk ≈ E W j + E P jk B

σ 2 T jk ≈ σ 2 W j + σ 2 P jk
B
Performance measures for the combined network
We approximate the throughput times for the complete production system by consid-
ering every work center to be independent of the others. Then, the expectation and
variance of the throughput times through the job shop are:

M
M

M
M

E[Tk ] ≈ E T jk rk (i, j) σ 2 [Tk ] ≈ σ 2 T jk rk (i, j).
i=0 j=1 i=0 j=1
The assumption that work centers behave independent of each other is not valid in the
general case since this only holds for product-form networks. However, it is common
that queueing network analyzers make this assumption (Whitt 1983).
Reorder points and safety stocks
In this section, we compute the reorder points so that the target fill rate βk is satis-
fied. The computation is based on standard inventory theory, see, e.g., Silver et al.
(1998, Chapter 7, p.253). First, we characterize the average demand during the order
T E[Tk ]
throughput time: E X k k = E[A k]
.
Van Nyen (2005) derives an approximation for the variance of the demand during
the order throughput time for the case of Poisson demand, using approximative results
from renewal theory presented in De Kok (1991):
σ 2 [Tk ] E[Tk ]
T
σ 2 Xk k = +
(E[ Ak ]) 2 E[Ak ]
Now we can fit a distribution function on the two moments of the demand during the
order throughput time. In this research, we use the normal distribution
tocompute the
T
reorder points. The reorder point can be determined by: bk = E X k k + ssk . The

safety stock can be computed as ssk = z k σ X kTk . In this formula, z k is the so-called
safety factor. Silver et al. (1998, p. 736) present an accurate approximation method
for z k for a given target fill rate βk .
Optimization of objective functions
The approximate analytical model presented here can be used to compute the value of
the objective functions for a given vector of batch sizes . Several search algorithms
could be used to optimize the objective functions. In our research, we use a search
algorithm that modifies the batch size Q k of one product at a time. The algorithm
searches in a single direction until no further improvement is possible, while the val-
ues of the other batch sizes are kept fixed. Then, the batch size of another product is
changed until no further improvement is possible. This is repeated until there is no
product for which a further improvement is possible. This final solution cannot be
improved in any direction and is the (local) optimum ∗ . The search procedure has
been extensively tested and its performance proved to be satisfactory. See Van Nyen
(2005) for numerical results on the tests.
References
Achabal DD, McIntyre SH, Smith SA, Kalyanam K (2000) A decision support system for vendor managed
inventory. J Retailing 76:430–454
Andel T (1996) Manage inventory, own information. Logistics Today 37:54–58
Aviv Y, Federgruen A (1998) The operational benefits of information sharing and vendor managed inventory
(VMI) programs. Working paper. Olin School of Business, Washington University, St Louis
Axsäter S (2001) A note on stock replenishment and shipment scheduling for vendor-managed inventory
systems. Manage Sci 47:1306–1310
Bell W, Dalberto L, Fisher M, Greenfield A, Jaikumar R, Kedia P, Mack R, Prutzman P (1983) Improving
the distribution of industrial gases with an on-line computerized routing and scheduling optimizer.
Interfaces 13:4–23
Bertazzi L, Paletta G, Speranza MG (2002) Deterministic order-up-to level policies in an inventory routing
problem. Transport Sci 36:119–132
Bertazzi L, Paletta G, Speranza MG (2005) Minimizing the total cost in an integrated vendor-managed
inventory system. J Heuristics 11:393–419
Bertrand JWM (1985) Multiproduct optimal batch sizes with in-process inventories and multi work centers.
IIE Trans 17:157–163
Burke M (1996) It’s time for vendor managed inventory. Ind Distribution 85:90
Cachon G (2001) Managing a retailer’s shelf space, inventory, and transportation. Manufact Service Opera
Manage 3:211–229
Campbell AM, Clarke L, Kleywegt A, Savelsbergh MWP (1998) The inventory routing problem. In: Crai-
nic TG, Laporte G (eds) Fleet management and logistics. Kluwer, Boston, pp 95–113
Çetinkaya S, Lee CY (2000) Stock replenishment and shipment scheduling for vendor-managed-inventory
systems. Manage Sci 46:217–232
Chaouch BA (2001) Stock levels and delivery rates in vendor-managed inventory programs. Product Oper
Manage 10:31–44
Chen F (2003) Information sharing and supply chain coordination. In: de Kok AG, Graves SC (eds) Hand-
books in operations research and management science vol 11. Supply chain management: design,
coordination and operation. Elsevier, Amsterdam
Cheung KL, Lee HL (2002) The inventory benefit of shipment coordination and stock rebalancing in a
supply chain. Manage Sci 48:300–306
Clark TH, McKenney JL (1994) Campbell Soup Company: a leader in continuous replenishment innova-
tions. Harvard Business School Case. Harvard Business School, Harvard University, Cambridge
Cottrill K (1997) Reforging the supply chain. J Business Strategy 18:35–39
Cox DR (1962) Renewal theory. Methuen, London
De Kok AG (1991) Basics of inventory management: part 1 Renewal theoretic background. Research
Memorandum FEW 520. Department of Economics, Tilburg University, Tilburg
Disney SM, Towill DR (2002) A procedure for the optimization of the dynamic response of a vendor
managed inventory system. Comput Ind Eng 43:27–58
Disney SM, Potter AT, Gardner BM (2003) The impact of vendor managed inventory on transport opera-
tions. Transport Res Part E 39:363–380
Dong Y, Xu K (2002) A supply chain model of vendor managed inventory. Transport Res Part E 38:75–95
Fry MJ, Kapuscinski R, Olsen TL (2001) Coordinating production and delivery under a (z,Z)-type vendor-
managed-inventory contract. Manufact Service Opera Manage 3:151–173
Gerber N (1991) Objective comparisons of consignment, just-in-time and stockless. Hospital Mat Manage
Quart 13:10–17
Hammond J (1994) Barilla SpA (A) and (B). Harvard Business School Case. Harvard Business School,
Harvard University, Cambridge
Holmström J (1998) Implementing vendor-managed inventory the efficient way. Product Inventory Manage
J 39:1–5
Huang GQ, Lau JSK, Mak KL (2003) The impacts of sharing production information on supply chain
dynamics: a review of the literature. Int J Product Res 41:1483–1517
Kaipia R, Holmström J, Tanskanen K (2002) VMI: what are you losing if you let your customers place
orders. Product Plan Control 14:17–25
Karmarkar US (1987) Lot sizes, lead times and in-process inventories. Manage Sci 33:409–418
Karmarkar US, Kekre S, Kekre S (1985) Lot sizing in multi-item, multi-machine job shops. IIE Trans
17:290–298
Kikkert I (2006) Supplier managed and owned inventory bij Neways Advanced Applications B.V. Master
Thesis. Technische Universiteit Eindhoven, The Netherlands
Kingman JFC (1961) The single server queue in heavy traffic. Proc Cambridge Philos Soc 57:902–904
Lambrecht MR, Ivens PL, Vandaele NJ (1998) ACLIPS: A capacity and lead time integrated procedure for
scheduling. Manage Sci 44:1548–1561
Lee HL, Padmanabhan V, Whang S (1997) Information distortion in a supply chain: the bullwhip effect.
Manage Sci 43:546–558
Mishra BK, Raghunathan S (2004) Retailer- vs. vendor-managed inventory and brand competition. Manage
Sci 50:445–457
Sahin F, Robinson EP (2002) Flow coordination and information sharing in supply chains: review, impli-
cations, and directions for further research. Decision Sci 33:505–536
Silver EA, Pyke DF, Peterson R (1998) Inventory management and production planning and scheduling.
Wiley, New York
Simchi-Levi D, Kaminsky P, Simchi-Levi E (2000) Designing and Managing the supply chain. McGraw-
Hill, Boston
Thomas DJ, Griffin PM (1996) Coordinated supply chain management. Eur J Oper Res 94:1–15
Tyan J, Wee HM (2003) Vendor managed inventory: a survey of the Taiwanese grocery industry. J Purchas
Supply Manage 9:11–18
Valentini G, Zavanella L (2003) The consignment stock of inventories: industrial case and performance
analysis. Int J Product Econ 81–82:215–224
Vandaele NJ, Lambrecht MR, De Schuyter N, Cremmery R (2000) Spicer off-highway products division-
brugge improves its lead-time and scheduling performance. Interfaces 30:83–95
Vandaele N, Van Nieuwenhuyse I, Claerhout D, Cremmery R (2007) Load-based POLCA: an integrated
material control system for multiproduct, multimachine job shops. Product Oper Manage (in press)
Van Nyen PLM, Bertrand JWM, Van Ooijen HPG, Vandaele NJ (2005) A heuristic to control integrated
multi-product multi-machine production–inventory systems with job shop routings and stochastic
arrival, set-up and processing times. OR Spectrum 27:399–434
Van Nyen PLM (2005) The integrated control of production-inventory systems. PhD Thesis. Technische
Universiteit Eindhoven. The Netherlands. Available online at: http://alexandria.tue.nl/extra2/
200513458.pdf
Waller M, Johnson ME, Davis T (1999) Vendor-managed inventory in the retail supply chain. J Business
Logistics 20:183–203
Whitt W (1983) The queueing network analyzer. Bell System Tech J 62:2779–2815
Whitt W (1994) Towards better multi-class parametric decomposition approximations for open queueing
networks. Ann Oper Res 48:221–24
Zheng Y, Zipkin PH (1990) A queueing model to analyze the value of centralized inventory information.
Oper Res 38:296–307
Zipkin PH (1986) Models for design and control of stochastic multi-item batch production systems. Oper
Res 34:91–104
Vendor-managed inventory and the effect of channel
power
Bogdan C. Bichescu · Michael J. Fry

OR Spectrum (2009) 31:195–228
DOI 10.1007/s00291-007-0102-7
Abstract We analyze decentralized supply chains that follow general continuous

review (Q, R) inventory policies subject to vendor-managed inventory agreements
where the supplier chooses the order quantity Q, and the retailer chooses the reor-
der point R. Within the VMI scenario, we explore the effect of divisions of channel
power on supply chain and individual agent performance by examining different game
theoretic models. Optimal policies and analytical results, including existence and uni-
queness proofs for equilibrium solutions under VMI, are derived. Numerical results
are provided to compare the effectiveness of VMI and to analyze different channel
power relationships under a variety of environmental conditions. We find that VMI
can result in considerable supply chain savings over traditional relationships and that
the relative division of channel power can significantly effect the performance of VMI.
Interestingly, we find that the greatest system benefits from VMI arise in asymmetric
channel power relationships, but that individual agents lack the incentive to assume a
leadership role.
Keywords Inventory · Game theory · Vendor-managed inventory · Channel power
1 Introduction
Increasing competition and the rapid adoption of advanced information technology

has prompted retailers and suppliers to reengineer their supply chains and examine
B. C. Bichescu (B)
Department of Statistics, Operations and Management Science, College of Business Administration,
The University of Tennessee, Knoxville, TN 37996, USA
M. J. Fry
Department of Quantitative Analysis and Operations Management, College of Business,
University of Cincinnati, Cincinnati, OH 45221-0130, USA

248 B. C. Bichescu, M. J. Fry
collaborative supply chain efforts to reduce costs and improve efficiency. Retailers’
sharing of point-of-sale (POS) data using electronic data interchange (EDI) systems
have become common practice. Vendor-managed inventory (VMI) has emerged in
this context as an initiative that takes the collaborative efforts beyond information
sharing and allows the supplier to exercise some amount of control on the actual
inventory levels at the retailer. Under a typical VMI agreement, the supplier controls
the order quantities delivered to the retailer, possibly subject to contractual limitations
specifying minimum service level requirements, etc. (see Fry et al. 2001). Wal-Mart
and Procter & Gamble (P&G) represent one of the first large-scale successes of such
VMI agreements. Their partnership began in 1985 and significantly improved P&G’s
on-time deliveries to Wal-Mart while increasing inventory turns (Buzzel and Ortmeyer
1995). VMI projects, including those implemented at Dillard Department Stores,
JCPenney and Wal-Mart have shown sales increases of 20–25 and 30% inventory
turnover improvements (Buzzel and Ortmeyer 1995).
Our goal in this paper is to examine the effect of channel power in vendor-managed
inventory agreements in a supply chain. Channel power refers to an agent’s ability to
control the decision making process; it can be a function of the agent’s relative size,
market presence, customer loyalty, etc.
In this paper, we analyze a VMI agreement between a supplier who delivers a single
product to a retailer. The supply chain follows a continuous review (Q, R) inventory
policy, according to which the retailer decides the reorder point R, and the supplier
determines the replenishment quantity Q (a similar model is used in Corbett 2001). In
this way, the supplier controls the delivery amount (Q) and the retailer retains some
control over service levels (R). The retailer incurs an inventory holding cost per item,
per unit time. The supplier incurs an inventory holding cost of his own as well as a fixed
replenishment cost per order. Penalty costs from stockouts are split between supplier
and retailer; this represents the situation where a retailer stockout leads to a loss of
customer goodwill (and possible lost future sales) at both the retailer and supplier.
Production at the supplier occurs at a deterministic linear rate. In cases where the
supplier’s on-hand inventory is insufficient to cover the retailer’s order, we assume
that the supplier can outsource the shortfall from a third party. There exists a positive
delivery lead time and customer demand during the lead time is normally distributed.
The effect of channel power on VMI performance has not been previously exami-
ned. In this paper, we analyze three distinct power relationships within VMI: a power-
ful supplier, a powerful retailer and equally powerful supplier and retailer. Industry
examples exist for all three VMI power scenarios (e.g., Wal-Mart as a powerful retailer,
Barilla SpA as a powerful supplier, etc.—see Buzzel and Ortmeyer 1995; Hammond
1994). We wish to analyze what effect these different channel power relationships have
on channel performance under VMI, as well as examining the effect on the individual
agent decisions and performance. We model channel power similar to Netessine and
Rudi (2004) as the ability of an agent to control the decision-making process in the
supply chain. Specifically, the more powerful firm moves first in a Stackelberg game.
These scenarios are modeled using a game-theoretic approach, with the equal-power
scenario being analyzed as a simultaneous-decision game and the powerful agent
(retailer or supplier) scenarios being analyzed as Stackelberg games. This represents
an important contribution of our paper, as we provide one of the few explicit formu-
Vendor-managed inventory and the effect of channel power 249
lations of channel power in the operations management literature, and the first, to our
knowledge, within the framework of VMI.
To better capture the overall benefits of the VMI agreement, we compare supply
chain performance under a VMI contract with the performance achieved under a
centralized supply chain and under a traditional retailer-managed inventory (RMI)
supply chain. In a centralized supply chain, there exists a single decision maker who
decides on both reorder point and replenishment quantity, while under RMI the retailer
alone determines both the reorder point and the replenishment quantity, but ignores
the supplier’s cost function.
We present analytical results for optimal policies in each scenario, including exis-
tence and uniqueness results for equilibrium solutions under VMI. We then perform a
numerical study to compare supply chain performance and generate insights. We find
that VMI results in considerable savings over RMI scenarios, in general, and that the
relative distribution of channel power can significantly affect the performance of VMI.
We examine the effect of scenario parameters on both the overall performance of VMI
and on the effect of channel power within VMI through a comprehensive, full-factorial
experiment. We also find that the lowest costs at the supply chain level result from
asymmetric channel power relationships, while individual agents lack incentive to take
a leadership role. We discuss implications and possible remedies for this finding.
The remainder of this paper is organized as follows. To better motivate our analysis,
the next section provides a review of the literature relevant to our work. We start our
analysis in Sect. 3, where we formulate the centralized model. Then, in Sects. 4 and 5
we present models for the VMI and the RMI scenarios. Section 6 presents analytical
models for the simultaneous and sequential gaming scenarios under VMI; Sect. 7
presents numerical results. Finally, we present conclusions and managerial insights in
Sect. 8.
2 Literature review
Our current work studies the impact of channel power on the performance of a supply
chain following a continuous review inventory policy under a general VMI agreement.
Thus, our work relates to several existing research streams, including papers that
examine VMI contracts, models for general (Q, R) inventory policies and works that
explore the implications of channel power.
We formulate models that relate to general continuous review inventory models.
Our approach closely follows the approximate cost formulation proposed by Hadley
and Whitin (1963) for a centralized system. This approach is amenable to analytical
solutions of the agents’ best response functions, which are crucial for our purpose of
analyzing the impact of channel power on performance. For this reason, we favor the
approximate model of Hadley and Whitin (1963) over the exact formulation proposed
by Zheng (1992) for a centralized supply chain. For recent developments on continuous
review models, we direct the reader to Federgruen (1993), Hopp et al.(1997), Zipkin
(2000) and Hill and Omar (2006). We build on these existing models by including the
game theoretic framework to account for channel power and extending the models to
both the retailer managed and VMI scenarios.
VMI programs, pioneered by companies such as Wal-Mart, P&G, Campbell Soup

(Clark and McKenney 1994) and Barilla SpA (Hammond 1994), etc. have become
popular in practice due to continued advances in information technology and increasing
supply chain competition. Correspondingly, there are many recent academic papers
that examine vendor-managed inventory. Cachon and Fisher (1997) and Clark and
Hammond 1997) use empirical evidence to compare the benefits of VMI and infor-
mation sharing. They conclude that most of the benefits of VMI could be achieved
by information sharing alone. Narayanan and Raman (1998) formulate a model that
compares RMI with a specific VMI agreement where the retailer basically rents space
to the supplier. Aviv and Federgruen (1998) investigate the benefits of VMI in a supply
chain with one supplier and N retailers under periodic review. The authors assume
that under VMI the timing and magnitude of shipments to the retailers is decided by
the supplier. They find that VMI is always more beneficial than information sharing
alone. Bernstein and Federgruen (2003) study a VMI agreement where, similar to our
paper, decision rights are split between retailer and supplier. However, Bernstein and
Federgruen model a scenario where the supplier controls inventory replenishment and
the retailer determines the product price, whereas we assume price to be exogenous
and we allow the supplier to determine replenishment quantity Q, and the retailer
to choose service level z. Bernstein and Federgruen also assume that the supplier
covers all holding costs, i.e., a consignment-type agreement, where we assume that
the supplier and retailer pay separate holding costs.
Fry et al. (2001) examine a specific type of VMI agreement called a (z, Z ) contract,
between a supplier and retailer. In this setting, the supplier controls the inventory reple-
nishment policy at the retailer subject to limits on service level and maximum inventory
chosen by the retailer (the z, Z quantities). The authors contrast the performance of
the VMI contract with RMI and identify the scenarios where VMI performs best in
comparison to RMI. Cachon (2001) examines coordination in two-echelon supply
chain with one supplier and N retailers. The supplier and retailers follow continuous
review (Q, R) policies, and, as in this research, the total backorder penalty cost is split
between supplier and retailer. Under VMI, the supplier is responsible for choosing the
policies at each of the retailers in the supply chain. Cachon shows that VMI achieves
the optimal solution only if the supplier and the retailers make fixed transfer payments
to participate in the VMI agreement.
Nagarajan and Rajagopalan (2004) also compare the performance of RMI and
VMI under continuous and periodic review policies. As part of the VMI agreement, the
authors explore various subsidizing schemes, e.g., the supplier subsidizes the retailer’s
penalty and holding costs or the supplier subsidizes retailer’s holding cost and the
retailer subsidizes supplier’s replenishment cost. These two parameter contracts are
shown to coordinate the channel under certain conditions. Wu et al. (2005) propose an
(α, h) VMI contract that coordinates a supply chain with one retailer and one supplier.
According to this contract, the supplier acts as a Stackelberg leader, manages the
retailer’s inventory and bears inventory carrying cost; the retailer decides the targeted
sales at the start of the selling season. The contract variable α determines how sales
and penalty costs are split between supplier and retailer; h indicates a holding cost
subsidy paid by the retailer to the supplier for each unit left unsold at the end of the
season.
In a related study, Corbett (2001) examines the impact of incentive conflicts and
information asymmetry on performance in a two-player decentralized supply chain,
which follows a continuous review (Q, R) policy. The author uses a principal-agent
approach to model scenarios where the principal, either the buyer or the supplier, acts
as the leader in the supply chain, but lacks full information on the agent’s costs, while
the agent follows and possesses full information. Corbett finds that in the absence of a
central planner with full information, no party can induce jointly optimal behavior for
all agents in the supply chain without sacrificing his own profits. This conclusion is
related to an important finding in our work, which shows that while the system prefers
one of the agents to lead, neither party has sufficient incentive to exercise channel
power on their own.
Consignment arrangements are often closely related to VMI agreements. Under
consignment, the responsibility of inventory decision making is transferred to the sup-
plier as in VMI, but ownership of goods is retained by the supplier until the moment
of sale (see Bolen 1988; Narayanan and Raman 1998 could also be considered a
consignment-type contract). These contracts are popular in a variety of industries in
various countries (Valentini and Zavanella 2003). Due to their characteristics, such
contracts are often appealing to and initiated by the retailers, who are typically repre-
sented as the powerful players in the supply chain (e.g., see Wang et al. 2004). Thus,
in addition to VMI, consignment agreements may represent a potential environment
where unequal splits of power are likely to exist.
An important contribution of our research is that it explores the impact of channel
power in a VMI agreement. The topic of channel power, defined as the ability of a firm
to influence the intentions and actions of another firm (see Emerson 1962) has seen
some examination in the social, political and marketing literature. However, to the best
of our knowledge, this topic has seen relatively little development in the operations
management literature and, as Cachon (2003) notes, additional research is needed on
this issue. Cachon (2003) states that the use of a profit reservation level and first mover
approaches are possible ways to model power. Here, we adopt the latter approach and
use a game-theoretic framework to characterize players’ actions. For empirical works
that study the issue of power within an operations management framework, we direct
the reader to Maloni and Benton (2000) and Benton and Maloni (2004). For papers
in the marketing literature that study the issue of power defined as the proportion of
channel profits that accrue to each of the channel members and use a game theoretic
approach to determine price, we direct the reader to Choi (1991) and Kadiyali et al.
(2000).
Netessine and Rudi (2004) compare a vertically integrated supply chain, a decentra-
lized supply chain and a drop-shipping supply chain. Under a drop-shipping contract,
the wholesaler, or the manufacturer, ships the product directly to the end customer;
thus, the retailer is relieved of any inventory responsibility and inventory-related costs.
Similar to the models presented here, various channel power structures are analyzed
for the drop-shipping model: a powerful wholesaler, a powerful retailer and equally
powerful wholesaler and retailer. The authors find that drop-shipping is most attractive
when the supply chain has a powerful retailer and least attractive when channel power
is equally split. The implications of channel power are also explored by Bichescu and
Fry (2007) in a supply chain setting where order quantity and shipping frequency are
decision variables and decision rights are split between a retailer and a supplier. Using
game-theoretic concepts, the authors find that a supplier-dominated channel not only
outperforms an equal split of power, but also approaches closely the performance of
the integrated channel. For a more comprehensive discussion on game theory applied
to supply chain models, we direct the reader to the excellent review by Cachon and
Netessine (2004).
3 The centralized scenario
In a centralized scenario a single decision maker makes all relevant decisions to mini-
mize total supply chain cost. Here we formulate a continuous review model for the
centralized scenario to find the optimal reorder point R, and optimal order quantity
Q, that minimize overall system costs. We assume that customer demand follows a
normal distribution, with mean µ and standard deviation σ . There exists a positive
delivery lead√time L, between supplier and retailer. Thus, the reorder point is defined
by µL + zσ L. Note that we use the security level, z, to represent service level and
to serve as a proxy for the reorder point. For analytical tractability we will generally
require z ≥ 0, but we will later discuss the case of z unrestricted.
Per-unit holding costs for on-hand inventory are incurred at both the retailer and
at the supplier, and are denoted as h R and h S , respectively. The retailer experiences
a linear inventory depletion rate (which in expectation is µ) and the supplier has a
deterministic linear production rate (for tractability of results this is set at µ). Cus-
tomer demand that cannot be satisfied from the available inventory at the retailer is
backordered, which results in a penalty cost of p per unit. We assume that when, due
to the temporal uncertainty of retailer orders, the supplier has not produced enough
to cover the retailer‘s order of Q units, the supplier can outsource the shortfall from
an uncapacitated third party at premium unit price b with negligible lead time. Thus,
the retailer always receives her orders in full.1 Each order shipped from the supplier
to the retailer results in a fixed order cost of S. Thus, the elements of the expected
centralized cost function per time unit corresponding to our model are detailed below.
√
• Holding costs (h R + h S ) Q2 + h R zσ L;
µ √
• Penalty costs Q pσ L(z), where (z) = {φ(z) − z[1 − (z)]}; φ(·) and (·)
represent the standard normal pdf and cdf, respectively;
µ
• Shipping costs Q S.
µ
• Shortfall costs b Q E (Q − T µ)+ , where T represents the random variable for
time between orders at the supplier.
µ
Lemma 1 The shortfall cost per period, b Q E (Q − T µ)+ , can be expressed as
bµK, where K > 0 is a scalar value, independent of Q.
Proof See Appendix B.

1 This assumption is common in both the related literature and in practice. Fry et al. (2001), Lee et al.
(2000), and Gavirneni et al. (1999) contain similar assumptions. This reflects the reality in many scenarios
where supplier will go to extraordinary lengths to insure full delivery to retailer such as in automotive
industry, electronics industry, etc.
Lemma 1 proves that the expected shortfall at the supplier is invariant to our deci-
sion variables, Q and z. Thus, no matter what ordering policies are followed, the
expected shortfall costs at the supplier are determined only by the demand distribution
parameters, µ and σ . Therefore, the expected centralized cost function is
µ √ µ Q √
K C (z, Q) = pσ L(z) + S + (h S + h R ) + h R zσ L + bµK. (1)
Q Q 2
Next, we explore the analytical properties of the centralized cost function. We use
Lemma 2 in the proofs of Propositions 1 and 4.
Lemma 2 θ (z) = 2φ(z)(z) − [1 − (z)]2 ≥ 0.
Proof Note that θ (0) > 0 and lim z→∞ θ (z) = 0. Further,
∂θ (z)
= −2zφ(z)(z) − 2φ(z)[1 − (z)] + 2φ(z)[1 − (z)]
∂z
= −2zφ(z)(z) ≤ 0, ∀z ≥ 0.
Thus, θ (z) is a nonincreasing function bounded below by 0 for z ∈ [0, ∞). Therefore,
θ (z) ≥ 0, ∀z ≥ 0.

Proposition 1 K C (z, Q) is convex in z and Q and jointly convex.
Proof In order to build the Hessian matrix, note that
∂ K C (z, Q) µ √ √
= pσ L [−1 + (z)] + h R σ L,
∂z Q
√
∂ K C (z, Q)
2 µpσ Lφ(z)
= ≥ 0.
∂z 2 Q
∂ K C (z, Q) µ √ µ hS + hR
= − 2 pσ L(z) − 2 S + ,
∂Q Q Q 2
√
∂ 2 K C (z, Q) 2µ[ pσ L(z) + S]
= ≥ 0.
∂ Q2 Q3
∂ 2 K C (z, Q) µ √
= 2 pσ L [1 − (z)] .
∂ Q∂z Q
Then,
2 √ √
∂ K C (z,Q) ∂ 2 K C (z,Q) µpσ Lφ(z) µpσ L[1−(z)]
∂z 2 ∂ Q∂z
= Q Q2
∂ 2 K C (z,Q) ∂ 2 K C (z,Q) µpσ √ L[1−(z)] √

∂ Q∂z 2µ[ pσ L(z)+S]
∂Q 2 2
Q Q3
µ2
= 4 p 2 σ 2 L 2φ(z)(z) − [1 − (z)]2
Q
√
+ 2 pSσ Lφ(z)
µ2 2 2 √
= p σ Lθ (z) + 2 pSσ Lφ(z) ≥ 0,
Q4
∀z ≥ 0 from Lemma 2.

Corollary 1 In the centralized scenario, the optimal service level is z C∗ (Q) =

√
−1 µp−h µp
RQ
and the optimal order quantity is Q ∗ (z) =
C
2µS+2µpσ L(z)
h R +h S . In
∗ ∗
addition, Q C (z) monotonically decreases in service level z, and z C (Q) decreases in
order quantity Q.
Proof The expressions for z C∗ (Q) and Q C∗ (z) are obtained by solving ∂ K C (z,Q) = 0
∂z
and ∂ K C∂(z,Q)
Q = 0 for z and Q, respectively. The interdependencies between the two
decision variables are found using the Implicit Function Theorem (IFT) as follows:
∗ (z) ∂ 2 K C (z,Q) √
∂ QC ∂ Q∂z µpσ L[1 − (z)] Q3
=− 2 =− × √
∂z ∂ K C (z,Q) Q2 2µ[ pσ L(z) + S]
∂ Q2
√
Qpσ L[1 − (z)]
=− √ ≤ 0 and
2[ pσ L(z) + S]
∂ 2 K C (z,Q) √
∂z C∗ (Q) ∂ Q∂z µpσ L[1 − (z)] Q
=− 2 =− × √
∂Q ∂ K C (z,Q) Q 2
µpσ Lφ(z)
∂z
2
1 − (z)
=− ≤ 0.
Qφ(z)

4 The VMI agreement
Here we model a VMI agreement, according to which the retailer is responsible for
choosing the optimal service level z, and the supplier chooses the optimal order quan-
tity, Q. We assume that penalty costs from out-of-stocks are split between the retailer
and supplier through a parameter α ∈ [0, 1] so that the retailer’s share of the penalty
cost per unit is αp and the supplier’s share is (1−α) p. This represents the reality where
a stock-out at the retailer results in a penalty to both the retailer and the supplier—
they both incur a loss of goodwill (see Cachon and Zipkin 1999, Nagarajan and
Rajagopalan 2004, among others for similar modeling assumptions). Note that accor-
ding to our VMI setting, the retailer does not transfer his inventory decision rights
entirely to the supplier, but retains a certain degree of autonomy by reserving the
right to choose the service level z. This is different from many previous VMI works,
which assume that the supplier has full control of the retailer’s inventory policy (e.g.,
Aviv and Federgruen 1998; Cachon 2001). Our model is reflective of reality where the
retailer retains some ability to control customer service levels even under VMI (see
Fry et al. 2001 and references therein for supporting evidence) and is closely related
to the split of decision rights modeled in Corbett (2001).
4.1 Retailer’s costs under VMI
In a decentralized supply chain, the retailer and supplier incur separate costs and, thus,
each faces a different objective function. According to the split ofdecision
rights des-
Q
√
cribed above, the retailer’s costs are represented by holding costs h R 2 + zσ L
√
µ
and penalty costs Q αpσ L(z) . Thus, the retailer’s cost function is

µ √ Q √
R
K VMI (z) = αpσ L(z) + h R + zσ L . (2)
Q 2
∗ (Q)) for the retailer to
Proposition 2 demonstrates that there is an optimal z (z VMI
choose when the supplier sets the order quantity, Q.
R (z) is convex in z and is minimized at z = z ∗ (Q).
Proposition 2 K VMI VMI
Proof We have
R (z)
∂ K VMI µ √ √
= αpσ L {−zφ(z) − 1 + (z) + zφ(z)} + h R σ L
∂z Q
µ √ √
= αpσ L [−1 + (z)] + h R σ L and
Q
∂ K VMI (z)
2 R
µ √
= αpσ Lφ(z) ≥ 0.
∂z 2 Q
R (z) is convex and there exists z = z ∗ (Q) that minimizes K R (z).
Thus, K VMI VMI VMI
∂ K VMI
R (z)
Solving ∂z = 0,

∗ −1 µαp − h R Q
z VMI (Q) = . (3)
µαp

Observation R (z) is convex in Q and is increasing in Q when

1 K VMI
√
Q ≥ 2µαpσh R L(z) and decreasing otherwise.
R (z)
The above result is easily proven from the first and second order conditions for K VMI
with respect to Q.
4.2 Supplier’s costs under VMI

The costs incurred by the supplier under VMI are holding costs Q2 h S , shipping costs
√
µ µ
Q S , penalty costs Q (1 − α) pσ L(z) , and shortfall costs (bµK). Thus, the
supplier’s cost function is defined as
µ √ µ Q
S
K VMI (Q) = (1 − α) pσ L(z) + S + h S + bµK. (4)
Q Q 2
S (Q) is convex in Q and is minimized at Q = Q ∗
Proposition 3 K VMI VMI (z).
Proof We have
∂ K VMI
S (Q)
µ √ µ hS
= − 2 (1 − α) pσ L(z) − 2 S + and
∂Q Q Q 2
S (Q)
∂ 2 K VMI 2µ √ 2µ
= 3 (1 − α) pσ L(z) + 3 S ≥ 0.
∂Q 2 Q Q
∂ K VMI
S (Q)
Setting ∂Q = 0 and solving for Q,
√
∗ 2µS + 2µ(1 − α) pσ L(z)
Q VMI (z) = . (5)
hS

From (5), Q ∗VMI > Q E O Q for α < 1 and z < ∞, where Q E O Q = 2µS h S is
the standard economic order quantity (first developed by Harris 1915). An intuitive
explanation for this, also noted by Hadley and Whitin (1963) for the centralized case,
is that expected backorders depend on z but√are independent of Q. Therefore, for
√ √
2µS+2µ(1−α) pσ L(z)− 2µS
α < 1 the supplier will order an additional √
hS
units due
to uncertainty in demand.
S (Q) is decreasing in z.
Observation 2 K VMI
Proof It is straightforward to note that
S (Q)
∂ K VMI µ √
= − (1 − α) pσ L [1 − (z)] ≤ 0.
∂z Q

5 Optimal policies under RMI
Here we analyze a traditional RMI scenario where the retailer decides both the order
quantity Q, and the service level z. Such a scenario represents a situation where the
retailer holds extensive channel power such that she can control all decisions related
to order delivery. Therefore, the retailer’s costs under RMI are composed of penalty
and holdingcosts and are identical to Eq. (2), except that the costs are now a function
of z and Q K RMIR (z, Q) as both are decision variables for the retailer. Proposition 4
and Corollary 2 detail the retailer’s optimal policy under RMI.
R (z, Q) is convex in z and Q and jointly convex.

Proposition 4 K RMI
Proof Note that

√
∂ K RMI
R (z, Q)
µαpσ L [−1 + (z)] √
= + h Rσ L
∂z Q
√
∂ K RMI (z, Q)
2 R
µαpσLφ(z)
= ≥ 0,
∂z 2 Q
√
R (z, Q)
∂ K RMI µαpσ L(z) h R
=− + ,
∂Q Q2 2
√
∂ K RMI (z, Q)
2 R
2µαpσ L(z)
= ≥0 and
∂ Q2 Q3
√
∂ K RMI (z, Q)
2 R
µαpσ L [1 − (z)]
= .
∂ Q∂z Q2
Then,
2 R
∂ K RMI (z,Q) ∂ 2 K RMI
R (z,Q) µαpσ √ Lφ(z) µαpσ √ L[1−(z)]

2 R∂z 2 ∂ Q∂z = √Q Q√2

∂ K RMI (z,Q) R (z,Q)
∂ 2 K RMI µαpσ L[1−(z)] 2µαpσ L(z)
∂ Q∂z ∂ Q2
Q2 Q3

µ2 α 2 p 2 σ 2 L
= 2φ(z)(z) − [1 − (z)] 2
Q4
µ α p2 σ 2 L
2 2
= θ (z) ≥ 0, ∀z ≥ 0 from Lemma 2.
Q4

Corollary 2 In the RMI scenario, the optimal service level chosen by the retai-
∗ (Q) = −1 µαp−h R Q and the optimal order quantity is Q ∗ (z) =
ler is z RMI µαp RMI
√
2µαpσ L(z)
hR . In addition, Q ∗RMI (z) monotonically decreases in service level, z, and
∗
z RMI (Q) decreases in order quantity, Q.
∂ K RMI
R (z,Q)
Proof The optimal solution under RMI is obtained by solving ∂z = 0 and
∂ K RMI
R (z,Q)
∂Q = 0 for z and Q, respectively. Using the IFT,
∂2 K R √
∂ Q ∗RMI (z) RMI
∂z∂ Q µαpσ L [1 − (z)] Q3
=− 2 R =− × √
∂z ∂ K RMI Q2 2µαpσ L(z)
∂ Q2
Q[1 − (z)]
=− ≤ 0 and
2
∗ (Q) ∂ K RMI
2 R √
∂z RMI ∂ Q∂z µαpσ L [1 − (z)] Q
=− 2 R =− × √
∂Q ∂ K RMI Q 2
µαpσ Lφ(z)
∂z 2
1 − (z)
=− ≤ 0.
Qφ(z)

The supplier’s costs under RMI are completely determined by the retailer’s choice
of Q ∗RMI and are given by K RMI
S (Q ∗ ), where
RMI
µ √ µ Q ∗RMI
S
K RMI (Q ∗RMI ) = (1 − α) pσ L(z) + S+ h S + bµK.
Q ∗RMI Q ∗RMI 2
Note that the optimal policy under RMI can be obtained using an iterative scheme
∗ (Q) and Q ∗ (z) until convergence is achieved.
that cycles between z RMI RMI
6 Effect of channel power under VMI
In this section, we explore the implications of the distribution of channel power bet-
ween supply chain agents. We consider three distinct channel power relationships:
a powerful retailer, a powerful supplier and equally powerful retailer and supplier.
Recall that our definition of channel power relates to an agent’s ability to control
the decision-making process in the supply chain. Thus, the powerful agent scenarios
assume that one of the agents, either the supplier or the retailer, has greater bargaining
power and therefore can control the decision making process by making his decision
first; the other agent follows and makes her decision subject to the leader’s choice.
These scenarios are modeled using a Stackelberg game, where the powerful player
acts as the Stackelberg leader. The equally powerful agent scenario assumes that nei-
ther agent has sufficient channel power to control supply chain decisions. In these
conditions, the supplier and the retailer make the inventory replenishment and service
level decisions simultaneously. This scenario is modeled using a simultaneous game
where the solution is characterized by a Nash equilibrium. Full information is assu-
med in all scenarios. This modeling approach has been used by others in the supply
chain research literature to describe channel power, e.g., Choi (1991) and Netessine
and Rudi (2004).
6.1 The case of equally powerful retailer and supplier
Here we assume that the retailer and the supplier simultaneously choose their optimal
strategies under full information to represent a shared power scenario. We first identify
when an equilibrium exists and then when the equilibrium is unique. Note first that the
retailer’s and supplier’s best-response functions are defined by relations (3) and (5),
respectively. From (3), define Q inv (z) as
µαp[1 − (z)]
Q inv (z) = . (6)
hR
From (5), and because z ≥ 0, the feasible range of order quantities is limited to
Q ∈ (Q E O Q , Q ∗VMI (0)). Furthermore, from (6) and z ≥ 0, Q ∈ (0, Q inv (0)). Because
Q inv (z) and Q ∗VMI (z) are monotonically decreasing in z, if Q inv (0) < Q ∗VMI (0), the
retailer’s and supplier’s response functions do not intersect, therefore, no equilibrium
exists in the feasible domain (see Fig. 1). However, if Q inv (0) ≥ Q ∗VMI (0), there
exists a unique interior Nash Equilibrium (see Fig. 2). Proposition 5 formally states
this result.
Proposition 5 For an equally powerful retailer and supplier, there exists a unique
Nash equilibrium in the feasible domain if Q inv (0) ≥ Q ∗VMI (0).
Proof Note that lim z→∞ Q inv (z) = 0 < lim z→∞ Q ∗VMI (z) = Q E O Q . Because
Q inv (z) and Q ∗VMI (z) are both strictly decreasing functions, it follows that if Q inv (0) ≥
Q ∗VMI (0) then Q inv (z) and Q ∗VMI (z) must intersect exactly once.

It can be shown that when z < 0 and Q inv (0) < Q ∗VMI (0), the two response
functions represented by Q inv (z) and Q ∗VMI (z) may not intersect or may intersect
twice, depending on the values of the environmental parameters. However, due to the
complexity of these response functions, we cannot obtain closed form conditions that
precisely characterize the existence of an equilibrium. Therefore, we require z ≥ 0
Fig. 1 No Nash equilibrium in the decentralized scenario µ = 20, σ = 12, p = 10, h S = 0.02, h R = 0.1,
S = 100, α = 0.5, L = 10
Fig. 2 Existence of a Nash equilibrium in the decentralized scenario µ = 20, σ = 12, p = 10, h S = 0.02,
h R = 0.05, S = 100, α = 0.6, L = 10
and Q inv (0) ≥ Q ∗VMI (0) for mathematical tractability in our analysis. We expect that
z ≥ 0 will hold in almost all practical scenarios; however, we note that, as suggested
in Zipkin (2000), when the unit backorder cost is sufficiently close to the unit holding
cost, a company may find it efficient to hold an amount of inventory that is less than
the mean demand during the leadtime.
6.2 The case of the powerful retailer
We model the scenario of a powerful retailer by allowing the retailer to act as a

Stackelberg leader and to choose z first. Assuming full information, the retailer will
take into account the supplier’s optimal strategy when choosing z. Thus, under this
scenario, the retailer will seek to minimize (2), where Q = Q ∗VMI (z). The following
analytical results describe the properties of the Stackelberg equilibrium.
Proposition 6 When the retailer is the Stackelberg
√ leader, there exists a unique
Stackelberg equilibrium, if S > 0.1575(1 − α) pσ L.

Note that for reasonable values of penalty cost p, demand standard deviation σ , and
lead time L, the condition stated in Proposition 6 holds, especially as α approaches
1 and thus (1 − α), approaches 0. In all our numerical trials the condition from
Proposition 6 is satisfied.
6.3 The case of the powerful supplier
In this section, we assume that the supplier is the Stackelberg leader and, thus, chooses
his optimal order quantity Q, first, knowing the retailer’s response function is repre-
sented by (3). The supplier’s optimal strategy is obtained by minimizing (4), when
∗ (Q).
z = z VMI
Proposition 7 When the supplier is the Stackelberg leader, there exists a unique
Stackelberg equilibrium.

7 Numerical results
In this section, we conduct a numerical study that provides insights on the performance
of VMI and how channel power affects supply chain performance under VMI. We
examine the effect of various environmental factors on VMI performance and the effect
of channel power through a comprehensive full-factorial experiment. The factors are
coefficient of variation of customer demand (CV), ratio of holding cost at the retailer
to holding cost at the supplier (HR), shipping cost (SC), unit penalty cost for stockouts
(P), shipping lead time (L) and retailer’s share of the penalty cost (α). Each of these
factors is assumed to take low, medium and high values as shown in Table 1 below and
are representative of practical values in the United States electronics and automotive
industries. The mean customer demand per period µ, and the unit holding cost per
period at the supplier h S , are held constant at 20 and 0.02, respectively, throughout
this study. Note that a consignment-type system could be considered by setting HR
≤ 1.0 (i.e., allowing for h S ≥ h R ). We have considered such situations in extended
numerical results and have found no differences in insights; thus, we present only
cases where HR > 1.0 here. Further, we assume that outsourcing cost b, is negligible
since the shortfall cost term, bµK is independent of both z and Q and, thus, invariant
to our decision making.
To capture how the performance of the centralized, VMI and RMI scenarios com-
pare, we use two measures, Decentralized Performance Gap (ϒ) and VMI Cost Reduc-
tion Factor ( ), where
VMI system cost − centralized system cost

ϒ= and
VMI system cost
RMI system cost − VMI system cost
= .
RMI system cost
Table 1 Data for the

Factor Level
experimental design
Low Medium High
CV 0.10 0.30 0.60

HR 1.50 2.50 5.00
P 10 15 20
SC 50 75 100
L 3 5 10
α 0.60 0.75 0.90
Thus, ϒ shows the performance gap between the centralized and VMI scenarios,
while measures the cost savings of moving from RMI to VMI. Note that our defi-
nition of ϒ is similar to the concept of price of anarchy, which is used in economics
and computer science to measure the extent to which uncoordinated selfish decisions
degrade the performance of a system compared to the global optimum.2 The effects
of channel power are specifically measured by the Stackelberg Cost Improvement
indicator, which captures, within the VMI model, the percentage cost reduction of
the powerful retailer (δ) and powerful supplier () scenarios over the equal power
scenario. This cost improvement is measured at both the agent and system level. Spe-
cifically, let δ R , δ S and δ, respectively, represent the retailer’s, supplier’s and supply
chain’s percentage cost improvements
when moving from an equal power scenario
to
R (Nash)−K R (Stackelberg, Retailer−led)
K VMI
a retailer-led VMI scenario e.g.,δ R = VMI
R (Nash) ; let
K VMI
R , S and represent similar cost improvements
for a supplier-led supply chain

K R (Nash)−K VMI
R (Stackelberg, Supplier−led)
e.g., R = VMI .
R K VMI (Nash)
7.1 Performance comparison of centralized, VMI and RMI scenarios
Tables 4 and 5 in Appendix A display the results of our numerical study concerning
how the performance of the centralized and decentralized scenarios compare. The
columns showing the Decentralized Performance Gap ϒ, and the VMI Cost Reduction
Factor , in Table 4 illustrate the relative performance of the centralized, VMI and
RMI scenarios. System costs are computed for each of the three different VMI channel
power scenarios, thus separate ϒ and values are reported for each scenario in Table 4.
Table 5 contains results on the optimal policies, represented by stocking factor z, and
order quantity Q, and the corresponding supply chain cost under the centralized, VMI
and RMI scenarios. To build a better understanding of the performance differences,
Table 4 displays the average, minimum and maximum cost improvement values at
each level of the six environmental factors. This allows for a clearer identification
of the range of savings offered by moving from RMI to VMI ( ) and from VMI to
centralized control (ϒ).
The results in Table 4 show that the performance penalty resulting from decentrali-
zed decision-making is significantly influenced by environmental factors, the perfor-
mance gap ϒ averaging 16.60% over all channel power scenarios and ranging from
a minimum of 6.43% to a maximum of 31.85%. Further, note that this performance
gap is consistently lower in the asymmetric power cases and lowest when the supplier
acts as the Stackelberg leader. Also, the performance gap in the supplier-led cases is
consistently lower then in the retailer-led cases (see Sect. 7.2 for a detailed discussion
of the impact of channel power on performance). In addition, according to Table 5,
the centralized scenario achieves customer service levels that outperform VMI in all
cases; however, these values are lower than those obtained in the RMI scenario. This
2 The price of anarchy is defined as the ratio of the Nash equilibrium solution to the system optimum. For
additional details see Papadimitriou (2001) and references therein.
is attributable to the position of absolute power that the retailer enjoys in the RMI
case, such that the retailer can obtain high service levels that minimize her costs while
ignoring supplier’s costs.
Table 4 shows that VMI leads to significant savings over RMI, regardless of the
channel power relationship. The VMI Cost Reduction Factor , averages 86.02%
over all scenarios. However, ranges from a minimum of 29.60% to a maximum
of 98.65%, indicating that the magnitude of savings offered from moving from RMI
to VMI are also highly dependent on scenario parameters. Table 5 indicates that the
retailer’s power to completely control the decision-making process under RMI leads
to much smaller order quantities (Q) than under any of the VMI scenarios or under
centralized control. Because the retailer has full power under RMI and disregards any
costs incurred by the supplier, she sets Q quite low resulting in high shipping costs
for the supplier and, hence, excessive system costs. These findings are consistent with
existing evidence in the literature, e.g., Fry et al. (2001), Nagarajan and Rajagopalan
(2004), which shows that a well-designed VMI contract outperforms traditional RMI
in many realistic scenarios.
Table 5 also shows that VMI scenarios tend to lead to order quantity values that
are higher than a centralized solution. Again, this reflects the greater power of the
supplier in all VMI scenarios (compared to RMI or to centralized control) since the
supplier controls the order quantity solution and he prefers larger Q values to reduce
shipping costs. However, the split of channel power between retailer and supplier in
VMI prevents the extreme solutions seen under RMI so that VMI leads to considerable
system cost savings. We also note that Q values remain fairly constant across the
channel power scenarios under VMI. This is because the supplier’s choice of Q is
somewhat insensitive to changes in z under VMI.3 Thus, the order quantity changes
very little across the different power scenarios under VMI, but there are significant
differences in the safety stock parameter z.
In other words, inventory policies will have similar order quantities across the dif-
ferent VMI power scenarios, but the policies will differ mainly in customer service
levels through different reorder points. Not surprisingly, the highest customer service
levels (z) are offered when the retailer has greater channel power; however, these ser-
vice levels are still less than those preferred in the centralized scenario. As mentioned
previously, both ϒ and vary significantly with the environmental factors. To better
understand when the penalty from decentralized decision-making is largest and when
moving from a traditional RMI setting to a VMI agreement would provide the highest
benefit, we perform a full-factorial experimental design whose main results regarding
ϒ and are summarized in Table 2, where strong positive (negative) direct effects
are represented by “↑” (“↓”), weak positive (negative) direct effects are represented
by “ ” (“”) and cases where a significant direct effect could not be identified are
represented by “−”.
∗
3 This can be verified by examining ∂ Q C (z) in the proof of Corollary 1 and noticing that as (z) → 1,
∂z
∗ (z)
∂ QC
∂z → 0. Because z values in equilibrium are relatively large (z > 1.2 in all VMI scenarios in Table 5,
∗ (z)
∂ QC
hence (z) > 0.88), we expect ∂z to be near zero.
Table 2 shows that holding cost ratio HR, demand uncertainty CV and delivery
lead time L, are the factors that have the strongest effect on the performance gap
between centralized and VMI scenarios. When holding costs at the retailer increase, the
centralized system responds with lower order quantities and more frequent shipments
to the retailer, while the VMI scenarios adopt an opposite strategy which leads to
higher holding costs and, thus, lower performance. Further, the larger order quantities
characteristic to the VMI scenarios allow for a better response under VMI to increases
in demand uncertainty relative to the centralized scenario. Thus, we find that the
performance of the VMI scenarios is close to first-best solutions represented by the
centralized case. The performance gap is largest when the holding cost ratio, shipping
cost and the split of penalty costs α, are high, and demand uncertainty and delivery lead
time are low. Under such environmental conditions, a coordination contract would be
most beneficial.
Figure 3 displays the significant direct effects on VMI Cost Reduction Factor .
According to Fig. 3, demand uncertainty CV, delivery lead time L, and holding cost
ratio HR, are factors with a strong negative impact on the cost savings of VMI. As these
factors increase RMI system costs decrease while VMI system costs (in all channel
power scenarios) increase. Thus, the VMI Cost Reduction Factor , decreases. Our
one-way analysis of variance (ANOVA) results show that the unit penalty cost for
stockouts P, and the retailer’s share of the penalty cost α, are not significant at the
5% level. However, while not explicitly shown here, we can report that multiple-way
Table 2 Direct effects on ϒ

Factor Decentralized performance gap (ϒ)
Equal power Retailer—Led Supplier—Led
CV ↓ ↓ ↓
HR ↑ ↑ ↑
P − −
SC
L ↓ ↓
α ↓ −
100%
Factors CV HR SC L
95%
Percentage Change in
90%
85%
80%
75%
70%
Low Med High
Factor Level
Fig. 3 Significant direct effects on

ANOVA shows significant interactions between P and CV and HR and between α

and CV and HR. Thus, these cross-effects may mask the direct effects of P and α on
, explaining why P and α are non-significant. Note also that higher penalty costs
result in higher system costs under all scenarios. Thus, we notice that system costs
increase proportionally in the VMI and RMI scenarios as P increases, representing
an alternative explanation as to why P has little influence on the overall variance of
. Increasing shipping cost SC, clearly leads to higher RMI system costs because the
retailer is completely insensitive to shipping costs, thus Q and z remain unchanged as
SC increases, resulting in higher supplier, and system, costs. Therefore, the effect of
higher shipping costs is stronger under RMI than in the VMI scenarios, so increases
with SC.
In conclusion, we find that VMI consistently leads to savings over RMI, regardless
of channel power allocations; however, relative cost savings are highest when demand
uncertainty, holding cost ratio and delivery lead time are low and shipping cost is high.
Here we must point out that while we believe our RMI model to be an appropriate
approximation of system behavior for comparison purposes, the actual cost values
incurred under RMI in our model are most likely worst-case estimates. In our model
we assume that the retailer acts with complete disregard to the supplier’s costs under
RMI. Such a strategy may be optimal for the short-term, but obviously would not
be sustainable over the long-term. In reality, a retailer may implicitly increase order
quantities to help suppliers offset excess shipping costs (which is why, in reality, many
suppliers require minimum shipping quantities, which are also not part of our model).
7.2 Channel power analysis
Our main results regarding the effect of channel power on supply chain performance
under VMI are captured in Table 6 in Appendix A. All values in Table 6 are positive,
indicating that unequal channel power VMI scenarios are always at least as efficient
as the equally powerful VMI scenario both at the supply chain and individual agent
levels. This result is consistent with findings in the economics literature, e.g., Boyer
and Moreaux (1987), Amir and Stepanova (2006) and in operations, e.g., Netessine
and Rudi (2004). An intuition for this result can be developed by comparing the cen-
tralized and decentralized solutions and the agents’ best response functions. We know
that the best response functions are monotonically decreasing and that the centrali-
zed system achieves high service levels with relatively low order quantities. Further,
previous work has shown that the Stackelberg leader is better off than under the Nash
equilibrium (see for example Netessine and Rudi 2004). Therefore, the leader will
select a solution closer to the integrated solution. Thus, if the retailer is the leader, she
will select a higher z compared to the Nash solution. This leads to lower costs for the
supplier as well, as its cost function is decreasing in the leader’s choice (see Obser-
vation 2). Similarly, when the supplier leads, the order quantity Q, will be smaller
compared to the Nash equilibrium. Given that the retailer’s cost function is increa-
sing in the supplier’s decision in all our numerical trials (specifically, the condition in
Observation 1 is satisfied for optimal ranges of z and Q in all cases), the supplier will
be better off following in the Stackelberg game.
More specifically, our results show that the cost improvement of the retailer-led
VMI scenario over simultaneous gaming ranges from ∼ 0 to 11.17% and the cost
improvement under the supplier-led scenario ranges from ∼ 0 to 16.72%. Furthermore,
≥ δ in each parameter scenario, thus the supply chain consistently performs best
when the supplier acts as the Stackelberg leader.
We are interested next in identifying which scenario parameters have the greatest
effect on determining when channel power has the greatest impact on supply chain
savings under VMI. The effects of channel power are shown in Table 6 in the columns
labeled δ and . According to the results in Table 6 we find that the largest cost
improvement over the equal power scenario is 16.72% which is achieved when the
supplier is the Stackelberg leader. Note that this is also the case where a retailer-led
supply chain achieves the greatest cost reduction, but the savings are only 11.18%. To
further explore the effect of scenario parameters on channel power savings, we again
perform a full-factorial experiment, this time measuring Stackelberg Cost Improve-
ment (δ and ). The direct effects of the factors are shown in Table 3. We find that
holding cost ratio HR, and coefficient of variation CV, have the strongest effect on
Stackelberg Cost Improvement and that this applies to both the retailer-led (δ) and
supplier-led () scenarios. Lead time L, has a moderate positive impact, while α has
a significant negative impact and shipping cost SC, appears to be insignificant. Again,
the parameters that are significant and their effects are consistent across supplier-led
and retailer-led VMI scenarios and at the agent and system level (with the exception
of penalty cost being insignificant in its effect on δ S ). Thus, it appears that moving
from an equal power VMI scenario to an asymmetric channel power relation is most
beneficial when demand variability, holding costs, and lead times are high, penalty
costs are low and the supplier’s share of penalty costs are large (small α). Furthermore,
this applies equally to whether the system is moving to greater retailer channel power
or greater supplier power; however, system cost savings are somewhat greater when
channel power is concentrated at the supplier (as can be seen in Table 5).
Interestingly, Table 6 also shows that δ R < R and S < δ S for each parameter
scenario. In other words, the agent incurs lower cost when s/he is the follower than
when s/he is the leader in the Stackelberg game. Proposition 8 somewhat formalizes
this result from the supplier’s perspective. Note that in all our numerical results (see
Table 5) service levels and order quantities are ordered such that z R ≥ z S ≥ z N
and Q N ≥ Q R ≥ Q S , where the subscript indicates the channel power distribution
Table 3 Direct effects on

Factor Stackelberg cost improvement
Stackelberg cost improvement
δR δS δ R S
CV ↑ ↑ ↑ ↑ ↑ ↑
HR ↑ ↑ ↑ ↑ ↑ ↑
P −
SC − − − − − −
L
α ↓ ↓ ↓ ↓ ↓ ↓
under VMI (N = Nash, S = powerful supplier, R = powerful retailer). Proposition 8

analytically demonstrates that if this condition holds, then the supplier always prefers
to be the follower.
Proposition 8 If z R ≥ z S ≥ z N and Q N ≥ Q R ≥ Q S , then

S
K VMI (z N , Q N ) ≥ K VMI
S
(z S , Q S ) ≥ K VMI
S
(z R , Q R ).
Proof Note that

√
S (Q)
∂ K VMI µ(1 − α) pσ L [−1 + (z)]
= ≤ 0. (7)
∂z Q
Combining (7) with z S ≥ z N , we have K VMI S (z , Q ) ≥ K S (z , Q ). Given

N N VMI S N
that Q S represents the supplier’s best response to a given service level z S , we have
S (z , Q ) ≥ K S (z , Q ). Further, because of (7) and z
K VMI S N VMI S S R ≥ zS,
K VMI (z S , Q S ) ≥ K VMI (z R , Q S ). Since Q R is the supplier’s best response when the
S S
S (z , Q ) ≥ K S (z , Q ). Therefore, K S (z , Q ) ≥
retailer chooses z R , K VMI R S VMI R R VMI N N
K VMI (z S , Q S ) ≥ K VMI (z R , Q R ).
S S

Our numerical results indicate that we can make an even more general statement
here since in all results
R
K VMI (z N , Q N ) ≥ K VMI
R
(z R , Q R ) ≥ K VMI
R
(z S , Q S ) and
K VMI (z N , Q N ) ≥ K VMI (z R , Q R ) ≥ K VMI (z S , Q S ),
where K VMI represents the cost to the supply chain as a whole. Thus, our results imply
that agents prefer to be the follower rather than the leader in a Stackelberg game, thus,
there exists a first-mover disadvantage under VMI. This finding is consistent with the
results shown in Gal-Or (1985) and Amir and Stepanova (2006) for their duopoly
models, Netessine and Rudi (2004) for their drop-shipping model and Corbett (2001)
for his analysis of information asymmetry on performance. Desiraju and Moorthy
(1997) also report cases of a first-mover disadvantage in a price leadership scenario,
e.g., a retailer who acts as the Stackelberg leader announces a retail price; the manu-
facturer (the follower in the game) will charge a wholesale price as close as possible
to the retail price, leaving just enough retail profit to keep the retailer interested. We
notice a similar behavior in our setting, where the Stackelberg follower extracts the
majority of the cost savings resulting from asymmetric channel power, leaving the
leader with just enough cost benefits to motivate him to move from equal power.
The above result leads to the interesting conclusion that at the agent level, channel
power does not seem to translate into lower costs. Thus, agents are not motivated
to exercise channel power even when they have it. However, at the system level,
costs are clearly lowest when there is some form of asymmetric channel relationship,
regardless of whether power is concentrated at the retailer or the supplier. Therefore,
in order to achieve the greatest supply chain efficiency, one of the agents must be
enticed to exercise channel power (i.e., assume Stackelberg leadership). This can be
accomplished through some sort of transfer payment between agents (see Cachon 2001
for a discussion of VMI and side payments to achieve channel coordination) or perhaps
a pairing of VMI with additional collaborative efforts such as holding cost subsidies
(see Nagarajan and Rajagopalan 2004 for an example of such an arrangement).
8 Conclusions
In this paper we analyze a decentralized supply chain operating according to a vendor-

managed inventory (VMI) agreement and a continuous review (Q, R) inventory policy.
Under VMI, the supplier determines the order quantity amount Q, to send to the
retailer, while the retailer retains control of the reorder point R. To capture the effect
of channel power in the supply chain on the performance of VMI, we develop three
different models for VMI corresponding to (1) a supply chain with a powerful supplier
that can lead the decision making process; (2) a supply chain with a powerful retailer;
and (3) a scenario where both agents have approximately equal channel power. The
powerful agent scenarios are modeled as Stackelberg games with the powerful agent
as the Stackelberg leader; the equal-power scenario is modeled as a simultaneous
decision game. Optimal policies are found for each scenario in the form of equilibrium
solutions for Q and R. Models are also developed for the centralized scenario where
one decision maker chooses both Q and R and a traditional retailer-managed inventory
(RMI) scenario for comparison purposes.
Our analysis shows that VMI leads to supply chain savings in many scenarios,
regardless of the channel power relationship. We find that VMI produces the greatest
savings over RMI when demand uncertainty, holding cost ratio and delivery lead time
are low and shipping cost is high. However, we find that the amount of savings can
be significantly affected by channel power in the supply chain. In all cases system
costs are lower when there is an asymmetric power relationship than under an equal
power scenario. Interestingly, however, the lowest costs are incurred at the agent level
by being a follower and not a leader in the Stackelberg game. Thus, while the system
prefers either the retailer or supplier to lead, neither has incentive to exercise channel
power on their own. Therefore, some additional incentive must be provided to an agent
to accept the leadership role in order to reduce system costs. This additional incentive
could take the form of side payments or additional collaborative mechanisms. We also
find that the lowest system costs are incurred when the supplier is the Stackelberg
leader. This appears to be contrary to the current trend of channel power shifting from
suppliers to retailers, such as is described in a Wall Street Journal article discussing
the shift in channel power from suppliers such as Levi Strauss to retailers such as
Wal-Mart (Wall Street Journal, June 17, 2004). Based on our findings, such power
concentration at the retailer may, at least in VMI systems, lead to costs that are higher
than a system where power is concentrated at the supplier, although such a system will
still be more efficient than a supply chain where power is equally distributed between
supplier and retailer.
This research represents an initial step to better understand channel power. However,
we believe that, due to its far reaching implications, this topic deserves additional
attention in the operations literature. Thus, future research can investigate settings
where multiple competing retailers sell the product of one supplier to explore whether
competition can mitigate the first-mover disadvantage that we note in our work (e.g.,
see Bernstein and Federgruen 2003, 2004, 2005). Alternatively, one may consider the
case of multiple suppliers selling substitutable or complementary products through a
common retailer (see for instance Netessine and Rudi 2003, Wang 2006) or may allow
customer demand to be price sensitive, as explained in Petruzzi and Dada (1999).
Another possible extension is to relax the assumption of full information and analyze
cases where the supply chain agents may hold asymmetric information on each other’s
cost parameters (e.g., Corbett 2001; Corbett et al. 2004). In all cases, it is expected
that these extensions will lead to considerable modeling challenges. However, these
extensions incorporate many of the complications encountered in practice and should
lead to models that better represent the actual influence that channel power exerts on
supply chain agents’ decision making.
Appendix A
Table 4 System performance comparison of centralized and decentralized scenarios (%)
Factor Factor Measure Decentralized performance gap (ϒ) VMI cost reduction factor ( )
value Nash Ret–Led Sup–Led Nash Ret–Led Sup–Led
Avg 18.3 18.1 17.9 96.44 96.45 96.46

0.1 Min 9.1 9.0 9.0 90.93 91.12 91.25
Max 30.1 29.4 29.0 98.66 98.66 98.66
Avg 17.3 16.7 16.1 87.73 87.87 87.99
CV 0.3 Min 8.0 8.0 7.8 67.21 69.17 70.36
Max 30.6 28.2 27.2 95.56 95.56 95.56
Avg 16.1 15.0 13.9 73.21 73.77 74.24
0.6 Min 6.7 6.7 6.4 29.60 37.46 41.37
Max 31.9 26.5 24.6 90.35 90.35 90.36
Avg 17.4 16.7 15.9 85.37 85.66 85.90
10 Min 6.7 6.7 6.4 29.60 37.46 41.37
Max 31.9 29.4 29.0 98.60 98.60 98.60
Avg 17.2 16.6 16.0 85.85 86.08 86.27
P 15 Min 6.7 6.7 6.4 33.23 39.34 42.46
Max 29.9 29.3 29.0 98.63 98.63 98.63
Avg 17.1 16.5 16.0 86.15 86.35 86.52
20 Min 6.7 6.7 6.4 35.20 40.51 43.24
Max 29.8 29.3 29.0 98.66 98.66 98.66
Avg 8.9 8.7 8.5 90.72 90.75 90.78
0.03 Min 6.7 6.7 6.4 69.34 69.75 70.00
Max 9.6 9.5 9.4 98.66 98.66 98.66
Avg 15.3 14.9 14.4 87.43 87.54 87.63
Table 4 continued
Factor Factor Measure Decentralized performance gap (ϒ) VMI cost reduction factor ( )
value Nash Ret–Led Sup–Led Nash Ret–Led Sup–Led
HR 0.05 Min 11.2 11.1 10.7 58.26 59.60 60.40

Max 16.8 16.5 16.4 98.21 98.21 98.21
Avg 27.6 26.2 25.0 79.22 79.80 80.28
0.1 Min 19.4 19.1 16.7 29.60 37.46 41.37
Max 31.9 29.4 29.0 97.08 97.08 97.09
Avg 17.6 17.1 16.6 90.07 90.19 90.30
3 Min 7.9 7.8 7.7 63.55 65.92 67.33
Max 30.7 29.4 29.0 98.66 98.66 98.66
Avg 17.3 16.7 16.1 86.75 86.95 87.12
L 5 Min 7.4 7.4 7.2 51.25 55.24 57.47
Max 31.0 29.2 28.8 98.23 98.23 98.23
Avg 17.1 16.2 15.4 80.24 80.64 80.98
10 Min 6.7 6.7 6.4 29.60 37.46 41.37
Max 31.9 28.9 28.3 97.42 97.42 97.43
Avg 16.9 16.2 15.5 83.03 83.33 83.58
50 Min 6.7 6.7 6.4 29.60 37.46 41.37
Max 30.6 29.0 28.6 98.09 98.09 98.09
Avg 17.3 16.6 16.0 86.22 86.44 86.64
SC 75 Min 7.1 7.1 6.9 41.15 47.34 50.60
Max 31.3 29.2 28.9 98.45 98.45 98.45
Avg 17.5 16.9 16.3 88.13 88.31 88.47
100 Min 7.4 7.4 7.2 48.60 53.82 56.68
Max 31.9 29.4 29.0 98.66 98.66 98.66
Avg 18.3 16.9 15.9 85.14 85.68 85.99
0.6 Min 8.5 7.4 6.8 29.60 37.46 41.37
Max 31.9 29.4 28.9 98.62 98.62 98.62
Avg 17.1 16.6 15.9 85.89 86.05 86.25
α 0.75 Min 7.3 7.0 6.5 37.44 39.54 42.07
Max 29.5 29.3 29.0 98.64 98.64 98.64
Avg 16.4 16.3 16.0 86.34 86.36 86.45
0.9 Min 6.7 6.7 6.4 41.35 41.61 42.67
Max 29.2 29.2 29.0 98.66 98.66 98.66
Table 5 Optimal scenario policies
Factor VMI
Factor Centralized Equal power Powerful retailer Powerful supplier RMI
value z Q Cost z Q Cost z Q Cost z Q Cost z Q Cost
0.10 1.785 202.8 15.7 1.279 386.3 19.8 1.512 384.9 19.7 1.285 382.9 19.7 3.226 2.6 629.5
CV 0.30 1.778 205.6 16.9 1.269 392.1 21.1 1.506 387.6 20.9 1.287 381.5 20.7 2.868 8.4 195.0
0.60 1.768 209.9 18.7 1.253 401.2 23.0 1.497 391.7 22.5 1.291 379.6 22.2 2.617 18.0 95.7
10 1.605 206.4 17.0 1.045 394.3 21.2 1.312 388.4 20.9 1.072 380.7 20.7 2.777 10.0 297.9
P 15 1.799 206.1 17.1 1.297 393.0 21.3 1.531 388.0 21.1 1.317 381.5 20.9 2.919 9.6 307.8
20 1.927 205.9 17.2 1.458 392.4 21.4 1.674 387.8 21.2 1.474 381.8 21.0 3.016 9.4 314.6
0.03 1.950 246.2 13.2 1.601 387.4 14.5 1.716 386.3 14.5 1.607 382.9 14.5 3.099 9.1 318.8
HR 0.05 1.795 209.6 16.1 1.327 390.8 19.0 1.524 387.7 18.9 1.340 382.1 18.8 2.930 9.6 308.0
0.10 1.586 162.4 22.0 0.872 401.5 30.3 1.277 390.2 29.8 0.915 378.9 29.3 2.683 10.3 293.5
3 1.780 204.8 16.6 1.271 390.5 20.7 1.508 386.8 20.5 1.286 381.9 20.4 3.002 6.8 401.5
L 5 1.778 205.8 17.0 1.268 392.6 21.2 1.506 387.8 20.9 1.287 381.4 20.7 2.916 9.0 306.2
10 1.773 206.2 17.9 1.255 396.9 22.3 1.501 389.7 21.9 1.284 380.5 21.7 2.794 13.2 212.6
50 1.859 170.6 14.5 1.371 325.5 18.0 1.598 320.6 17.8 1.393 314.3 17.6 2.904 9.7 205.6
SC 75 1.769 208.1 17.2 1.257 397.0 21.5 1.497 391.8 21.2 1.278 385.0 21.0 2.904 9.7 306.8
100 1.703 239.6 19.5 1.171 457.2 24.4 1.421 451.8 24.2 1.191 444.7 23.9 2.904 9.7 407.9
0.60 1.777 206.1 17.1 1.127 401.7 21.7 1.532 390.4 21.1 1.168 379.2 20.8 2.831 9.9 301.8
α 0.75 1.777 206.1 17.1 1.280 391.8 21.2 1.507 388.2 21.1 1.297 381.7 20.8 2.909 9.6 307.1
0.90 1.777 206.1 17.1 1.393 386.1 21.0 1.477 385.6 21.0 1.398 383.0 20.9 2.971 9.5 311.4
Table 6 Impact of channel power (%) on system and agent cost
Factor Factor Measure Ret-Led.cost improve.(δ) Sup-Led. cost improve.()
value δR δS δ R S
Avg 0.23 0.38 0.28 0.85 0.01 0.55

0.1 Min 0.00 0.00 0.00 0.06 0.00 0.02
Max 1.85 2.87 2.13 4.92 0.14 3.60
Avg 0.63 1.13 0.79 2.28 0.08 1.53
CV 0.3 Min 0.01 0.01 0.01 0.16 0.00 0.07
Max 5.04 8.62 5.98 12.59 1.25 9.61
Avg 1.12 2.24 1.47 3.92 0.31 2.76
0.6 Min 0.01 0.03 0.02 0.28 0.00 0.13
Max 9.13 17.29 11.17 20.70 4.79 16.72
Avg 0.77 1.41 0.97 2.70 0.17 1.86
10 Min 0.00 0.01 0.00 0.07 0.00 0.03
Max 9.13 17.29 11.17 20.70 4.79 16.72
Avg 0.64 1.22 0.82 2.28 0.12 1.57
P 15 Min 0.00 0.00 0.00 0.06 0.00 0.03
Table 6 continued
Factor Factor Measure Ret-Led.cost improve.(δ) Sup-Led. cost improve.()
value δR δS δ R S
Max 7.38 14.73 9.15 17.17 3.32 13.83

Avg 0.58 1.12 0.75 2.07 0.10 1.42
20 Min 0.00 0.00 0.00 0.06 0.00 0.02
Max 6.55 13.49 8.19 15.39 2.71 12.40
Avg 0.13 0.30 0.22 0.90 0.01 0.44
0.03 Min 0.00 0.00 0.00 0.06 0.00 0.02
Max 0.72 1.98 1.33 3.96 0.20 2.14
Avg 0.37 0.78 0.53 1.73 0.05 1.05
HR 0.05 Min 0.01 0.01 0.01 0.10 0.00 0.06
Max 2.10 5.09 3.22 7.77 0.72 5.13
Avg 1.48 2.67 1.79 4.42 0.33 3.35
0.1 Min 0.03 0.05 0.04 0.25 0.00 0.18
Max 9.13 17.29 11.17 20.70 4.79 16.72
Avg 0.51 0.91 0.64 1.83 0.07 1.24
3 Min 0.00 0.00 0.00 0.06 0.00 0.02
Max 5.46 9.45 6.50 13.51 1.49 10.37
Avg 0.63 1.18 0.81 2.26 0.11 1.54
L 5 Min 0.00 0.01 0.00 0.07 0.00 0.03
Max 6.80 12.21 8.19 16.33 2.45 12.77
Avg 0.87 1.71 1.13 3.02 0.22 2.12
10 Min 0.00 0.01 0.01 0.10 0.00 0.04
Max 9.13 17.29 11.17 20.70 4.79 16.72
Avg 0.71 1.40 0.93 2.52 0.16 1.75
50 Min 0.00 0.01 0.00 0.07 0.00 0.03
Max 9.13 17.29 11.17 20.70 4.79 16.72
Avg 0.65 1.23 0.84 2.33 0.13 1.60
SC 75 Min 0.00 0.01 0.00 0.06 0.00 0.03
Max 8.75 15.60 10.52 20.22 4.11 16.06
Avg 0.62 1.12 0.78 2.20 0.11 1.50
100 Min 0.00 0.00 0.00 0.06 0.00 0.02
Max 8.55 14.65 10.16 19.99 3.79 15.72
Avg 1.49 2.71 1.88 4.38 0.32 3.03
0.6 Min 0.07 0.12 0.10 0.37 0.00 0.16
Max 9.13 17.29 11.17 20.70 4.79 16.72
Avg 0.44 0.91 0.58 2.03 0.07 1.38
α 0.75 Min 0.02 0.04 0.03 0.18 0.00 0.08
Max 2.55 5.95 3.36 9.39 1.01 7.40
Avg 0.06 0.13 0.08 0.64 0.01 0.43
0.9 Min 0.00 0.00 0.00 0.06 0.00 0.02
Max 0.32 0.89 0.45 2.91 0.10 2.26
Appendix B
Proof of Lemma 1 The retailer places an order to the supplier once every Q units of
customer demand. Thus, the time between orders at the supplier is a random variable,
T defined as Q/D, where D represents customer demand rate with mean Q/µ.4 Then,
applying Theorem 2.1.2 in Casella and Berger (1990) for D ∈ (0, ∞), the probability
distribution function of T is

Q −µ
1 −
t Q
f T (t) = √ e 2σ 2 . (B.1)
σ 2π t2
Using (B.1) and recalling that the supplier’s production

rate is µ, the expected
shortfall per cycle at the supplier, E (Q − T µ)+ , becomes
Q Q
µ
µ
Q
E (Q − T µ)+ = (Q − t) f T (t) dt = Q FT − t f T (t) dt. (B.2)
µ
0 0
Using the identity

Q Q Q
FT (t) = P(T ≤ t) = P ≤t =1− P D ≤ = 1 − FD ,
D t t
Eq. (B.2) becomes
Q

µ Q −µ
Q 1 − t
E (Q − T µ)+ = Q [1 − FD (µ)] − √ e 2σ 2 dt.
σ 2π t
0
Q/t−µ
Substituting ξ = σ and noting that dt = − (ξ σQσ
+µ)2
dξ , gives
⎛ ⎞
∞
1 µ
E (Q − T µ)+ = Q ⎝ − φ(ξ ) dξ ⎠ . (B.3)
2 σξ + µ
0
∞
µ µ
So, b Q E (Q − T µ)+ = bµ 21 − 0 σ ξ +µ φ(ξ ) dξ = bµK.

Proof of Proposition 6 The existence of the Stackelberg equilibrium when the retailer
is the leader follows from the continuity of the retailer’s cost function. Uniqueness of
4 Given that T is not continuous at D = 0, we require, for tractability, that D > 0 For reasonable values
of the coefficient of variation, the probability of a negative demand is sufficiently small to satisfy this
assumption.
the equilibrium is satisfied if we can show that the retailer’s cost function is quasi-
convex. The retailer wishes to minimize
√
αpσ µLh S (z)
R
K VMI (z) = √ √
2 S + (1 − α) pσ L(z)
√ √ √
hR µ
+√ S + (1 − α) pσ L(z) + h R zσ L.
2h S
R . For ease of
We proceed next to get the first and second order conditions for K VMI
exposition, define
√
(z)
A= √ and B = S + (1 − α) pσ L(z),
S + (1 − α) pσ L(z)
so that
√ √
αpσ µLh S hR µ √
R
K VMI (z) =A √ +B√ + h R zσ L.
2 2h S
Then,
√ √
−[1 − (z)] S + (1 − α) pσ L(z) + (1−α) √ pσ L(z)[1−(z)]
√
∂A 2 S+(1−α) pσ L(z)
= √
∂z S + (1 − α) pσ L(z)
√
[1 − (z)][2S + (1 − α) pσ L(z)]
= − √ and
2[S + (1 − α) pσ L(z)]3/2
√ √ √
∂2A 2 φ(z)[2S+(1−α) pσ L(z)]+(1−α) pσ L[1−(z)]2 [S +(1−α) pσ L(z)]
= √
∂z 2 4[S+(1−α) pσ L(z)]5/2
√ √
3(1 − α) pσ L[1 − (z)]2 [2S + (1 − α) pσ L(z)]
− √
4[S + (1 − α) pσ L(z)]5/2
√ √
(1 − α) pσ L[S + (1 − α) pσ L(z)] 2φ(z)(z) − [1 − (z)]2
= √
4[S + (1 − α) pσ L(z)]5/2
√ √
4φ(z)S[S + (1 − α) pσ L(z)] − 3(1 − α) pσ L S[1 − (z)]2
+ √ .
4[S + (1 − α) pσ L(z)]5/2
Furthermore,
√
∂B (1 − α) pσ L[1 − (z)]
=− √ and
∂z
2 S + (1 − α) pσ L(z)
2
√ √
√
(1−α) pσ L[1−(z)]
2(1 − α) pσ Lφ(z) S + (1 − α) pσ L(z) − √ √
∂ 2B S+(1−α) pσ L(z)
= √
∂z 2 4[S + (1 − α) pσ L(z)]
√ 2 √
(1 − α) pσ L 2φ(z)(z) − [1 − (z)]2 + 2(1 − α) pσ Lφ(z)S
= √
4[S + (1 − α) pσ L(z)]3/2
≥ 0.
∂ 2 K VMI
R (z)
After recombining terms and algebraic manipulation, the numerator of ∂z 2
becomes
√
αpσ h S Lµ S + (1 − α) pσ L(z)
√
× (1 − α) pσ L 2φ(z)(z) − [1 − (z)]2 + 4φ(z)S
√
−3α(1 − α) p 2 σ 2 L µSh S [1 − (z)]2

√ √ √ 2
+h R µ[S + (1 − α) pσ L(z)] (1 − α) pσ L
√
× 2φ(z)(z) − [1 − (z)]2 + 2(1 − α) pσ Lφ(z)S
√ √
= S + (1 − α) pσ L(z) 2φ(z)(z) − [1 − (z)]2 (1 − α) pσ L
√ √
× pασ h S Lµ + h R µ(1 − α) pσ L
√
−3α(1 − α) p 2 σ 2 L µSh S [1 − (z)]2
√
+ S + (1−α) pσ L(z) 4φ(z)Spασ h S Lµ+2h R Lµ(1−α) pσ φ(z)S
√ √
= (1 − α) p 2 σ 2 L µ 2φ(z)(z) − [1 − (z)]2 S + (1 − α) pσ L(z)
× [αh S + (1 − α)h R ]
√
+2 pσ LµSφ(z) S + (1 − α) pσ L(z) [2αh S + (1 − α)h R ]
√
−3α(1 − α) p 2 σ 2 L µSh S [1 − (z)]2 , (B.4)
∂ 2 K VMI
R (z) √ √
and the denominator of ∂z 2
is 4 2h S [S + (1 − α) pσ L(z)]5/2 ≥ 0. Thus,
∂ 2 K VMI
R (z)
in search of a condition for ∂z 2
≥ 0, we focus on (B.4) and impose that it be
positive. A sufficient condition is that the sum of its last two terms be positive, which
translates into
√
(1 − α) pσ L 3αh S [1 − (z)]2 − 2φ(z)(z)[2αh S + (1 − α)h R ]
S≥ . (B.5)
2φ(z)[2αh S + (1 − α)h R ]
However, for mathematical tractability, we do not use (B.5) directly but rather an
upper bound for (B.5) obtained from the inequality 3[1 − (z)]2 − 2φ(z)(z) ≤
2[1 − (z)]2 . Thus, we have
√
(1 − α) pσ L[1 − (z)]2 [αh S − (1 − α)h R ]
S≥ , (B.6)
2φ(z)[2αh S + (1 − α)h R ]
as a sufficient condition.
α hR
It is straightforward to note that if (1−α) ≤ hS , then αh S − (1 − α)h R ≤ 0. In this
α hR
case, condition (B.6) is satisfied ∀S ≥ 0. If, however, (1−α) ≥ h S , it can be shown that
[1−(z)]2 αh S −(1−α)h R
terms φ(z) and 2αh S +(1−α)h R
are bounded above by 0.63 and 0.5, respectively.
Thus, (B.6) is equivalent to
√ 0.63 √
S ≥ (1 − α) pσ L 0.5 = 0.1575(1 − α) pσ L . (B.7)
2

Proof of Proposition 7 The supplier’s cost function is continuous and thus there exists
a Stackelberg equilibrium for the case of the powerful supplier. The supplier’s cost
function is
µ √ ∗ µ Q
S
K VMI (Q) = (1 − α) pσ L(z VMI (Q)) + S + h S + bµK.
Q Q 2
We focus next on showing that the equilibrium is unique. Using the IFT,
∗ (Q))
∂φ(z VMI ∗ (Q) z ∗ (Q))[1 − (z VMI
∗ (Q))]
∂φ(z) ∂z VMI
= = VMI
∂Q ∂z ∂Q Q
z ∗ (Q))h R
= VMI , and
µαp
∗ (Q))
∂(z VMI ∗ (Q) ∗ (Q))]2
∂(z) ∂z VMI [1 − (z VMI
= = ∗
∂Q ∂z ∂Q Qz VMI (Q)
h 2R Q
= ∗ (Q)) .
(µαp)2 φ(z VMI
Then,
√ h2 Q2 √ ∗
∂ K VMI
S (Q) µ(1 − α) pσ R
L (µαp)2 φ(z ∗ − µ(1 − α) pσ L(z VMI (Q))
VMI (Q))
=
∂Q Q2
µS hS
− +
Q2 2
√ 2 2 √ ∗ ∗
(1 − α)σ Lh R Q − (µαp)2 (1 − α)σ L(z VMI (Q))φ(z VMI (Q))
= ∗
Q 2 α 2 µpφ(z VMI (Q))
µS hS
− + ,
Q2 2
√ √
(1 − α)σ Lh 2R Q − µα(1 − α) pσ ∗
L(z VMI ∗
(Q))z VMI ∗
(Q)h R µp Q 2 α 2 φ(z VMI (Q))
S (Q)
∂ 2 K VMI
= ∗
∂ Q2 [µpφ(z VMI (Q))]2 (Qα)4
√ √
− (1 − α)σ Lh 2R Q 2 − (µαp)2 (1 − α)σ L

∗ ∗
×(z VMI (Q))φ(z VMI (Q))
∗ ∗

2Qµα 2 pφ(z VMI (Q)) + α Q 2 (z VMI (Q))h R
× ∗
[µpφ(z VMI (Q))]2 (Qα)4
∗
2Sp 2 φ(z VMI (Q))2 α 3 µ3
+ ∗ .
[µpφ(z VMI (Q))]2 (Qα)4
∂ 2 K VMI
S (Q) ∂ 2 K VMI
S (Q)
We want ∂ Q2
≥ 0. Note that ∂ Q2
≥ 0 if its numerator is positive. Thus,
∂ 2 K VMI
S (Q)
we now focus on C(Q), where ∂ Q2
= C∂(Q)
Q2
. After algebraic manipulation,
√ ∗ ∗
C(Q) = α(1 − α)σ L 2(µαp)3 (z VMI (Q))φ(z VMI (Q))2

∗ ∗ ∗
−h 2R Q 2 µαpφ(z VMI (Q)) − h 3R Q 3 z VMI (Q) + 2Sp2 φ(z VMI (Q))2 α 3 µ3 .
∗ (Q), Q ∈ (0, µαp ). Now,

Due to the existence conditions for z VMI 2h R
∗ ∗
lim φ(z VMI (Q)) = lim (z VMI (Q)) = 0, so lim C(Q) = 0. Furthermore,
Q→0 Q→0 Q→0
∗ ∗ ∗
limµαp z VMI (Q) = 0, limµαp φ(z VMI (Q)) = lim (z VMI
µαp
(Q)) = δ > 0. Thus
Q→ 2h Q→ 2h Q→ 2h
R R R

√ δ
limµαp C(Q) = α(1 − α)σ L(µαp)3 2 × δ 3 − + 2φ(0)2 Sµ3 α 3 p 2 > 0.
Q→ 2h 4
R
Note that δ ≈ 0.39894 > 0. Thus, the extreme points of C(Q) are positive. Then,

dC(Q) ∗
√ ∗ ∗
= 4h R z VMI (Q)α(1 − α)σ L 4(µαp)2 (z VMI (Q))φ(z VMI (Q)) − h 2R Q 2
dQ
h 3R Q 3 ∗ (Q))
Spα 2 µ2 φ(z VMI
+ ∗ (Q))z ∗ (Q) + √ .
4µαpφ(z VMI VMI α(1 − α)σ L
Let,
∗ ∗
D(Q) = 4(µαp)2 (z VMI (Q))φ(z VMI (Q)) − h 2R Q 2 .
Then,

1
lim D(Q) = 0 and lim D(Q) = (µαp)
µαp
2
4×δ −
2
> 0.
Q→0 Q→ 2h 4
R
Furthermore,
dD(Q) ∗ ∗
= 2h 2R Q + 4h R µαpz VMI (Q)(z VMI (Q)) ≥ 0.
dQ
d C(Q)
Thus, D(Q) ≥ 0 ∀Q ∈ (0, µαp
2h R ), which implies d Q ≥ 0. Hence, all critical points
∂ 2 K VMI
S (Q)
of C(Q) are ≥ 0; therefore, C(Q) ≥ 0 and ∂ Q2
≥ 0. It follows that K VMI
S (Q) is
convex and a unique equilibrium exists.

References
Amir R, Stepanova A (2006) Second-mover advantage and price leadership in Bertrand duopoly. Games
Econ Behav 55:1–20
Aviv Y, Federgruen A (1998) The operational benefits of information sharing and vendor managed inventory
(VMI) programs. Working paper, Olin School of Business, Washington University, St. Louis, MO
Benton W, Maloni M (2004) The influence of power driven buyer/seller relationships on supply chain
satisfaction. J Oper Manage (23):1–22
Bernstein F, Federgruen A (2003) Pricing and replenishment strategies in a distribution system with com-
peting retailers. Oper Res 51(3):409–426
Bernstein F, Federgruen A (2004) Dynamic inventory and pricing models for competing retailers. Naval
Res Logis 51(2):258–274
Bernstein F, Federgruen A (2005) Decentralized supply chains with competing retailers under demand
uncertainty. Manage Sci 51(1):18–29
Bichescu B, Fry M (2007) A numerical analysis of supply chain performance under split decision rights.
Omega. doi:10.1016/j.omega.2007.04.001
Bolen WH (1988) Contemporary retailing, 3rd edn. Prentice Hall, Englewood Cliffs
Boyer M, Moreaux M (1987) On stackelberg equilibria with differentiated products: the critical role of the
strategy space. J Ind Econ 36(2):217–230
Buzzel R, Ortmeyer G (1995) Channel partnerships streamline distribution. Sloan Manage Rev (36):85
Cachon GP (2001) Stock wars: inventory competition in a two-echelon supply chain with multiple retailers.
Oper Res 49(5):658–674
Cachon GP (2003) Supply chain coordination with contracts. In: Graves S, de Kok A (eds) Handbooks in
operations research and management science, vol 11: Supply chain management: design, coordination
and operation. Elsevier, Amsterdam pp 229–339
Cachon GP, Fisher M (1997) Campbell Soup‘s continuous replenishment program: evaluation and enhanced
inventory decision rules. Prod Oper Manage 6(3):266–276
Cachon GP, Netessine S (2004) Game theory in supply chain analysis. In: Simichi-Levi D, Wu SD,
Shen Z-J (eds) Supply chain analysis in the eBusiness Era. Kluwer, Boston
Cachon GP, Zipkin P (1999) Competitive and cooperative inventory policies in a two stage supply chain.
Manage Sci 45(7):936–953
Casella G, Berger RL (1990) Statistical Inference, 1st edn. Brooks/Cole, Belmont
Choi C (1991) Price competition in a channel structure with a common retailer. Market Sci 10(4):271–296
Clark TH, Hammond J (1997) Reengineering channel reordering processes to improve total supply chain
performance. Prod Oper Manage 6(3):248–265
Clark TH, McKenney JL (1994) Campbell Soup Company: a leader in continuous replenishment inno-
vations. Harvard Business School Case # 195124. Harvard Business School, Harvard University,
Cambridge
Corbett CJ (2001) Stochastic inventory systems in a supply chain with asymmetric information: cycle
stocks, safety stocks, and consignment stock. Oper Res 49(4):487–500
Corbett CJ, Zhou D, Tang CS (2004) Designing supply contracts: contract type and information asymmetry.
Manage Sci 50(4):550–559
Desiraju R, Moorthy S (1997) Managing a distribution channel under asymmetric information with perfor-
mance requirements. Manage Sci 43(12):1628–1644
Emerson RM (1962) Power-dependence relations. Ame Soc Rev (27):31–41
Federgruen A (1993) Centralized planning models for multi-echelon inventory systems under uncertainty.
In: Graves SC, RinnooyKan AHG, Zipkin PH (eds) Handbooks in operations research and management
science, vol 4 (Logistics of production and inventory). Elsevier, Amsterdam pp 133–173
Fry MJ, Kapuscinski R, Olsen TL (2001) Coordinating production and delivery under a (z, Z )-type Vendor-
Managed Inventory contract. Manufact & Serv Oper Manage 3(2):151–173
Gal-Or E (1985) First mover and second mover advantages. Int Econ Rev 26:649–653
Gavirneni S, Kapuscinski R, Tayur S (1999) Value of information in capacitated supply chains. Manage
Sci 45(1):16–24
Hadley G, Whitin TM (1963) Analysis of inventory systems. Prentice-Hall, Englewood Cliffs
Hammond J (1994) Barilla SpA (A) and (B). Harvard Business School Case # 694046. Harvard Business
School, Harvard University, Cambridge
Harris F (1915) Operations and Costs. (Factory Management Series). A.W. Shaw, Chicago, pp 48–52
Hill R, Omar M (2006) Another look at the single-vendor single-buyer integrated production-inventory
problem. Int J Prod Res 44(4):791–800
Hopp W, Spearman M, Zhang R (1997) Easily implementable inventory control policies. Oper Res
45(3):327–340
Kadiyali V, Chintagunta P, Vilcassim V (2000) Manufacturer-retailer channel interactions and implications
for channel power: an empirical investigation of pricing in a local market. Market Sci 19(2):127–148
Lee HL, So KC, Tang CS (2000) The value of information sharing in a two-level supply chain. Manage Sci
46(5):626–643
Maloni M, Benton W (2000) Power influences in the supply chain. J Bus Logis 21(1):49–73
Nagarajan M, Rajagopalan S (2004) Coordination issues in vendor managed inventory systems. Working
paper. University of Southern California, Marshall School of Business, Los Angeles
Narayanan VG, Raman A (1998) Assignment of stocking decision rights under incomplete contracting.
Working paper, Harvard Business School, Harvard University, Cambridge
Netessine S, Rudi N (2003) Centralized and competitive inventory models with demand substitution. Oper
Res 51(2):329–335
Netessine S, Rudi N (2004) Supply chain structures on the internet and the role of marketing-operations
interaction. In: Simchi-Levi D, Wu S, Shen Z (eds) Supply chain analysis in the eBusiness Era chap.
19. Kluwer, Amsterdam
Papadimitriou CH (2001) Game theory and mathematical economics: a theoretical computer scientists intro-
duction. In: Proceedings of the 42nd IEEE symposium on foundations of computer science (FOCS)
pp 4–8, Las Vegas, NV, October 2001. IEEE Computer Society Press, Los Alamitos
Petruzzi NC, Dada M (1999) Pricing and the newsvendor problem: a review with extensions. Oper Res
47(2):183–194
The Wall Street Journal (2004) In bow to retailers’ new clout, Levi Strauss makes alterations. Wall Street J
pp A1. http://online.wsj.com/article/SB108742485606139344.html
Valentini G, Zavanella L (2003) The consignment stock of inventories: industrial case and performance
analysis. Int J Prod Econ 81-82:215–224
Wang Y (2006) Joint pricing-production decisions in supply chains of complementary products with uncer-
tain demand. Oper Res 54(6):11101127
Wang Y, Jiang L, Shen Z (2004) Channel performance under consignment contract with revenue sharing.
Manage Sci 50(1):34–47
Wu X, YK R, Dang C (2005) Coordinating a supply chain with a vendor-managed inventory contract.
Working paper. City University of Hong Kong, Hong Kong
Zheng Y-S (1992) On properties of stochastic inventory systems. Manage Sci 38(1):87–103
Zipkin P (2000) Foundations of inventory management. McGraw-Hill, New York
Value chain management for commodities: a case study
from the chemical industry
M. Kannegiesser · H.-O. Günther · P. van Beek ·

M. Grunow · C. Habla

OR Spectrum (2009) 31:63–93
DOI 10.1007/s00291-008-0124-9
Abstract We present a planning model for chemical commodities related to an

industry case. Commodities are standard chemicals characterized by sales and supply
volatility in volume and value. Increasing and volatile prices of crude oil-dependent
raw materials require coordination of sales and supply decisions by volume and value
throughout the value chain to ensure profitability. Contract and spot demand differen-
tiation with volatile and uncertain spot prices, spot sales quantity flexibility, spot sales
price–quantity functions and variable raw material consumption rates in production
are problem specifics to be considered. Existing chemical industry planning models
are limited to production and distribution decisions to minimize costs or makespan.
Demand-oriented models focus on uncertainty in demand quantities not in prices. We
develop an integrated model to optimize profit by coordinating sales quantity, price
M. Kannegiesser · H.-O. Günther (B)

Department of Production Management, Technical University of Berlin,
Wilmersdorfer Str. 148, 10585 Berlin, Germany
P. van Beek
Management Studies Group and Operations Research and Logistics Group,
Wageningen University, Hollandseweg 1, 6706 KN Wageningen, The Netherlands
M. Grunow
Department of Manufacturing Engineering and Management, Technical University of Denmark,
Building 425, 2800 Kgs. Lyngby, Denmark
C. Habla
Department of Enterprise-Wide Software Systems, The Fern Universität in Hagen,
Universitätsstr. 1, 58097 Hagen, Germany

284 M. Kannegiesser et al.
and supply decisions throughout the value chain. A two-phase optimization approach
supports robust planning ensuring minimum profitability even in case of worst-case
spot sales price scenarios. Model evaluations with industry case data demonstrate the
impact of elasticities, variable raw material consumption rates and price uncertainties
on planned profit and volumes.
Keywords Value chain management · Sales and supply network planning · Demand
uncertainty · Commodities · Chemical industry
1 Introduction
The chemical industry is one of the key global industries with product sales of e
1,776 billion globally in 2004 (CEFIC 2005). In this article, we focus on the segment
of chemical commodities. Commodities are mass products produced and sold in high
volumes with standardized quality and few variants. Price is the key buying criterion
for customers. Examples are standard polymers, certain types of intermediate products
or basic chemicals. Sales prices for theses commodities are volatile and can change
regularly, e.g., weekly or monthly based on negotiations between the company and its
customers.
Prices for raw materials can also change regularly. Specifically, many key raw
materials in the chemical industry showed a severe rise in prices due to the increase
of the crude oil price over the last years. Raw price volatility and increases have to be
considered in sales and supply planning of commodity products to ensure profitability
of the business. Therefore, the focus on demand and supply volume planning alone is
not sufficient since a feasible volume plan might not be profitable for the company due
to the volatility of supply costs and sales prices. The monthly planning process needs
to support integrated decisions on volume and values, specifically on sales quantities
and prices considering available supply volumes and raw material costs. In this paper,
an integrated planning model related to a real-life case from the European chemical
industry is presented.
In our investigation, we consider a simplified intra-organizational value chain net-
work of a company producing chemical commodities. The industry context of this
case is a company operating a complex, multi-stage value chain network producing
polymers that also require several intermediate products as raw material. The company
is operating at several production sites and is serving different sales locations. The
business is a commodity business where raw materials and finished products are char-
acterized by market price and volume volatility. Annual production volumes exceed 1
Mio. tons. In this study we focus on the monthly sales and operations planning process
for the entire value chain network for a planning horizon of 6–12 months.
Figure 1 shows a section of the network. The company has grouped multiple cus-
tomers in regional or industry-specific sales locations. Two production resources are
located in one production location, from where sales locations are served. One market-
facing multi-purpose resource produces multiple finished commodity products. The
second single-purpose resource produces the intermediate product for the multi-
purpose resource in continuous production mode. The intermediate product produced
Value chain management for commodities 285
Procurement Production Sales
Production location 1 Sales location 1

continuous multi-purpose
Procurement Sales location 2
R1 R2
location 1
...
raw material intermediate finished
Sales location 9
Legend:
= procurement & sales location
l∈L
= production location
= resource r∈R
= product p∈P
= material flow
periods: monthly planning bucket t ∈T
Fig. 1 Section of the considered value chain network
on resource R1 requires a raw material product procured from an external procurement

location.
The planning problem at hand shows a number of characteristics that are typical of
the chemical industry.
• Spot and contract business differentiation is an important issue in the chemical
industry specifically in commodity business.
• Price and volume volatility for chemical commodities in sales and procurement is
more significant than in other industries, e.g., in discrete parts manufacturing.
• The entire production system is organized as a multi-stage network with multi-
purpose and continuously operated production resources.
• Material flows are predominantly divergent with intermediates used in multiple
subsequent products.
• Raw material consumption rates in production are variable depending on the degree
of capacity utilization.
These characteristics can be found, for example, in basic chemicals and/or polymer
production, while fine chemical and pharmaceutical production can be seen as a spe-
cialty type of business relying on smaller quantities and complex batch production
mode. The simplified network as shown in Fig. 1 focuses on the interaction between
procurement, production and sales. The problem at hand is an excerpt from the global
value chain planning problem of a polymer producing company. In our investigation
we focus on the interaction between key business functions in the global value chain
context. The model developed represents a prototype which is used by the planners
to better understand volume and value dynamics from sales to procurement and their
impact on profit in a value chain network. In a later stage, the company intends to
introduce the Supply Network Planning module of an advanced planning software
system for operative planning (cf. Dickersbach 2006). To reduce the complexity of
the prototype model, several standard features such as inventory records and transpor-
tation are excluded mainly because they do not have a high profit impact compared
to sales and procurement issues. Exchange rates and risk hedging inventories are also
excluded here though they do represent further important issues in the investigated
global value chain network which will be included in the final implementation of the
value chain planning model.
Traditionally, supply network planning models focus on the flow of goods in the
network while assuming sales and procurement prices as being fixed. Revenue man-
agement, however, represents a topic which has recently gained considerable interest
both in practice and in academia. For an application in the iron and steel industry
and a discussion of dynamic pricing in the US automotive industry, cf. Spengler et al.
(2007) and Biller et al. (2005), respectively. Key issues of revenue management are
dynamic pricing strategies as well as accept and reject decisions to make more effec-
tive use of resources. Booking and pricing systems of airlines, hotels, car rentals,
telecommunication systems and cargo transportation are just a few popular examples
of revenue management, cf. Gosavi et al. (2007), Bartodziej et al. (2007), Lee et al.
(2007), Defregger and Kuhn (2007), Reiner and Natter (2007). These papers focus
on revenue maximization based on pricing and decisions to influence the demand for
services such as airline seats, rental car capacity or hotel rooms which are in limited
supply. Active sales and pricing decisions investigated in revenue management are
principally relevant for the industrial planning problem considered in our paper. How-
ever, in contrast to service industries we deal with physical products and the complex
decision-making process in a global chemical value chain.
This paper aims at integrating ideas of revenue management into supply network
planning to optimize profit throughout the entire intra-organizational value chain net-
work. We choose “value chain management” as an overall term for the integration of
demand-oriented management concepts such as revenue management as well as sup-
ply-oriented logistics management concepts which primarily focus on material flows.
Specifically, our modeling approach reflects the following key issues:
• For chemical commodities as well as for many other industrial products (e.g., fer-
tilizers or animal feed products), contract and spot demand can be distinguished.
While sales prices and quantities are fixed for contract demand, spot market sales
can be highly variable with regard to both price and quantity. We develop a value
chain planning model that, in addition to production and distribution planning, also
supports pricing and sales decisions for spot demand.
• Similar to sales commodity markets, raw materials can be procured either based on
fixed contracts with suppliers or on the spot market. In the latter case, the company
has to decide on the procurement quantity taking the volatility of procurement
prices into account. Our modeling approach also reflects these issues which are of
increasing importance in many industries specifically confronted with increasing
raw material prices.
• Empirical investigations have shown that both spot sales prices for commodities
as well as procurement prices for raw materials are characterized by high uncer-
tainty. Modeling these prices as independent random variables, as it is assumed in a
large number of academic contributions, is not always realistic because the drivers
behind the market development, e.g., development of crude oil prices, are ignored.
Hence, our approach is based on scenario analysis which utilizes human exper-
tise to forecast market developments in combination with subjective probability
measures.
• Finally, it is shown how technological flexibility with respect to consumption rates
of raw material and feasible processing modes of the chemical production equip-
ment can be used in order to balance sales market demands and procurement
opportunities.
The overall objective of the proposed optimization model is to maximize profit by coor-
dinating sales turnover with quantity and prices as well as supply decisions throughout
the value chain. Model evaluations with industry case data demonstrate clearly the
applicability of the value chain optimization model. The model has been developed
and implemented together with the company proving the industry case and also the
problem requirements and assumptions such as contract and spot demand.
The remainder of this article is organized as follows. The next section provides
an overview of the relevant literature. In Sect. 3, a mixed-integer linear optimization
model for sales and supply planning in intra-organizational value chain networks is
developed. Section 4 presents a case study evaluation based on a real application from
the European chemical industry.
2 Literature review
In the academic literature a wealth of papers dealing with demand and supply network
management has been published. For an overview and classification, see, e.g., Thomas
and Griffin (1996), Stadtler (2005) and Tang (2006). Some of these papers focus on
demand, others on supply aspects of the problem. Among the demand-focused papers
emphasis is given either on demand forecasting, demand uncertainty, or pricing deci-
sions.
The objective of demand forecasting is to predict future demand quantities as accu-
rate as possible based on historical data. For an overview of demand forecasting within
supply chain management see Kilger and Wagner (2008) and Meyr (2008). The clas-
sical approach towards demand forecasting does not apply to the considered chemical
commodity business, where contract demand is certain and spot demand does not
need to be fulfilled. In addition, the development of demand does not follow historical
demand patterns, but is rather influenced by future raw material prices as investigated
by Asche et al. (2003) for crude-oil related products.
The paper by Gupta and Maranas (2003) represents one example for dealing with
demand uncertainty in the chemical industry. The authors propose a demand and sup-
ply network planning model to minimize costs. Production decisions are made “here
and now” and demand uncertainty is balanced with inventories independently incorpo-
rating penalties for safety stock and demand violations. Demand quantity uncertainty
is modeled as a normally distributed continuous random variable with known mean and
standard deviation and penalty costs are charged for unfilled demand. This approach,
however, is not suitable in our commodity case, since spot demand and factors such as
demand price uncertainty for chemical commodities or fluctuating raw material and
crude oil prices have to be considered. Another example from the chemical engineer-
ing literature has been given by Chen and Lee (2004). They develop a multi-company
demand and supply network planning model to maximize profit under demand uncer-
tainty and pricing decisions. Demand uncertainty is modeled with quantity scenarios
and probabilities. A two-phase optimization strategy is developed to reach robust
plans. Pricing decisions are modeled with fuzzy logic considering satisfaction levels
of buyer and seller assuming collaboration and preference transparency between both
parties. This assumption, however, is not valid in the spot sales commodity business
considered in our investigation.
Chakravarty (2005) develops an optimization model for global network design
decisions incorporating sales quantity and price decisions. Chakravarty uses demand
curves, where demand quantity is a function of price, and sales turnover is decided
using quadratic optimization. The model scope of profit optimization incorporating
variable sales prices and supply quantities as well as costs is similar to the considered
problem more on a macro network design level rather than on a monthly planning
level for a chemical industry value chain. In addition the assumption of a monopo-
listic market constellation, where the company is able to influence demand by price
setting reflected in the demand curves is not valid in the considered case.
In contrast to demand planning, the supply side of chemical industry value chains
has been widely investigated especially with focus on production planning and sched-
uling. Examples of papers dealing with industrial applications are Blömer and Günther
(2000), Neumann et al. (2002), Kallrath (2002a,b) or on multi-site supply network
planning with given demand, cf. Timpe and Kallrath (2000), Grunow (2001),
Grunow et al. (2003), and Berning et al. (2002). Production scheduling for batch
and campaign production and synchronization of production plans across plants con-
sidering sequence and production mode constraints are major subjects in this field of
research. The specific aspect of variable raw material consumption, which is essential
in the industrial application considered in our investigation, has not sufficiently been
addressed in the literature so far.
Procurement planning in general and spot and contract procurement planning in the
chemical industry particularly have recently been investigated in a number of papers.
For instance, Stadtler (2008) discusses general tasks of purchase planning integrated
in overall supply chain management at the order level. Recent papers discuss pro-
curement strategies for spot and contract markets. Reiner and Jammernegg (2005)
develop a risk-hedging model and compare different procurement strategies including
speculation inventories. Marquez and Blanchar (2004) present extended procurement
strategies based on real-options to optimize contract portfolios considering in-tran-
sit and warehouse inventories. Seifert et al. (2004) underline the importance of spot
procurement next to contract procurement and show the advantage, if a fraction of
demand is based on spot market procurement.
So far, models presented in the academic literature focus either on demand or on
supply aspects. In the academic literature we did not find any realistic value chain
planning model that integrates sales and supply decisions by volume and value in
a price-volatile chemical commodity business, although this planning problem is of
high importance not only in the chemical commodity industry.
3 Sales and supply planning model
To support decision making in the considered intra-organizational value chain network

a mixed-integer linear programming (MILP) model is proposed. Maximizing profit
throughout the entire value chain is seen as the overall objective function. Principally,
the value chain profit consists of the following constituents:
Profit = spot sales turnover depending on variable sales prices and quantities
+ fixed contract sales turnover
− spot procurement costs depending on variable procurement prices and
quantities
− fixed procurement costs
− variable production costs depending on variable raw material consumption
and processing mode
The first two elements of the profit function and related constraints are reflected by
the sales model introduced in Sect. 3.1. The supply model presented in Sect. 3.2 con-
siders all other issues related to procurement and production. To solve the model, two
different optimization strategies are proposed (see Sect. 3.3).
3.1 Sales model
3.1.1 Demand and sales planning for chemical commodities
In the industrial application considered, the central task is to plan monthly sales vol-
umes and values in the network for 6–12 months. The planning process starts with a
monthly demand forecast of quantities and prices submitted by the “Sales and Market-
ing” department of the company. The forecast aggregates demand of single customers
at the sales location level resulting in a cumulated demand quantity and a weighted
average price. The planning objective is to maximize profit considering available pro-
duction and procurement capabilities, sales prices and supply costs. The planning
result is a tactical sales and operations plan with sales quantities and prices as well as
production and procurement quantities per month. The planning problem shows some
specifics as described in the following.
Contract and spot sales quantity management

Contract and spot demand can be distinguished in chemical commodities markets.
Contract demand is based on agreements between the company and customers with
sales quantities and prices being fixed for a defined period. Contract demand quanti-
ties and prices are fulfilled as forecasted and are deterministic. Spot demand is also
forecasted by quantity and price. However, spot demand does not need to be fulfilled
completely since the company can make active sales decisions on the acceptance or
rejection of spot sales requests. The spot price can be bilaterally negotiated, requested
by the customer directly or set by the company. In the latter case the customer reacts
with a quantity bid. In any case prices are negotiated bilaterally between company and
customer. Double auction mechanisms with multiple buyers and sellers submitting
offers and bids cleared in one market price are not considered in this context.
Demand Sales
Available total supply
cut
spot spot push
demand quantity sales quantity
spot ↑ - ↓
spot sales price
demand price
contract contract
demand price sales price - - -
contract contract
demand quantity sales quantity
1 2 3 Periods 1 2 3 Periods
Price effects: - demand price = sales price ↓ demand price ≥ sales price ↑ demand price ≤ sales price
Legend: sales quantity deviation to demand quantity
Fig. 2 Principle of contract and spot demand and sales
Spot sales quantities are flexible and can be lower or higher than the forecasted
quantities for various reasons as shown in Fig. 2. Firstly spot sales are lower than the
demand quantity if the spot demand quantity exceeds the available supply and the
company needs to make monthly volume quotation decisions cutting volumes first on
an overall sales location level and then also on a detailed, individual customer level.
Merely for illustration reasons, the available supply is shown to be constant in Fig. 2.
Of course in the real application the available supply can vary, e.g., due to variations
in procurement quantities and production capacity. Secondly spot sales are reduced if
spot demand prices are too low compared to raw material costs forcing the company
to make a loss when supplying the customer. Hence the spot demand forecast has a
bid character as in single-sided auctions competing for limited supply. The bid can be
successful and is fully supplied or it can be partly or even fully rejected depending on
the available supply quantity and the bid price. Like in stock markets and exchanges,
bids need not be necessarily executed in the marketplace if the bid volume and price
cannot be cleared with a suitable offer.
Note that there are no penalties in spot business as it can be found in supply network
planning, where an artificial penalty is applied if demand cannot be met. These often
subjective penalties are not related to actual business agreements or actual monetary
penalties negotiated between the company and the customer. In our case customers
either have a fixed contract or they do flexible spot business on a tactical level. This
flexibility, however, does not destabilize the respective value chain operations since it
is limited to the tactical planning level and does not impact the operational order level.
Customers have a very early information and commitment on a monthly level whether
they receive the requested spot quantities or not. If a spot customer has received a con-
firmation, the supplier delivers the related orders accurately and with high reliability.
To summarize, in the considered industrial application demand is not regarded as a
given monolithic quantity to be fulfilled in the traditional supply chain management
sense but is defined as a mix of fixed contract demand and flexible spot demand.
Spot sales price–quantity functions and elasticities

Spot sales decisions have intuitive price effects as shown in Fig. 2: higher average
prices are achieved when cutting spot sales quantities or lower prices are required
when pushing additional quantities into the market. However, we do not assume a
monopolistic market situation, where the company is able to influence or dominate
market prices. In our case, price is a result of the spot sales quantity decision made
by the company. Since the price is an average price across several customer forecasts
grouped into one sales location, it is intuitive that the average price increases when sales
quantities are lower than the demand forecast quantity. It is assumed that customers
with lowest prices are cut first. Hence the average price across the remaining customers
increases. In this context, competitor behavior has no influence on this price–quantity
function. Competitors have influence on the overall market prices and the available
supply. However, the model focuses on bilateral negotiations between the company
and its customers. This business relationship is confidential, i.e., the competitor does
not know about spot quantities ordered by the customers and the corresponding spot
sales prices. The competitor does not even know to what extent a customer is supplied
on contract or spot basis. Since the business relationships are kept confidential the
competitor is not able to take specific reactions.
Spot sales price uncertainty

Spot demand quantity and prices are uncertain in the commodity business for the con-
sidered planning horizon. Since price is the main buying criterion, mid-term demand
quantity is mainly influenced by the price level. Additionally, commodity suppliers
often make supply volume decisions before sales prices are finally fixed due to com-
plex multi-stage production systems, large lead times in production and raw material
supply, lack of change-over flexibility in production with production plans fixed for
one month, planned shut-downs for maintenance as well as long transportation lead
times specifically in global value chain networks. Therefore, supply volumes in com-
modity business are fixed prior to sales prices in the market. Hence, the spot sales price
remains an uncertain parameter. In our investigation, the spot sales price is consid-
ered as uncertain leading to different price and sales turnover scenarios for the same
sales quantity. Therefore, contract demand quantity and price as well as spot demand
quantities are treated deterministic while spot sales prices are considered stochastic.
3.1.2 Derivation of price–quantity functions for spot demand
In the value chain network investigated the following entities have to be considered in
the formulation of the sales model:
• Products include finished products sold on the market, intermediate products pro-
duced and raw materials procured.
• Locations represent the nodes of the value chain network such as sales, production
or procurement locations.
• The planning horizon is divided into discrete time buckets (periods), months by
default.
Demand and sales are planned for all valid product–sales location combinations
{ p, l} ∈ P L S and a medium-term planning horizon covering periods t ∈ T .
Given the simplifying assumption that transit times between production and sales
locations and thus inventory balances can be neglected, the model can be separated by
time periods and periods could even be neglected. The same is true for locations. How-
ever, from an industry-practice perspective, it is important to show the monthly devel-
opment and interrelationships of sales and operation figures thus making the dynamics
in the value chain across business functions and the volatility in profits, prices and
volumes more transparent. Thus, instead of performing single period experiments a
model formulation is suggested that integrates all activities within the entire planning
horizon. Locations are essential as reference points for the aggregation of demand and
the determination of the price–quantity functions. In the real application, transit times
as well as intermediate inventories, safety stocks, etc. can easily be embedded into the
model formulation.
Demand input data comprise the demand forecast provided by the “Sales and Mar-
keting” organization of the company. Demand forecasts are aggregated from a single
customer level to an aggregated sales location level. The contract demand forecast
indicates the total demand quantities of all relevant products for each period and prod-
uct–location combination. In addition, the corresponding average sales price can be
derived from the customer contracts. Owing to the usual contract terms, total sales
turnover achieved from contract sales is fixed.
In contrast, total sales turnover achieved from spot sales depends on the decisions of
the company on spot prices and sales quantities for each period and product-location
combination. As explained in the previous subsection, the company receives quantity
and price bids from its spot market customers. In addition, the local “Sales and Market-
ing” units forecast expected bids for future periods. It is important that all spot sales
opportunities are forecasted as total demand bids regardless of whether production
capacity needed to fulfill this demand is available or not. It should be noted that the
resulting average spot sales price increases if the spot demand exceeds the production
capacity and the company selects the spot demand bids with the best spot sales prices.
This relationship is expressed by the elasticity ε defined as −ε = (p/ p) : (x/x).
Here, the elasticity can be interpreted as the change of the average spot price p with
respect to the change of the spot sales quantity x. Forecasting individual customer
spot demand for 6–12 months is more difficult than forecasting the overall spot mar-
ket demand. The latter is essential in order to evaluate if spot demand exceeds own
supply. To model the relationship between spot price and quantities, we show how
adequate price–quantity functions can be derived from the forecasted customer bids.
The derivation of price–quantity functions is based on the following major assump-
tions:
• The relationship between spot sales price and spot sales quantity can be modeled as
a linear function within the feasible minimum and maximum quantities defined by
the management of the company. Of course, the price–quantity relationship could
also be modeled using a non-linear function depending on the actual price–quantity
bids the company receives. In our case we found that the linear function showed a
sufficient statistical fit based on the real data provided by the company.
• External factors affecting the spot demand quantity, e.g., competitor actions, are
not considered, i.e., spot sales demand only depends on spot sales price for each
period and product–location combination.
The detailed steps of the algorithm for determining the price–quantity function for
spot sales demand and a numerical example are provided in Table 1. To keep the
presentation simple, we assume one single period and one individual product–loca-
tion combination, i.e., the corresponding indices are omitted. Given are spot demand
quantity qc and price forecast pc for individual customers c ∈ C (step 1). The spot
demand forecast by price and quantity represent future sales opportunities defined by
the “Sales and Marketing” unit of the company. Historical sales can provide some
guidance. However, anticipating future price trends in the market also depends on
future demand and the raw material price development. Note that the forecast does not
have to be necessarily discussed with the customer but can and should be based on the
market knowledge of the “Sales and Marketing” organization also reflecting targets
and new sales opportunities which “Sales and Marketing” wants to actively pursue in
the market. Next, all price forecasts are sorted in non-increasing order giving ranks
r = 1, . . . , R (step 2). In step 3, demand forecast quantities qc are summed up to a
cumulated spot demand quantity Q r for each rank r = 1, . . . , R with Q R being the
total demand quantity across all forecasts In step 4, the corresponding average spot
demand price forecast Pr for each rank r = 1, . . . , R is determined with PR being the
average price across all forecasts. In the following steps 5 and 6, the quantity share
Q r /Q R and the average price ratio Pr /PR of each rank r = 1, . . . , R are determined.
Table 1 Algorithm to determine the price–quantity function of spot demand and numerical example
Algorithmic steps Customer
A B C D
1. List individual customers c ∈ C with spot demand quantity qc and price forecast pc
Quantity (t) 100 200 100 200
Price (e/t) 100 90 80 70
2. Sort forecasts in non-increasing order of price using ranks r = 1, . . . , R
Rank 1 2 3 4
3. Determine cumulated spot demand quantity Q r for rank r = 1, . . . , R
Quantity (t) 100 300 400 600
4. Determine average spot demand price Pr for rank r = 1, . . . , R
Ø Price (e/t) 100 93.3 90 83.3
5. Determine quantity share Q r /Q R of rank r = 1, . . . , R
Quantity (%) 17 50 67 100
6. Determine average price ratio Pr /PR of rank r = 1, . . . , R
Price (%) 120 112 108 100
7. Perform linear regression for price ratios and quantity shares
Regression y = −0.2407·x + 1.2408; R2 =1.00
8. Determine price elasticity
Elasticity ε = 0.2407
In step 7 a linear regression for price ratios with respect to quantity shares is carried
out giving the price–quantity function. Finally, the spot demand elasticity is obtained
as the negative slope of the regression function (step 8). Note that the elasticity is
determined based on linear regression considering the quantity shares and average
price shares and not the absolute quantity and average prices in the price–quantity
function.
This proposed algorithm requires a sufficient number of individual customer bids
or forecasts within one sales location and thus relies on effective support by the local
“Sales and Marketing” units. If the number of price–quantity bids is not sufficient and
the regression is not accurate enough, elasticities cannot be directly used for decision
making. In this case, elasticity is assumed to be 0 meaning no price effects are included
in the model and calculated profits are lower and more cautious than in reality. If all
customers have the same spot prices, the average price is equal to the individual prices
and the elasticity is equal to 0 meaning that no average price effects occur in case of vol-
ume reductions. In the investigated example from the chemical industry, we observed
that price elasticities were volatile and ranked mainly between 0.1 and 0.5 different by
month, product and locations analyzed for 12 months. The number of customers for
one product and one location varied each month between 10 and 36. The R-squared
value for the linear regression varied monthly between 0.4 and 0.99. Without having
conducted a full elasticity analysis across the entire portfolio, the analysis helps to
prove market perceptions such as a higher elasticity exists in one market compared to
another market or comparing elasticity between products being perceived to have a
different elasticity. The statistical quality of the linear regression analysis in selected
months was considered as good in terms of the number of customers involved and the
R-squared value proving the applicability of the approach. Alternatively, a quadratic
regression of the sales turnover curve could be applied. This concept, however, does
not create the same basis for understanding in the “Sales and Marketing” organization
of the company since elasticity is the parameter known in “Sales and Marketing” to
discuss and understand price–quantity dynamics in the market rather than discussing
quadratic regression parameters that cannot be well understood and translated into
direct price-quantity-relations.
Another issue of considerable practical importance in commodity markets is the
uncertainty of market prices arising from a great number of external factors. In our
case study investigation, price uncertainty is reflected by alternative price scenarios
s ∈ S. In the real application, scenarios have to be defined for each product–loca-
tion combination. To keep the presentation simple, we again consider only one single
product–location combination.
To model the volatility of market prices, a price factor δs for spot demand price,
e.g., 0.8, 1.0 and 1.2, and a corresponding subjective scenario probability ωs valid for
the entire planning horizon have to be defined by management. Typically three sce-
narios “worst”, “best” and “average” are used in order to limit the complexity and
keep the scenario planning pragmatic. The price scenario philosophy of the company
is to have only one single sales plan with quantity x0 that is executed in the market at
different price levels ps . In addition we assume identical price–quantity functions, i.e.,
identical spot demand elasticity for all price scenarios meaning that the price factor
δs is impacting all customers homogenously not changing their spot demand volume.
p (x) p ( x) ⋅ x
Elasticity s∈S
∆p x0 s = 2 „best“
−ε = ⋅
∆x ps
s = 1 „average“
∆p
ps s = 3 „worst“
∆x „best“
„average“
„worst“
x x
x0 x0
x min x max x min x max
Fig. 3 Price–quantity function and sales turnover curve for individual price scenarios
Figure 3 illustrates the concept of scenario-based price–quantity functions, which

basically describe the dependency of sales price p on quantity x. With price–quantity
function p(x) the resulting sales turnover is given as p(x) · x. The scenario-based
price–quantity functions have different slopes but the elasticity considering relative
average price shares and relative quantity shares is identical in each scenario. In addi-
tion to given input data, sales control data are defined by the planner executing sales
and marketing business rules to set the boundaries for spot sales quantities. Control
parameters x min and x max indicate the minimum and maximum spot demand that
needs to be fulfilled as shown in Fig. 3.
The concept of scenario-dependent sales turnover functions represents a signifi-
cant advantage of demand price scenarios compared to demand quantity scenarios,
since the company does not have to manage different volume scenarios creating high
complexity in all areas of planning from sales to procurement. Moreover, the scenario
price factors can be directly applied to model the sales turnover in the objective func-
tion of the optimization model without affecting quantity constraints of the model.
This advantage might change the perspective on demand uncertainty from quantity
scenarios towards price scenarios related to a defined sales quantity. This is even more
practicable, since prices can be changed faster in practice compared to production
volumes and material flows. In particular in the production of chemical commodities,
considerable changeover times of the processing equipment have to be considered.
Moreover, transportation lead times and limitations on transit stock often reduce the
flexibility to adjust production quantities and redirect material flows on short notice.
3.1.3 Linear approximation of spot sales turnover
Since spot price and quantity depend on each other according to the linear price–
quantity function, the profit function is quadratic. In the following, we show how a
piecewise linear approximation of the sales turnover function can be achieved. This
y
actual turnover curve
~
y4
~
y turnover gradient
3
approximated
turnover
~
y2
add additional
partial quantity points i ∈ N
to improve approximation
~
y1
q~1 q~2 q~3 q~4 x
Spot sales quantity
x min x max
Fig. 4 Linear sales turnover approximation approach
approach is based on the concavity property of the sales turnover function and the
limited region of sales quantity flexibility to be considered. For a review of linear
approximation techniques for non-linear functions see Kallrath (2002b). For the opti-
mization problem investigated in this paper, some application-specific features have
to be considered which are described in the following. As in the previous subsection,
we skip the indices for periods and product–location combinations in order to improve
the understandability of the presentation.
The sales turnover approximation approach illustrated in Fig. 4 is based on partial
quantity points subdividing the sales turnover curve into multiple sections, for which
sales turnover is linearly approximated. As explained in the previous subsection, x min
and x max are given as management-defined control parameters, which indicate the
minimum and maximum spot demand that needs to be fulfilled, respectively. The set
of partial quantity points i ∈ N has four elements by default: 0, x min , Q R and x max ,
where Q R indicates the total quantity of all forecasted customer quantities (see the
algorithm for determining the price–quantity function in the previous subsection).
Note that x max > Q R expresses the possibility of gaining additional spot market
qauantity at lower sales prices. In the case of x max = Q R only forecasted orders are
considered. The three non-zero points are fixed and indexed by i min for x min , i mid
for Q R and i max for x max . The approximation can be improved by adding additional
partial quantity points i + between i min , i mid and i mid , i max , respectively. Partial spot
sales quantities q̃i are determined at each partial quantity point i ∈ N . Corresponding
partial spot sales turnover ỹi values are calculated for each partial spot sales quantity
q̃i using the exact sales turnover function. Partial spot sales turnover between two
partial quantity points is approximated based on the spot sales turnover gradient of
the linear connection for the partial quantity section j = 1, . . . , N − 1 between two
partial quantity points.
Since the sales turnover curve is concave and the linear sales turnover gradients
decrease monotonically, no integer variables are required to decide which partial quan-
tity section is filled first. The objective function to maximize sales turnover will ensure
to fill the partial quantity sections from left to right. It should be noted that the linear
sales turnover approximation does not depend on the individual sales price scenario.
3.1.4 Constraints of the sales model
Before the constraints that make up the sales model are presented, the respective nota-
tion has to be defined. Note that some variables, e.g., for modeling aggregate sales
figures, are introduced to improve the readability of the model formulation. These
variables could be replaced by the corresponding expressions.
Indices, index sets
p∈P products
l∈L locations
l ∈ LS sales locations
i∈N partial quantity points
j = 1, . . . , N − 1 partial quantity sections
t∈T periods
{ p, l} ∈ P L Sal valid product–sales location combinations
for sales products
Parameters
Sc
q plt contract demand quantity forecast for product–sales location
combination { p, l} and period t
Ss
X plt , X plt minimum and maximum spot sales quantity for product–sales
Ss
location combination { p, l} and period t, respectively

τ jSsplt spot sales turnover gradient of the linear sales turnover
approximation for partial quantity section j, product–sales
location combination { p, l} and period t
q̃iSs
plt partial spot sales quantity at partial quantity point i for
product-sales location combination { p, l} and period t
p Sc
plt contract sales price for product–sales location
Decision variables
S
x plt total sales quantity for product–sales location
Ss
x plt spot sales quantity for product–sales location
x̃ Ss
j plt partial spot sales quantity for partial quantity section j
for product–sales location combination { p, l} and
period t
Ss
y plt spot sales turnover for product–sales location
ỹ Ss
j pltpartial spot sales turnover for the partial quantity
section j for product–sales location
In the following, only the constraints of the sales model are presented. Additional
constraints of the supply model are presented in Sect. 3.2.3. Finally, the objective
function for maximizing the profit for the entire value chain network is defined in
Sect. 3.3. It should be noted that constraints of the sales model do not depend on the
individual spot sales scenario.
The total sales quantity is obtained as the sum of contract sales and spot sales
quantity:
S
x plt = x plt
Ss
+ q plt
Sc
∀{ p, l} ∈ P L Sal , t ∈ T (1)
The spot sales quantity is limited between minimum and maximum boundaries:
Ss
plt ≤ x plt ≤ X plt ∀{ p, l} ∈ P L , t ∈ T
X Ss Ss Sal
(2)
The total spot sales quantity equals the sum of the partial spot sales quantities:

N −1
Ss
x plt = j plt ∀{ p, l} ∈ P L , t ∈ T
x̃ Ss Sal
(3)
j=1
Partial spot sales quantities need to fit in the respective section between consecutive
partial spot sales quantities:
j plt ≤ q̃i plt − q̃i−1, plt ∀{ p, l} ∈ P L , i ∈ N , i > 1, j = 1, . . . , N − 1, t ∈ T

x̃ Ss Ss Ss Sal
(4)
Partial spot sales turnover is given as the product of partial quantity and partial sales
turnover gradient:
j plt = τ j plt · x̃ j plt ∀{ p, l} ∈ P L , j = 1, . . . , N − 1, t ∈ T

ỹ Ss Ss Ss Sal
(5)
The spot sales turnover equals the sum of the partial spot sales turnovers:

N −1
Ss
y plt = j plt ∀{ p, l} ∈ P L , t ∈ T
ỹ Ss Sal
(6)
j=1
Further constraints, for example, on sales contract quantity rules and flexibility are
possible but excluded here.
3.2 Supply model
3.2.1 Procurement and consumption of raw materials for chemical commodities
Variable raw material consumption rates in production

Raw material consumption in production is traditionally treated as constant based
on given recipe factors. Recipe in the chemical industry is a synonym for the bill-
of-material in discrete parts manufacturing and includes all input products with their
respective input fraction required to produce one unit of one or several output products
in a production process. However, in chemical production the degree of raw material
consumption rates and hence the recipe factors often depend on the processing mode of
the equipment, which can be employed at different utilization or throughput levels. In
this case the recipe is not composed of static input factors but of recipe functions, which
express the relationship between the input consumption and the output quantity pro-
duced. Hence the problem of how to decide on raw material consumption and how to
balance volatile raw material costs with sales quantities and prices needs to be solved.
Spot and contract procurement

Raw materials are procured either based on fixed contracts or on the spot market.
Differences in spot and contract prices have been observed in many business sectors,
cf., Reiner and Jammernegg (2005). In analogy to the demand side, procurement con-
tracts are fixed by quantity and price with the objective to ensure a basic supply of raw
materials. Spot procurement supports the requirements of the company for flexibility
in supply and sales planning facing uncertain market prices. By utilizing spot procure-
ment the company can decide the actual procurement quantity with certain flexibility
around the offered quantity. Price levels for contracts and spot business differ and
are volatile.
3.2.2 Modeling flexible recipes
Key issues of the supply model are to decide on the variable raw material consump-
tion rates in production and on spot procurement quantities. Both issues are highly
interrelated, i.e., high production rates determine the amount of raw material that has
to be supplied. Moreover, raw material costs per output ton produced can grow with
higher production utilization and throughput rates. In the overall context of value chain
optimization, production rates have to comply with decisions reflected by the sales
model, e.g., on spot sales quantities and prices.
In the following, the basic principle of flexible recipes is presented. To keep the
explanations simple, we consider only one single type of finished product that is pro-
duced from one single raw material on one resource at a specific location during a
given period, i.e., indices for input and output products, resources, locations, and peri-
ods are omitted. In the real application, however, there are multiple input products.
Typically, one input product represents the main feed into the production process while
the others are auxiliary substances which can be procured on short notice.
Let C denote the production capacity of the resource measured in tons of output
per period and let x in and x out indicate the input of raw material and output of finished
products, respectively. Capacity utilization is defined as U = x out /C. Minimum utili-

zation rates and the capacity as maximum utilization rate have to be maintained. Even
in periods with extremely low demand, production processes must run at a minimum
utilization rate to ensure process stability and product quality. A complete shut-down
of an asset is technically feasible, for example, in the case of planned maintenance
or in emergency cases but not considered as a planning option in regular operations.
In many types of chemical mass production, raw material consumption depends on
the utilization rate of the equipment employed. Hence, linear recipe functions can
be derived, which indicate the input of raw material required to produce the desired
amount of output.
In Table 2 the derivation of linear recipe functions is explained using a numerical
example. Utilization rates are given in steps of 20% assuming that all rates are used
with equal probability. Capacity is given at 1,440 tons per day. The next two rows
indicate pairs of input and output quantities for each utilization rate. These figures can
be derived from the technological parameters of the production equipment. The recipe
factor is defined as the ratio of input to output quantities. Note that recipe factors only
refer to the main raw material and do not include other input materials. This explains
the value of the recipe factor of less than 1.0 for U = 20%. Finally, linear regression is
applied with respect to the recipe factors. As a result, a variable consumption factor of
a = 1.3 and a constant factor of b = −144 are obtained based on the given utilization
rates and the underlying technological parameters.
The special case of a static recipe is given for b = 0. In this case, the raw material
consumption does not change with capacity utilization. Otherwise the recipe factor
grows with increasing resource utilization. The linear recipe function for the example
of Table 2 is illustrated in Fig. 5. As a reference case, the static recipe is shown. Linear
recipe functions are one type of recipe found in chemical industry. Of course other
forms of recipe functions are possible depending on the consumption pattern analyzed
for a specific resource.
In the supply model production input and output quantities depend on each other
according to the recipe functions. Output and utilization decisions determine the raw
material quantities to be supplied at each location. As mentioned before, raw materi-
als are procured in both spot and contract mode. While contract procurement needs
to be executed as agreed, spot procurement is flexible with minimum and maximum
quantities for each type of raw material and each product–location combination. The
company decides on spot procurement quantities within these intervals.
Table 2 Linear recipe function example
Utilization rate U 20% 40% 60% 80% 100%
Capacity C 1,440 1,440 1,440 1,440 1,440

Raw material input quantity x in 230 605 979 1,354 1,728
Production output quantity x out 288 576 864 1,152 1,440
Recipe factor (x in /x out ) 0.80 1.05 1.13 1.18 1.20
Linear regression w.r.t. recipe factors x in = 1.3 · x out − 144; a = 1.3, b = −144
R 2 = 0.99
Recipe 1.5
factor
Linear
recipe
1.0
Static
recipe
0,5
0
0% 20% 40% 60% 80% 100%
Utilization
Fig. 5 Static recipe and linear recipe function
3.2.3 Constraints of the supply model
Before the constraints which make up the supply model are presented, the respective
notation has to be defined. Note again that some aggregate (redundant) variables are
introduced to improve the readability of the model formulation.
Indices, index sets
p∈P products
l∈L locations
r∈R resources
t∈T periods
l ∈ LP production locations
l ∈ LO procurement locations
l ∈ LS sales locations
{ p, r } ∈ P R in , P R out valid input/output product–resource combinations
{ p, l} ∈ P L in , P L out valid input/output product–production
location combinations
{ p, l} ∈ P L Proc valid product–procurement location
combinations for procured products
{ p, l} ∈ P L Sal valid product–sales location combinations
for sales products
Parameters
CrPt capacity of production resource r in period t
UrP min minimum utilization rate of production resource r
a pr , b pr parameters of the linear recipe function for
product–resource combination { p, r }
dt number of production days in period t
c Pvar
pr variable production cost per unit for product–resource
combination { p, r }
Pspot
c plt average cost rate per unit for spot procurement for
product–procurement location combination { p, l} in period t
c Pcon
plt average cost rate per unit for contract procurement of
product–procurement location combination { p, l} in period t
Pcon
q plt contract procurement quantity for product–procurement
location combination { p, l} in period t
Pspot Pspot
X plt , X plt minimum and maximum spot procurement quantity for
product–procurement location combination { p, l}
in period t, respectively
Decision variables
Pout
x pr production output quantity for product–resource combination
t
{ p, r } in period t
Pin
x pr production input quantity for product–resource combination
t
{ p, r } in period t
Pout
x plt production quantity for product–production location combination
{ p, l} in period t
Pin
x plt secondary demand in product–production location combination
{ p, l} in period t
Pspot
x plt procurement spot quantity for product–procurement location
combination { p, l} in period t
Proc
x plt procurement quantity for product–procurement location combination
{ p, l} in period t
S
x plt total sales quantity for product–sales location
v pr
Pvar
t variable production costs for product–production location
combination { p, l} in period t
v plt
Proc procurement costs for product–production location combination
{ p, l} in period t
In the following, the constraints of the supply model are presented.
Capacity and minimum utilization rate limit the total production quantities of all
products produced on the resource in a specific period:

UrP min · CrPt ≤ t ≤ Cr t
Pout
x pr P
∀r ∈ R, t ∈ T (7)
{ p,r }∈P R out :r =r
The input quantity of intermediate or raw material products required depends on

the production rate of the resource and the linear recipe function which is determined
on a tons per day basis. Hence the number of production days needs to be considered
in constraint (8).
⎛ ⎞

Pin
x pr ⎝
t = a pr · r t ⎠ + b pr · dt
x pPout ∀{ p, r } ∈ P R in , t ∈ T (8)
{ p ,r }∈P R out
Production output and input quantities are aggregated at the location level.

Pout
x plt = x pPout
r t ∀{ p, l} ∈ P L out , t ∈ T (9)
{ p ,r }∈P R out : p = p

Pin
x plt = r t ∀{ p, l} ∈ P L , t ∈ T
x pPin in
(10)
{ p ,r }∈P R in : p = p
Given the simplified network with two dedicated resources at a single production
location these constraints are not required. However, for practical reasons it is impor-
tant to keep locations and resources separated, since key effects such as flexible recipes
are related to specific resources and their technology rather than to an entire production
location.
The total variable production costs are obtained as product of production quantity
and variable production cost rate:
v pr t = c pr · x pr t ∀{ p, r } ∈ P R out , t ∈ T
Pvar Pvar Pout
(11)
Total procurement costs are calculated based on variable spot procurement quanti-
ties and fixed contract procurement quantities:
Pspot Pspot
v plt
Pvar
= (x plt · c plt ) + (q plt
Pcon
· c Pcon
plt ) ∀{ p, l} ∈ P L
Proc
,t ∈ T (12)
Total procurement quantities are obtained by summing up spot and contract pro-
curement quantities:
Pspot
plt = x plt
x Proc + q plt ∀{ p, l} ∈ P L Proc , t ∈ T
Pcon
(13)
Total spot procurement quantity is limited between the minimum and maximum
boundaries:
Pspot Pspot Pspot
X plt ≤ x plt ≤ X plt ∀{ p, l} ∈ P L Proc , t ∈ T (14)
The following equation balances total supply quantities consisting of production

and procurement quantities with total demand consisting of total sales quantity and sec-
ondary demand of production based on the assumption of single sourcing. In practice,
material balances also include inventories and transportation quantities which are
important in global networks with several weeks lead times and considerable transit
inventories. Since we focus on the integration of business functions in value chains,
these issues are beyond the scope of this paper.

t +
Pout
x pl x Proc
pl t
l ∈L P :l =l l ∈L O :l =l
∀{ p, l} ∈ P L out , P L in ,
= Pin
x pl t +
S
x pl t (15)
P L Proc , P L Sal , t ∈ T
l ∈L P :l =l l ∈L S :l =l
3.3 Optimization strategies
The objective of the proposed modeling approach is to maximize profit for the entire
value chain network. It is assumed that the company behaves risk-averse in face of
the price uncertainty and seeks to ensure minimum profits. Two optimization strate-
gies can be applied incorporating spot sales price scenarios to reflect price uncertainty
(cf. Chen and Lee 2004):
• One-phase optimization: maximize expected profit across one or multiple price

scenarios. This approach corresponds to the classical “expected value” maximiza-
tion known from decision theory.
• Two-phase optimization: maximize expected profit across multiple price scenarios
taking into account the constraint that a given minimum profit value is reached.
From a practical point of view, this approach seems to be more appropriate in
situations where a high variability of profit can be expected and the risk of low
profit outcomes shall be minimized.
The one-phase optimization strategy considers one or multiple spot price scenarios.
Each scenario (see Sect. 3.1.2) is characterized by the spot price factor δ plst , which
expresses possible spot price levels, e.g., 0.8, 1.0, and 1.2, for each relevant product–
location combination { p, l}, period t and scenario s ∈ S. Each scenario is assigned a
subjective probability ωs . While supply decisions remain unchanged, the various spot
price scenarios lead to multiple sales turnover scenarios that are realized with the same
spot sales quantity. Since price scenarios are represented by specific price factors, they
can be directly applied to model spot sales turnover in the objective function.
The expected profit determines the average profit across all price scenarios weighted
with their scenario probability ωs . With the notation defined in Sects. 3.1.4 and 3.2.3
the expected profit function can be defined as follows:
(16)
The expected profit across multiple scenarios provides a more realistic picture of
the future profit situation compared to one single scenario. However scenarios are
consolidated and expressed as a single value based on their probability weights. The
planner would have no information about potential worst case profits and might like
to sacrifice expected profit opportunities for safety in exchange. This is addressed by
the two-phase optimization approach.
The two-phase optimization strategy (see Figure 6) first maximizes the minimum
scenario profit z min , which is lower or equal to all single scenario profits z s , where z s
Phase 1 Phase 2
max z min max z exp

z min → z min*
subject to subject to
zs ≥ z min
, ∀s ∈ S zs ≥ z min*
and all other constraints and all other constraints
Fig. 6 Two-phase optimization strategy
is defined as follows:
⎡

zs = ⎣ Ss
y plt · δ plst + plt · q plt
p Sc Sc
t∈T { p,l}∈P L S { p,l}∈P L S

⎤

− v pr t −
Pvar
v Proc
plt
⎦ (17)
{ p,r }∈P R out { p,l}∈P L Proc
This first phase determines the best minimum profit z min from all scenarios. z min is
then fixed as baseline profit z min ∗ for the second phase of the optimization, where the
expected profit z exp is maximized across all scenarios given the condition that each
scenario profit reaches the minimum scenario z min ∗ . This concept aims to obtain more
robust solutions considering probabilistic demand quantity scenarios.
4 Case study evaluation
The optimization model presented in the previous section was implemented in ILOG
OPL Studio 3.71 using CPLEX 9.1 as solver and was tested with industry case data
on an Intel Pentium 4 PC with 1598 MHz and 256 MB RAM. Table 3 indicates the
number of entities included in the case study evaluation.
For confidentiality reasons data from the company are sanitized in a way that data
used for the case study evaluation are generated reflecting realistic dimensions of the
investigated business application. However, data used in the simulation show the same
scale. Several numerical experiments were carried out in order to analyze the impact
of integrating sales and supply decisions by volumes and values based on the devel-
oped value chain planning model. Numerical results are presented in the following
subsections.
4.1 Price scenario experiments
In the first experiment we compare the optimization strategies introduced in subsec-

tion 3.3 for different spot price scenarios. Two alternative demand spot price scenarios
“best case” and “worst case” with equal probability of 0.25 are defined in addition to
Table 3 Number of entities in

Basic elements Number of entities
the case study evaluation
Products 50
- Finished 48
- Intermehdiate 1
- Raw material 1
Locations 11
- Sales 9
- Production 1
- Procurement 1
Resources 2
- Continuous 1
- Multi-purpose 1
Periods 6
1-phase optimization 2-phase optimization

Index Index
200 200
less
100 robust, 100 more
more robust,
extreme less
solutions extreme
solutions
0
0 Period
Period 1 2 3 4 5 6
1 2 3 4 5 6
Best profit index Best profit index
Expected profit index Expected profit index
Worst profit index Worst profit index
Sales quantity index Sales quantity index
Fig. 7 Comparison of the 1-phase and 2-phase optimization strategies
the standard scenario with probability 0.5. The best case assumes a continuous price
increase while the worst case assumes a continuous price decrease. Consequently the
expected profit is the average of the best and worst case scenario results and equiva-
lent to the standard scenario result in this special case. Numerical results are shown in
Figure 7 for a planning horizon of six periods. Results of the one-phase optimization
strategy show relatively constant sales quantities and expected profits slightly below
the index value of 100. The results of the first period are indexed at 100 in order to
compare the results of the subsequent periods with the first period.
Executing this sales plan can lead to very positive best-case scenario profits
but also to very poor profits, if the worst-case price scenario occurs. Less extreme
plans can be reached with the two-phase optimization strategy: scenario profits are less
variable and the worst case scenario results are comparatively better than in the
Fig. 8 Elasticity model reaction Quantity

test results
Profit
index indices
400 120
110
300
220 100
200
187
155
125 90
89 100
100
80
0 70
0.0 0.2 0.4 0.6 0.8 1.0 Elasticity
Profit index
Sales index
Production index
Procurement index
one-phase-optimization strategy. The overall value chain plan in sales, production

and procurement is more cautious with lower sales quantities and lower expected
profits as the pay-off for better minimum profits.
To conclude, the two-phase optimization results in lower average profits. In the real
application, planners might also vary the subjective weights for the different scenar-
ios or set alternative minimum profit levels. This way additional information on the
robustness of the obtained solution and a better understanding of the complex rela-
tionships between volumes and values in a price-volatile commodity business could
be gained.
4.2 Spot price elasticity test
The price elasticity of spot demand for all finished products is varied in multiple
scenarios from 0 to 1. The base plan has the elasticity 0.2 and results for the elas-
ticity value of 0.2 are indexed at 100. Elasticity of 0.2 means that the average sales
price increases by 2%, if sales volumes are decreased by 10% and vice versa. Experi-
ments are conducted applying the sales turnover approximation method with 24 partial
quantity points to reach a high accuracy of the approximation as it will be evaluated in
subsection 4.4. Experimental results shown in Figure 8 reveal that different elasticities
lead to different optimal profits and quantities in sales, production and procurement,
since sales volume-dependent average price effects are considered in the model.
In the specific case the base plan with an elasticity of 0.2 leads to a situation of
under-utilization of production capacity, since high raw material costs can not always
be compensated by sales prices. Higher elasticities lead to higher sales volumes, capac-
ity utilization and profit increase since the relative sales volume increase can be realized
with a lower relative average sales price decrease. In this situation it is profit-optimal to
increase production and push additional sales volume into the market with lower sales
prices. Full utilization is reached for elasticities of 0.6. Consequently an elasticity of
0.0 leads to lower profits, lower sales volumes and production utilization, since the
insufficient sales price level does not change with sales quantity decisions.
Note that higher elasticity leads to higher sales and production volumes in this
specific case of under-utilization due to the specific raw material prices and recipe
functions. In case of full-utilization and different raw material prices and recipe func-
tions, higher elasticities can also lead to a reduction of sales and production volumes
if reduction of sales volumes and the respective increase of the average price is profit-
optimal compared to supply costs. To conclude, the test demonstrates the influence
elasticity can have on commodity sales and supply decisions and resulting profits.
Our numerical results reveal that considering average price effects reflected by
elasticities can have significant influence on the overall volume plan. Hence, a profit-
optimal supply and production plan does not necessarily maximize capacity utiliza-
tion. Therefore, firms should not try to change or reduce elasticities but consider them
in their sales and production planning taking the profit impact of price effects into
account. Focusing on volumes alone and not considering existing price elasticity will
lead to suboptimal plans and reduced profits.
4.3 Raw material price experiments
The influence of raw material prices on profit and utilization is investigated in the
third experiment. Production capacity appears to be a bottleneck not sufficient to serve
demand with full spot sales flexibility and elasticity of 0.2. Prices for the raw material
required in the intermediate production process are varied around a basis index of 100
from 80 to 140. Two raw material recipe scenarios are considered: a static recipe fac-
tor of 1.2 and a linear recipe function, where raw material consumption rates increase
from 0.8 at 50% utilization to 1.2 at 100% utilization. Figure 9 shows the results of
the raw material price scenario experiments.
It is obvious from Fig. 9 that profit decreases in all cases, the more procurement
prices increase. However the static recipe leads to comparatively higher sales and pro-
duction volumes and lower profits compared to the case with the linear recipe function.
The reason is that the static recipe represents the maximum factor of the recipe func-
tion that does not change. In comparison raw material consumption rates and costs
can be decreased in the case of linear recipe functions by reducing the production
utilization. Therefore, all volume indices are reduced in the case of linear raw material
consumption, since raw material costs due to higher prices can be saved lowering
production and raw material consumption. The opposite effect occurs if the maximum
value of the recipe function values is higher than the static recipe factor. In both cases
raw material unit costs cannot be directly allocated to production output as basis for
product profitability and contribution margin analysis since raw material quantities
and costs depend on overall value chain planning decisions. Our numerical results
reveal that a recipe function with different raw material consumption rates depending
on production utilization has a major impact on the optimal profit and on capacity
Linear recipe function Static recipe

Profit Quantity Profit Quantity
index indices index indices
400 400
100 100
300 300
50 50
200 200
142 0 143 0
120 121
100 100
100
83 70 79
60 53 -50 100 59 -50
42 28
0 -100 0 -100
80 90 100 110 120 130 140 Raw 80 90 100 110 120 130 140 Raw
Profit index material Profit index material
Sales index price index Sales index
Production index price index
Production index
Procurement index Procurement index
Fig. 9 Raw material price model reaction test results
Table 4 Sales turnover approximation model performance test
Partial quantity points 4 6 8 14 24 44 64
Constraints (thou.) 27 37 48 79 131 234 338

Variables (thou.) 24 35 45 76 128 232 335
Solution time (s) 1 2 2 5 28 118 276
Profit gap (%) 10.79 2.35 1.05 0.25 0.07 0.01 Basis
utilization. Not considering these dynamics would endanger a company’s profitability

when focusing only on maximizing production utilization.
4.4 Accuracy of the sales turnover approximation
Finally, the accuracy of the piecewise linear sales turnover approximation method
is tested using the industry test data set and elasticities of 0.2. The number of partial
quantity points is varied from 4 to 64 as shown in Table 4. Numerical results reveal that
already 24 partial quantity points are sufficient to reach 99.93% of the objective func-
tion value obtained for the very accurate approximation based on 64 partial quantity
points. The approximation is even more accurate if sales quantity flexibility is close to
the forecast point. The approximation is less accurate, if spot sales quantities can be
cut entirely as in the test data, since the sales turnover curve has highest gradients near
the point of origin, where the gap between actual and approximated sales turnover is
highest.
Considering the tactical planning purpose, run times of 1 min or less are acceptable
in practice. These short run times even enable a planner to evaluate different scenarios,
i.e., running the model with different parameter settings. To utilize the scenario mode
of the optimization model, a small number of partial quantity points would suffice
thus permitting solutions within only a few seconds.
4.5 Industrial application
As mentioned before, the presented optimization model has been developed as a pro-
totype model to support the introduction of the Supply Network Planning module of
an advanced planning system. In particular, our model helps to determine the required
scope of the Supply Network Planning module implementation, to reveal the necessity
of customizing the standard advanced planning software, and to evaluate the possible
benefits for the company. Hence, the focus of our model formulation was to reflect the
company’s key optimization problem, namely balancing the consumption of a basic
price-volatile raw material which is processed in continuous production mode with
output volumes of more than 1 Mio. tons per year and coordinating the respective
sales, production and procurement activities.
Prior to the implementation of an enhanced optimization model, the industrial com-
pany started a major business reorganization project in order to improve the coordina-
tion of business functions from procurement to sales for their global production sites
and sales representations. In the course of this project several of the key instruments
included in our model formulation were put into practice, in particular, the concept
of spot and contract demand management and the instruments of demand elasticities
and turnover functions as well as linear raw material recipe functions. Major effects
of their application are the following.
• Changing the planning philosophy from pure demand fulfilment, which can be
seen as the traditional supply chain management orientation, towards focussing on
the global value chain profit by introducing demand management concepts based
on the differentiation between spot and contract demand with active spot sales
decisions helped the company to turn around the loss-making business unit into a
highly profitable one.
• Recognizing the effects of spot demand elasticity and applying them in an inte-
grated value chain planning effort provided the company additional insights into
the dynamics of the global markets they are operating in. In fact, several markets
show very high elasticities with significant price differences for the same product
while demand on other markets is fairly insensitive to sales prices. These insights
helped the company to make better sales decisions. Specifically, in the case of
supply shortages decisions on cutting spot sales volumes could directly be derived
from the model calculations.
• Incorporating linear raw material consumption functions and variable prices into
the value chain planning model directly identified potential cost savings of several
Mio. $ per year. Formerly, the company used to fully utilize production capac-
ity. After gaining insights from the model application into the interdependencies
between procurement and sales volume and prices and the use of different produc-
tion modes, the company recognized that this is not necessarily the profit-optimal
production strategy. Now managers seek to determine differentiated profit-optimal
utilization levels for their key production assets at three global sites. As a result,
production volumes are shifted from less resource-efficient assets to the more
efficient ones in the global network.
• The basic model is used by three global value chain planners for monthly plan-
ning of a global business unit. The model helped them to better understand the
profitability levers in the value chain network from procurement to sales.
Further benefits are expected from introducing the Supply Network Planning module
of an advanced planning software system to be used jointly with the Demand Planning
and ATP/CTP modules which have already been implemented.
5 Summary and outlook
In this paper a model is presented to coordinate sales and supply decisions for com-
modities in a chemical industry value chain. Price–quantity functions, volatile and
uncertain prices, flexible quantities in sales, procurement and production as well as uti-
lization-dependent recipes create complex interdependencies which make it extremely
difficult for the human planner to determine profit-optimal network-wide sales and
supply plans even for small-sized value chain networks. Price–quantity function elas-
ticities support decisions toward sales volume reductions or increases considering the
effect of increasing or decreasing average prices. We evaluated the piecewise linear
approximation approach to decide on sales turnover with sales price and volumes
as variables. The approximation delivered very accurate results within short solution
times and thus can be seen as an efficient approach to solve the underlying quadratic
optimization problem. Variable raw material recipes have a direct impact on volumes
and values, if raw material prices cannot be compensated by sales prices. Applying
two-phase optimization strategies for sales price scenarios leads to more robust plans
ensuring target profitability even in case of worst-case prices with the pay-off of more
cautious and lower expected profits.
The model presented in this paper has been implemented by the company as basis
for numerical investigations. The company has extended the basic model with further
features such as inventory balances and transportation activities as well as exchange
rates and further specifics of chemical commodity production such as throughput
smoothing. Contract and spot sales planning has been implemented in their APS-
based demand planning system and procurement planning for key raw materials have
been established by their global purchasing department. Implementing these integrated
sales and supply planning tools has shown major effects on the overall profitability
of the business unit. Specifically the spot price mechanism used to better coordinate
sales and supply decisions showed a major impact for the company.
Integrating sales and supply decisions throughout the value chain poses new
interdisciplinary research questions as an outlook. Neither supply network planning
minimizing costs to fulfill given demand nor revenue management maximizing rev-
enue based on a given supply adequately addresses the problem of managing an
industrial value chain end-to-end by volumes and value. Business rules for selling
production output profit-optimal in contract or spot business as well as alternative
methods to model price–quantity functions considering the impact on supply and
profit are potential further areas of research. The overall research focus may shift from
supply flexibility and cost minimization towards end-to-end supply, sales and pricing
decisions to utilize the value chain in the most profitable way.
References
Asche F, Gjölberg O, Völker T (2003) Price relationships in the petroleum market: an analysis of crude oil
and refined product prices. Energy Econ 25: 289–301
Bartodziej P, Derigs U, Zils M (2007) O&D revenue management in cargo airlines—a mathematical pro-
gramming approach. OR Spectrum 29: 105–121
Berning G, Brandenburg M, Gürsoy K, Mehta V, Tölle FJ (2002) An integrated system solution for supply
chain optimization in the chemical industry. OR Spectrum 24: 371–401
Biller S, Chan LMA, Simchi-Levi D, Swann J (2005) Dynamic pricing and the direct-to-customer model
in the automotive industry. Electron Commerce Res 5: 309–334
Blömer F, Günther H-O (2000) LP-based heuristics for scheduling chemical batch process. Int J Product
Res 38: 1029–1051
CEFIC (2005) Facts and figures—the European chemical industry in a worldwide perspective: July 2005.
The European Chemical Industry Council, Brussels
Chakravarty AK (2005) Global plant capacity and product allocation with pricing decisions. Eur J Oper
Res 165: 157–181
Chen C, Lee W (2004) Multi-objective optimization of multi-echelon supply chain networks with uncertain
product demands and prices. Comput Chem Eng 28: 1131–1144
Defregger F, Kuhn H (2007) Revenue management for a make-to-order company with limited inventory
capacity. OR Spectrum 29: 137–156
Dickersbach JT (2006) Supply Chain management with APO: structures, modelling approaches and imple-
mentation peculiarities. 2nd edn, Springer, Berlin
Gosavi A, Ozkaya E, Kahraman AF (2007) Simulation optimization for revenue management of airlines
with cancellation and overbooking. OR Spectrum 29: 21–38
Grunow M (2001) Management von Produktions- und Distributionsnetzwerken in der chemischen Industrie.
In: Sebastian H, Grünert T (eds) Logistikmanagement—supply chain management and electronic busi-
ness. Teubner, Stuttgart, pp 323–335 (in German)
Grunow M, Günther H-O, Yang G (2003) Plant co-ordination in pharmaceutics supply networks. In: Günther
HO, van Beek P (eds) Advanced planning and scheduling solutions in process industry. Springer, Ber-
lin, pp 261–294
Gupta A, Maranas CD (2003) Managing demand uncertainty in supply chain planning. Computers Chem
Eng 27: 1219–1227
Kallrath J (2002) Planning and scheduling in the process industry. OR Spectrum 24: 219–250
Kallrath J (2002b) Gemischt-ganzzahlige Optimierung: Modellierung in der Praxis. Vieweg, Braunschweig,
(in German)
Kilger C, Wagner M (2008) Demand planning. In: Stadtler H, Kilger C (eds) Supply chain management
and advanced planning. 4th edn. Springer, Berlin, pp 133–160
Lee LH, Chew EP, Sim MS (2007) A heuristic to solve a sea cargo revenue management problem. OR
Spectrum 29: 123–136
Marquez AC, Blanchar C (2004) The procurement of strategic parts. Analysis of contracts with suppliers
using a system dynamics simulation model. Int J Product Econ 88: 29–49
Meyr H (2008) Forecast methods. In: Stadtler H, Kilger C (eds) Supply chain management and advanced
planning. 4th edn. Springer, Berlin, pp 505–516
Neumann K, Schwindt C, Trautmann N (2002) Advanced production scheduling for batch plants in process
industries. OR Spectrum 24:251–279
Reiner G, Jammernegg W (2005) Bewertung unterschiedlicher Beschaffungsstrategien für Risk-Hedging
Supply Chains unter Berücksichtigung intermodaler Transportprozesse. In: Günther H-O, Mattfeld
DC, Suhl L (eds) Supply Chain Management und Logistik: Optimierung, Simulation, Decision
Support. Physica, Heidelberg, pp 115–134
Reiner G, Natter M (2007) An encompassing view on markdown pricing strategies: an analysis of the
Austrian mobile phone market. OR Spectrum 29: 173–192
Seifert RW, Thonemann UW, Hausman WH (2004) Optimal procurement strategies for online spot markets.
Eur J Oper Res 152: 781–799
Spengler T, Rehkopf S, Volling T (2007) Revenue management in make-to-order manufacturing—an appli-

cation to the iron and steel industry. OR Spectrum 29: 157–171
Stadtler H (2005) Supply chain management and advanced planning—basics, overview, and challenges.
Stadtler H (2008) Purchasing and material requirements planning. In: Stadtler H, Kilger C (eds) Supply
chain management and advanced planning. 4th edn. Springer, Berlin, pp 217–229
Tang CS (2006) Perspectives in supply chain risk management. Int J Product Econ 103: 451–488
Thomas DJ, Griffin PM (1996) Coordinated supply chain management. Eur J Oper Res 94: 1–15
Timpe CH, Kallrath J (2000) Optimal planning in large multi-site production networks. Eur J Oper Res
126: 422–435
MILP-based campaign scheduling in a specialty
chemicals plant: a case study
Marcus Brandenburg · Franz-Josef Tölle

OR Spectrum (2009) 31:141–166
DOI 10.1007/s00291-007-0084-5
Abstract Supply chain management in chemical process industry focuses on

production planning and scheduling to reduce production cost and inventories and
simultaneously increase the utilization of production capacities and the service level.
These objectives and the specific characteristics of chemical production processes
result in complex planning problems. To handle this complexity, advanced planning
systems (APS) are implemented and often enhanced by tailor-made optimization algo-
rithms. In this article, we focus on a real-world problem of production planning arising
from a specialty chemicals plant. Formulations for finished products comprise several
production and refinement processes which result in all types of material flows. Most
processes cannot be operated on only one multi-purpose facility, but on a choice of
different facilities. Due to sequence dependencies, several batches of identical pro-
cesses are grouped together to form production campaigns. We describe a method
for multicriteria optimization of short- and mid-term production campaign scheduling
which is based on a time-continuous MILP formulation. In a preparatory step, deter-
ministic algorithms calculate the structures of the formulations and solve the bills of
material for each primary demand. The facility selection for each production campaign
is done in a first MILP step. Optimized campaign scheduling is performed in a second
step, which again is based on MILP. We show how this method can be successfully
adapted to compute optimized schedules even for problem instances of real-world
size, and we furthermore outline implementation issues including integration with an
APS.
M. Brandenburg (B)
Beiersdorf AG, Unnastr. 48, 20245 Hamburg, Germany
F.-J. Tölle
Bayer Business Services GmbH, 51368 Leverkusen, Germany

316 M. Brandenburg, F.-J. Tölle
Keywords Supply chain management · Campaign planning and scheduling ·

Chemical process industry · MILP · Advanced planning systems
A list of symbols
Master data/problem parameters

F Set of facilities
f ∈F Facility
P Set of products
p∈P Product
A Set of processes
a∈A Process
Af ⊆ A Set of processes that can be operated on a facility f ∈ F
Fa ⊆ F Set of facilities that can operate process a ∈ A
z a, f Cycle time of process a ∈ A operated on facility f ∈ F
ba, f Batch size of process a ∈ A operated on facility f ∈ F
ap ∈ A Unique process (apart from refinement) that produces product
p∈P
Pa− , Pa+ ⊆ P Set of input, output products of process a ∈ A
− and δ +
δa, Input, output amounts of product p ∈ P for process a ∈ A
p a, p
(fractions of batch size ba, f )
B p = (A p , Pp , F p ) Formulation of product p ∈ P
Pp = ∪ p∈A p Pa+ Set of output products of the processes in A p
Ap Processes that have to be operated to produce product p ∈ P
F p = ∪a∈A p Fa Set of facilities on which processes of A p can be operated
sta,a Duration of the set up activity required between processes
a, a ∈ A
R⊆A Set of processes which have the refinement property
p∗ Off spec product
a∗ Refinement process
R∗ Set of refinement processes
Transactional data/instance parameters

E Set of order elements
∈E Order element
π Product ordered by order element ∈ E
q Order quantity of order element ∈ E
t Due date of order element ∈ E
c Production campaign
ts Starting time of a campaign
tc Completion time of a campaign
n Number of batches of a campaign (“campaign size”)
v Set up indicator of a campaign (v = 1 ⇔ set up activity is
performed before campaign c starts)
MILP-based campaign scheduling in a specialty chemicals plant: a case study 317
C Set of all campaigns linked to order element ∈ E (“campaign chain”)

C∗ Set of all possible campaign chains for order element ∈ E
i( p, t) Inventory level for product p ∈ P at time t ≥ 0
tf Earliest availability date of facility f ∈ F
Feasible solution
T Vector of earliest availability times (t f ) f ∈F
I Set of initial inventory levels {i( p, 0) | p ∈ P}
S Schedule
Algorithm 1
Bπ Formulation for product π ∈ P
lp Manufacturing level of product p ∈ P
la Manufacturing level of process a ∈ A
lmax Maximum manufacturing level of all processes a ∈ Aπ
F Set of all possible facility combinations for order ∈ E
Fi Set of i-th possible facility combination for order ∈ E
Algorithm 2
ip Algorithm parameter
i min
p Algorithm parameter
αn Algorithm parameter
na Algorithm parameter
MILP 1
x > (C i ) Selection indicator for campaign C i ∈ C
w Workload variable
Algorithm 3
wc,c Minimum delay times between campaigns c, c ∈ C
γk Algorithm parameter
MILP 2
ts , ts ≥ 0 Starting times of campaigns c, c ∈ ∪∈E C

tc , tc ≥ 0 Completion times of campaigns c, c ∈ ∪∈E C
α ≥ 0 Lateness of order ∈ E
xc,c ∈ {0, 1} Sequence indicator
m≥0 Timespan
hc(X ) Holding cost for products in X ⊆ P
tc() Lateness cost of demand element ∈ E
sca,a Cost for set up activity between processes a, a ∈ A
mc Timespan cost
T Duration of all campaigns in the schedule T = ∈E c∈C n · z a, f
1 Introduction
Supply chain management (SCM) has turned out to be not only a good lever for
cost cutting and cash generating, but also provides a competitive advantage within
different industries. This has made most companies focus on SCM and has resulted in
a higher orientation to the SCM core processes plan, source, make, deliver and return,
as described by Supply Chain Council (2004), which comprise strategic, tactical and
operational levels. Many disciplines help to improve opportunities for SCM excellence.
In particular, IT progress enabled the development of advanced planning systems
(APS) which enhanced transactional enterprise resource planning (ERP) systems and
replaced classical MRP II systems.
Different characteristics of process industry push production planning and sche-
duling into the forefront of tactical and operational supply chain planning resulting
in special requirements for APS and, often, in tailor-made models for decision sup-
port. The large variety of industry-specific characteristics in chemical production and
the huge number of combinations occurring in the real world result in many different
planning problems of N P-hard complexity. These can be solved by applying
either standardized optimization methods with relatively limited optimization results
or problem-specific algorithms involving a relatively high development, implementa-
tion and maintenance effort. The latter are differentiated between exact or deterministic
methods of mathematical optimization (incl. MILP, MINLP, graph theory, constraint
programming) and meta-heuristics (incl. evolutionary strategies, tabu search, simula-
ted annealing), both divided into off-line and on-line algorithms.
It is beyond the scope of this article to review all the literature on production plan-
ning and scheduling in a process industry; therefore, we focus on only a small selection
of publications. A thorough introduction to the main concepts of APS is given by Drexl
et al. (1994) or Stadtler and Kilger (2005). Günther (2005) gives a good overview of
the architecture and applications of APS, Tempelmeier (2001) or Meyr et al. (2005a)
analyze the structure of a typical APS. Meyr et al. (2005b) compares selected commer-
cial APS, details on SAP-based SCM are given by Knolmayer et al. (2002), Bartsch
and Bickenbach (2001) or Kallrath and Maindl (2006). Reklaitis (1996) or Kallrath
(2003a) give good introductions to planning and scheduling in the process industry,
Neumann et al. (2003) describe resulting requirements to APS. Different case studies,
e.g. Altrichter and Caillet (2005), Reuter (2005) or Richter and Stockrahm (2005),
show the application of SAP APO in the process industry. Grossmann (2005) shows a
new perspective on enterprise-wide optimization. An overview of different optimiza-
tion concepts for scheduling problems is given by Drexl and Kimms (1997), Kolisch
and Padman (2001), Shah (1998), Pinto and Grossmann (1998), Grossmann et al.
(2002) and Mendez et al. (2006). Complexity issues are considered by Monma and
Potts (1989) and Pekny and Reklaitis (1998).
In this article, we describe an optimization method for a real-world problem of
short- and mid-term production planning and scheduling arising from a specialty che-
micals plant. The optimization method is based on a time-continuous MILP model
formulation. Section 2 provides the problem statement and outlines the solution
approach, Section 3 explains the model formulation and the optimization method.
Section 4 contains computation results and implementation issues, while Sect. 5

details the conclusions.
2 Problem statement
2.1 Application environment
The problem considered in this article arises from a specialty chemicals plant pro-
ducing fragrances, flavors and aroma chemicals. The plant comprises a number of
non-identical multi-purpose single-activity facilities on which production processes
can be operated in non-pre-emptive batch mode. To produce a finished good, different
products, processes and facilities have to be considered, all depending on the formu-
lation for the finished product. Apart from refinement, the process that produces a
product is unique, and each process and each product can be assigned to a unique
formulation (“formulation encapsulation”). The number of processes and manufactu-
ring levels are formulation-dependent, and both differ greatly from one formulation to
another. The facilities cannot be assigned to a certain process, formulation or manu-
facturing level, and most processes can be performed on a choice of different facilities,
resulting in batch sizes and cycle times that depend on the process as well as on the
selected facility. Some processes have a recycling property, i.e. an input product is
not transformed completely but a fraction of this amount is discharged unchanged
when the process ends. Other processes run in joint production and release more than
one output product. Some of these joint products—so-called off spec products—can
be upgraded in a refinement process to achieve a product that fully meets the quality
requirements of the output of the main process. These characteristics result in all types
of material flows (linear, convergent, divergent, cyclic).
On the basis of the state task network (STN) representation proposed by Kondili
et al. (1993), an example of a typical formulation scheme is depicted in Fig. 1. A set
up activity has to be performed between two processes scheduled on the same facility
if and only if these two processes are not part of the same formulation. To reduce
these set up activities, several batches of identical processes are grouped together
to form (production) campaigns, i.e. they are operated on the same facility without
interruptions or idle times in between. Storage constraints for products do not have to
be considered for either shelf-life or storage capacity.
2.2 Formal description of the problem
The problem addressed here can roughly be stated as follows: Given

1. a set of production facilities with corresponding capacities,
2. a set of products that can be processed on these facilities, associated constraints,
production parameters, and bills of material (BOM),
3. a set of demand elements (both forecasts and customer orders), and
4. a set of penalties and cost functions.
Provide a feasible schedule which simultaneously ensures that
Fig. 1 Scheme of a typical formulation
1. enough material at each manufacturing level is produced to satisfy all primary and
secondary demands,
2. all production constraints, resource availabilities, and business requirements are
respected,
3. good solutions in a practical sense with regard to the multiple conflicting objectives
are obtained, and
4. identical batches are grouped to form campaigns.
2.2.1 Master data/problem parameters
Let F denote the set of facilities and P the set of products. Furthermore, let A be the
set of processes with subsets A f ⊆ A of processes that can be operated on a facility
f ∈ F and subsets Fa ⊆ F of facilities that can operate process a ∈ A. A process
a ∈ A operated on facility f ∈ F has a fixed cycle time z a, f and a fixed batch size
ba, f , both depending on the specific process and the chosen facility. Let a p ∈ A be
the unique process (apart from refinement) that produces product p ∈ P with input
and output products given by the subsets Pa− and Pa+ ⊆ P respectively in amounts
− and δ + respectively (fractions of batch size b
δa, p a, p a, f ). A formulation for finished
product p ∈ P is denoted by B p = (A p , Pp , F p ) with a set A p ⊆ A of processes
that have to be operated to produce p ∈ P, a set Pp = ∪ p∈A p Pa+ of output products
of the processes in A p and a set F p = ∪a∈A p Fa of facilities on which the processes
of A p can be operated. For each pair of processes a, a ∈ A, let sta,a denote the
duration of the set up activity required between a and a . Let R ⊆ A denote the set
of processes which have the refinement property, i.e. for a ∈ R there exist output
products p, p ∗ ∈ Pa+ \Pa− ( p ∗ off spec product) and a refinement process a ∗ ∈ A
with the properties p ∈ Pa+∗ \Pa−∗ and p ∗ ∈ Pa−∗ . Let R ∗ denote the set of all refinement
processes.
2.2.2 Transactional data/instance parameters
The demand elements are reflected by a set E = {(π , q , t ) | π ∈ P; q > 0; t > 0}

of order elements = (π , q , t ) requiring a quantity q > 0 of product π ∈ P
at due date t > 0. A campaign c = (ts , tc , n, v, a, f ) consists of n ∈ N batches of
process a ∈ A f operated without interruption or idle time on facility f ∈ Fa between
the starting time ts and the completion time tc and is sometimes preceded by a set up
activity (v = 1 ⇔ set up activity is performed before campaign c starts). For order
element = (π , q , t ) ∈ E, let C = {c = (ts , tc , n, v, a, f ) | ts > 0, tc > 0, n ∈
N, v ∈ {0, 1}, a ∈ Aπ , f ∈ Fa } be the set of all campaigns linked to the demand
element ∈ E, in the following such a set will be called a campaign chain. Due to the
fact that |Fa | > 1 for most of all a ∈ A p , there might be more than only one possible
campaign chain for a demand element ∈ E. Let C∗ denote the set of all possible
campaign chains for ∈ E. Let i( p, t) denote the inventory level for product p ∈ P
at time t ≥ 0 and let t f denote the earliest availability date for facility f ∈ F.
A campaign c ∈ S with processes a ∈ A is called formulation successor of a
campaign c ∈ S with processes a ∈ A if c consumes products that are produced by
c, i.e. if Pa− ∩ Pa+ = ∅. A campaign c ∈ S is defined as the schedule successor of a
campaign c ∈ S if c and c are scheduled on the same facility f ∈ F in such a way
that (i) ts ≤ ts and (ii) te∗ < ts or te < ts∗ for all other campaigns c∗ scheduled on f .
2.2.3 Properties of a feasible solution
Given an instance (E, T, I ) with a set E of order elements ∈ E, a vector T =

(t f ) f ∈F of earliest availability dates for the facilities f ∈ F and a set I = {i( p, 0) | p ∈
P} of initial inventory levels, a schedule S = {(ts , tc , n, v, a, f ) | ts > 0, tc > 0, n ∈
N, v ∈ {0, 1}, a ∈ A, f ∈ Fa } consisting of campaign elements c ∈ S is called
feasible if the following conditions hold:
1. Primary demands are satisfied regarding quantity produced allowing backlogging

for each product, i.e. for each π ⊆ , ∈ E:

+ −
i(π, 0) + δa,π · ba, f · n − δa,π · ba, f · n ≥ q
c∈S c∈S ∈E,π⊆
2. Inventory levels are always non-negative, i.e. for all p ∈ P, t ≥ 0:

i( p, t) ≥ 0
3. At each facility f ∈ F, at most one activity (process or set up) is scheduled at a
time, i.e. for each pair of campaigns c, c ∈ S scheduled on f ∈ F with ts ≤ ts :
tc ≤ ts − v · sta,a
4. Set up activities are scheduled if required, i.e. for each campaign c and its schedule
successor c scheduled on f ∈ F with ts ≤ ts :
v = 1 ⇔ {a, a } ⊆ A p ∀ p ∈ P (a and a do not belong to the same formulation)
5. The earliest availability date is respected for each facility, i.e. for each f ∈ F and
each c ∈ S scheduled on f :
ts − v · sta ,a ≥ t f ∀a ∈ A
Besides the mathematical aspect of solution quality, which can be easily measured by
an objective function, the acceptance of the solution by the supply chain planner in
the plant is a decisive quality criterion. Some key user requirements could be consi-
dered easily by constraints or specific properties of the solution approach, e.g. the
requirement to have at most one campaign for each reaction within a campaign chain.
Other requirements were realized by fine tuning the solution approach, e.g. defining
appropriate parameter settings for the objective function in accordance with the key
users.
2.3 Solution methodology
Numerous approaches to solve such problems exist, and these apply different mathe-
matical optimization methods and representations of the production process and the
planning horizon. Benchmark problems are described by e.g. Kondili et al. (1993),
Shah et al. (1993), Papageorgiou and Pantelides (1993) or Kallrath (2003a). A sur-
vey of time-continuous versus time-discrete approaches is given by Floudas and Lin
(2004). A process representation based on state task networks (STN) is introduced by
Kondili et al. (1993), Pantelides (1994) proposed the resource task network (RTN).
Applications of STN or RTN are evaluated by e.g. Shah et al. (1993), Schilling and
Pantelides (1996), Dimitriadis et al. (1998) and Giannelos and Georgiadis (2002).
MILP-based mathematical optimization concepts are reviewed by Floudas and Lin
(2005), applications are shown by e.g. by Blömer and Günther (1998, 2000), Burkard
et al. (1998a,b), Hui et al. (2000), Gupta and Karimi (2003), Yi and Reklaitis (2003)
and Burkard and Hatzl (2005, 2006). Kallrath (2003b) presents an MILP-based method
for combined strategic and operational planning, Karimi and McDonald (1997a,b)
apply MILP for integrated mid-term planning and short-term scheduling. Meyr (2004)
introduces a MILP-based approach for combined lot sizing and sequencing.
Decomposition techniques are introduced by Harjunkoski and Grossmann (2002),
Maravelias (2006) or Castro et al. (2005). Maravelias and Grossmann (2004a,b) or
Timpe (2003) present hybrid MIP/CP algorithms, Schulz et al. (1998) or Alle et al.
(2003) apply MINLP-based methods. Campaign planning and scheduling is discussed
in detail by Papageorgiou and Pantelides (1996a,b) or Oh and Karimi (2001a,b), in-
dustrial applications are described by Berning et al. (2003) applying genetic algorithm
techniques or by Grunow et al. (2003a,b) based on MILP formulations.
We will present a time-continuous MILP-based solution approach for campaign
scheduling. The basic idea is as follows:
1. Identical batches for the same primary or secondary demand are grouped to form
a campaign.
2. For each primary demand ∈ E, a campaign chain C , i.e. a set of campaigns, is

determined in such a way that
– C contains at most one campaign for each process,
– including initial inventory, the amounts produced by these campaigns are suf-
ficient to satisfy the primary demand q and all secondary demands arising
from , and
– no campaign of C feeds other campaigns that do not belong to C .
In the following, the property described by the last two bullet points will be called
order encapsulation of a campaign chain. The formulation encapsulation property
of the considered problem facilitates the generation of order specific campaign
chains. In the event that some of the processes of Aπ can be operated on a choice of
different facilities, different campaign chains are calculated independently from
each other, one chain for each possible facility combination. For each possible
chain, this calculation determines the number of batches of which each campaign
consists. The two possible campaign chains for the example formulation shown
above are depicted in Fig. 2 (demand quantity q = 20 t, no initial inventories
i( p, 0) = 0 for all p ∈ P).
3. For each primary demand ∈ E, exactly one set C of campaigns is selected to
satisfy . For each a ∈ Aπ this determines the facility f ∈ Fa on which a is
operated.
4. Starting times ts , completion times tc and set up activities v are determined for all
campaigns in ∪∈E C in such a way that the resulting schedule is feasible.
Steps 1 and 2 are realized by solving BOM for all orders and all possible facility
combinations; steps 3 and 4 are realized by solving MILPs. Complexity reduction is
the main reason for dividing the facility selection and the campaign scheduling into
two MILPs. Solving both problems simultaneously would significantly increase the
Fig. 2 Two possible campaign chains for the example formulation

number of variables and constraints or result in non-linearities requiring a completely

different solution approach. Furthermore, it would be even more difficult to apply the
solution method to real-world instances, which was already a challenging requirement.
3 Model formulation and solution approach
3.1 Campaign chain creation and selection
At first, all possible campaign chains C∗ for each demand element ∈ E are calcula-
ted. On the basis of the structure of each formulation, i.e. its products and processes and
their corresponding manufacturing levels, the sizes of all campaigns, i.e. the number
of batches of which each campaign consists, are calculated. This calculation is perfor-
med for all possible facility combinations taking into account the available inventory.
Having determined the sizes of all possible campaigns, just one set of campaigns is
selected for each demand element by solving MILP 1.
3.1.1 Campaign chain creation
Algorithm 1 is performed for each demand element = (π, q, te ) ∈ E in order to

determine the structure of the formulation Bπ = (Aπ , Pπ , ∪a∈Aπ Fa ) and the manu-
facturing levels l p and la for each product p ∈ Pπ respectively for each a ∈ Aπ :
Algorithm 1 (Determination of formulation structures)

Pπ = {π}, Aπ = {aπ }, P = {π}, lπ = 0
WHILE P = 0 DO
P = P \{ p}
Aπ = Aπ ∪ {a p }
IF a p ∈ R THEN
Aπ = Aπ ∪ {a ∗p }
Pπ = Pπ ∪ Pa+∗
p
FOR p ∈ Pa−p DO
P = P ∪ { p }
lp = lp + 1
Pπ = Pπ ∪ { p }
le = max{l p | p ∈ Pπ }
FOR a ∈ Aπ DO
la = max{l p | p ∈ Pa+ }
FOR k = 1 TO le DO
Ak = {a | a ∈ Aπ , la = k}
lmax = max{la | a ∈ Ak }
Having performed Algorithm 1 for each demand element ∈ E, the set F = {Fi | 1 ≤
i ≤ a∈Aπ |Fa |} of all possible facility combinations can now easily be determined by
the Cartesian product F = ×a∈Aπ Fa . Each of these facility combinations Fi results
in just one campaign chain C which can satisfy the primary demand and all of its
resulting secondary demands.
Algorithm 2 is performed for each ∈ E to determine the required sizes of these
campaigns. To keep them small, Algorithm 2 will ensure that possibilities for off spec
product refinement and consumption of available initial inventory are both used to the
maximum extent. For that reason, it is required to update the initial inventory i( p, 0)
after each iteration of Algorithm 2 on ∈ E. In the event that different chains with
different batch sizes are possible for the same demand element, this update will be
done with the minimum amount i min p of inventory that is left when scheduling one of
the possible chains.
A LGORITHM 2 (Calculation of campaign sizes)
FOR p ∈ Pπ DO
i min
p = i( p, 0)
FOR Fi ∈ F DO
FOR p ∈ Pπ DO
i p = i( p, 0)
qp = 0
qπ = q
FOR k = 0 TO lmax DO
FOR a ∈ Ak DO
FOR p ∈ Pa+ DO
IF i p < q p THEN
qp = qp − i p
ip = 0
ELSE
i p = i p − qp
qp = 0
n a = max{ δ + ·bp | p ∈ Pa+ , f a given by Fi }
q
a, p a, f a
n a∗ = 0
IF a ∈ R THEN
+ ·b
n a = min{n ∈ N | δa, p a, f a · n + αn · ba, f a ≥ q p }
+
n·δa, ·b
p ∗ a, f a
+i p∗
with αn = δa−∗ , p∗ ·ba ∗ , f ∗

a
+
q p −n a ·δa, p ·ba, fa
na∗ = δa+∗ , p ·ba ∗ , f ∗

a
+
i p∗ = i + δ p∗ ·b · n − δa−∗ , p∗ · ba ∗ , fa∗ · n a ∗
a, p ∗ a, f a a
FOR p ∈ Pa DO−
− − δ+ ) · b
q p = q p + (δa, +
p a, p a, f a · n a + δa, p · ba, f a
FOR p ∈ Pπ DO
i min
p = min{i min
p , i p}
FOR p ∈ Pπ DO
i( p, 0) = i min
p
3.1.2 Campaign chain selection
For each demand element, just one campaign chain fulfilling the primary demand as
well as all arising secondary demands has to be selected. The objective of this selec-
tion is to ensure balanced use of the available facilities instead of putting most of the
workload on only a few facilities. The selection is done by solving the MILP 1 that
minimizes the maximum facility-specific workload.
MILP 1 (Campaign Selection)
Minimize
Subject to
1. Just one campaign chain for each order element selected:

x(C i ) = 1 ∀ ∈ E
C i ∈C
2. Workload constraint considered:

n · z a, f · x(C i ) ≤ w ∀ f ∈ F
∈E C i ∈C c∈C i
Domains
w ≥ 0 Workload
x(C i ) ∈ {0, 1} Selection indicator.
Minimizing the workload has shown to be a reasonable objective for two reasons:
– A consideration of cost parameters in the MILP 1 is not necessary, because the

facilities have a comparable cost structure. In case that major differences in cost
structures occur, the constraints of type 2 could be weighted by a facility-specific
cost parameter.
– Balancing the workload allows to reduce the timespan of the campaign schedule
obtained by MILP 2. It furthermore allows a better synchronization of subsequent
campaigns resulting in less inventory and enables earlier order fulfilment which
helps reduce order tardiness.
After solving MILP 1 the selected facility combination Fi is determined for each
order ∈ E. Performing Algorithm 2 again for only the fixed facility combination Fi
ensures optimized use of available initial inventory because i min
p does not have to be
considered in this second algorithm run.
3.2 Campaign scheduling
Having fixed the sizes and facilities for each campaign, the last step is to determine
starting and completion times for each campaign in such a way that the resulting
schedule is feasible. One way to ensure inventory feasibility is to formulate mass
balances explicitly by hard constraints in the MILP-based step. To reduce the number
of constraints in MILP 2, we chose a different approach: for each demand element
∈ E, the chain C has the order encapsulation property. Due to this, the required
non-negativity of inventory levels is ensured implicitly if appropriate minimum delays
between a campaign c ∈ C and all of the formulation successors in C are respected,
inventory-based interdependencies between campaigns in different chains do not exist.
These delays are calculated using Algorithm 3 and then introduced, as well as capacity
limitations and set up requirements, into MILP 2 with hard constraints.
3.2.1 Calculation of minimum delay times
Algorithm 3 is performed for each demand element ∈ E and each campaign c ∈ C

to determine minimum delays between starting times of c and each of its formulation
successors in C . This calculation follows the assumption that, in the event that refi-
nement processes have to be considered in a formulation, these refinement processes
feed the last batches of their formulation successors.
Algorithm 3 (Minimum delay times)

Unless indicated otherwise, c and c indicate campaign objects (ts , tc , n, v, a, f ) and
(ts , tc , n , v , a , f ) respectively.
FOR c formulation successor of c DO
IF f = f THEN
wc,c = n · z a, f
ELSE
IFa ∈ R ∗ with n ∗ batches of main process a ∗ ∈ A on f ∗ ∈ F THEN
i p = i( p, 0) + δa+∗ , p · ba ∗ , f ∗ · n ∗
ELSE
i p = i( p, 0)
+ ·b +
(k−1)·δa, p a, f +i p −δa , p ·ba , f
γk = (δa− , p −δa+ , p )·ba , f

wc,c = max{z a, f · k − z a , f · min{γk , n } | 0 ≤ k < n}
IF γk < n THEN
IF wc,c < n · z a, f − γn+1 · z a , f THEN
wc,c = n · z a, f − γn+1 · z a, f
3.2.2 Campaign scheduling
MILP 2 determines starting and completion times ts and tc and also the campaign
sequence on each facility modeled by binary sequence indicators xc,c which require
virtual first and last dummy campaigns f d( f ) and ld( f ) on each facility f ∈ F.
Weighted by appropriate cost parameters, the objective function simultaneously mi-

nimizes inventory (implicitly by reducing delays between formulation predecessors
and successors), due date violations, set up activities and the timespan of the total
schedule. Different types of constraints ensure the feasibility of the schedule: Types 1
and 2 focus on timespan and lateness for the objective function, type 3 ensures that all
batches of a campaign are scheduled without idle time in a campaign. Types 4–6 en-
sure feasibility of the schedule with regard to delays, resource availabilities and set up
requirements. Types 7–10 ensure that the sequence indicators are set appropriately, i.e.
that each campaign has a unique schedule predecessor and a unique schedule successor.
MILP 2 (Campaign Scheduling)
Unless indicated otherwise, c and c indicate campaign objects (ts , tc , n, v, a, f ) and

(ts , tc , n , v , a , f ), respectively.
Minimize

hc(Pa+ ∩ Pa− ) · (ts − ts )
∈E c∈C c ∈C

+ tc() · α
∈E

+ sca,a · xc,c
∈E c∈C ∈E c ∈C
+mc · m
Subject to
1. Lateness of order ∈ E is at least as big as the difference between the completion
time of the last campaign of C and the due date for ∈ E:
tc − t ≤ α ∀ ∈ E, c ∈ C with a = aπ
2. Timespan m ends after completion of all campaigns:
tc ≤ m ∀ ∈ E, c ∈ C
3. No idle time within a campaign c ∈ C —completion time tc of campaign c equals

starting time ts of c plus duration of all n batches of c:
tc = ts + n · z a, f ∀ ∈ E, c ∈ C
4. Minimum delays wc,c between campaign c ∈ C and its formulation successor

c ∈ C have to be respected:
ts + wc,c ≤ ts ∀ ∈ E, c, c ∈ C with Pa+ ∩ Pa− = ∅

5. Earliest availability time has to be respected for each facility:
t f ≤ ts ∀ ∈ E, c ∈ C , f ∈ F
6. Campaigns c ∈ C and its schedule successor c ∈ C must respect set up

activities:
tc + sta,a − (1 − xc,c ) · T ≤ ts ∀ ∈ E, c ∈ C , ∈ E, c ∈ C with f = f
7. Each campaign c has just one schedule predecessor and just one schedule succes-
sor. Each first dummy campaign f d( f ) has just one schedule successor, each last
dummy campaign ld( f ) has just one schedule predecessor:
Predecessor

xc,c + x f d( f ),c = 1 ∀ ∈ E, c ∈ C
∈E c∈C

xc,ld( f ) + x f d( f ),ld( f ) = 1 ∀ f ∈ F
∈E c∈C
Successor

xc,c + xc,ld( f ) = 1 ∀ ∈ E, c ∈ C
∈E c ∈C

x f d( f ),c + x f d( f ),ld( f ) = 1 ∀ f ∈ F
∈E c ∈C
8. No campaign c is the schedule successor of itself:
xc,c = 0 ∀ ∈ E, c ∈ C
9. No schedule predecessor for the first dummy campaign f d( f ), no schedule suc-

cessor for the last dummy campaign ld( f ):
xc, f d( f ) = 0 ∀ ∈ E, c ∈ C , f ∈ F
xld( f ),c = 0 ∀ ∈ E, c ∈ C , f ∈ F
x f d( f ), f d( f ) = 0 ∀ f, f ∈ F
xld( f ), f d( f ) = 0 ∀ f, f ∈ F
xld( f ),ld( f ) = 0 ∀ f, f ∈ F
10. Schedule predecessor and schedule successor (incl. first or last dummy campaigns)
have to be operated on the same facility:
xc,c = 0 ∀ ∈ E, c ∈ C , ∈ E, c ∈ C with f = f
x f d( f ),c = 0 ∀ ∈ E, c ∈ C , f ∈ F with f = f
xc,ld( f ) = 0 ∀ ∈ E, c ∈ C , f ∈ F with f = f .
Domains
ts , ts ≥ 0 Starting times of campaigns c, c ∈ ∪∈E C
tc , tc ≥ 0 Completion times of campaigns c, c ∈ ∪∈E C
α ≥ 0 Lateness of order ∈ E
xc,c ∈ {0, 1} Sequence indicator of campaigns c, c ∈ ∪∈E C

and dummy campaigns c, c ∈ { f d( f ), ld( f ) | f ∈ F}
m ≥ 0 Timespan.
Parameters
hc(X ) Holding cost for products in X ⊆ P
tc() Lateness cost of demand element ∈ E
sca,a Cost for set up activity between processes a, a ∈ A
mc Timespan cost

T Duration of all campaigns in the schedule T = ∈E c∈C n · z a, f .
4 Realization and empirical results
4.1 Size of problem instances and computational limitations
Real-world instances of the planning problem considered are very large; some key
figures listed in Table 1 give an impression of their size. Although the solution method
described so far is straightforward, each MILP and each algorithm is calculated in one
step, it cannot be applied to such large instances without modifications.
Both MILP steps are complexity drivers, because their underlying problems Cam-
paign Selection and Campaign Scheduling are N P-hard (proofs base on reduction
to Partition rsp. Hamilton Path and are available from the authors on request). Real-
world instances result in not more than 200–300 binary variables for MILP 1 which
can therefore still be solved exactly within an acceptable time. On the contrary, com-
putation analysis has shown that a heuristic approach is required for MILP 2: Small
Table 1 Characteristics of real-world problem instances
Master data Transactional data
Min . Avg. Max. Min. Avg. Max.
|A| 154 |E| 85 92 100

|F| 26 Horizon 300 days
|P| 226 z a, f 8h 28.8 h 84 h
|A p | 1 2.5 6 sta,a 8h
|Fa | 1 1.4 4 |C∗ | 1 3.6 54
|A f | 1 8.7 26
|B| 62
instances were identified for which MILP 2, although comprising less than 1,000
binaries, could not be solved exactly in reasonable time. Real-world problem instances
result in 10,000–13,000 binary variables and therefore cannot be solved exactly.
4.2 Adaption of solution method
Due to the fact that the size of real-world instances does not allow MILP 2 to be solved
exactly, a solution heuristics for the campaign scheduling problem is required. Three
specific characteristics of the problem and of the model formulation enable a simple
but effective solution approach:
– The demand elements are distributed uniformly across the planning horizon. This
allows to partition the demand elements set E = E 1 ∪ · · · ∪ E n (E i ∩ E j = ∅) in
such a way that the order subsets are sorted by increasing due dates of their orders,
i.e. for ∈ E i , ∈ E j the following relationship holds: i < j ⇒ t ≤ t .
– Due to order encapsulation, the mapping of campaigns to demand elements is
unique. This allows to iteratively construct a schedule for all E i , because campaigns
that belong to a chain C , ∈ E i do not interact with campaigns from other orders
∈ E j , j = i.
– Finalizing the production for a demand element ∈ E significantly earlier does
not result in higher penalties.
Due to these characteristics, it is possible to solve MILP 2 iteratively without losing too
much solution quality: The schedule is determined by partitioning the set E of demand
elements to subsets E 1 , . . . , E n (E i ∩ E j = ∅) and then solving MILP 2 iteratively
for the respective subsets E j . During these iterations it is ensured that on each facility
the sequence of the scheduled campaigns follows the chronological sequence of the
partitioned demand subsets. For two campaigns c and c assigned to demand elements
∈ E i and ∈ E j and scheduled on facility f ∈ F, we have i < j ⇒ ts < ts . The
effects of this iterative approach of constructing an optimized schedule are outlined
in Fig. 3. After the last iteration, the whole MILP 2 is solved with fixed sequence
indicators xc,c , i.e. without changing the campaign sequence obtained before by the
iterations.
The described method is one possible way to solve the scheduling problem, but it is
not the only possibility for reducing the complexity of MILP 2. Two other approaches
based on rounding strategies for LP relaxations were examined: The strategy was to
solve the LP relaxation of MILP 2 iteratively and round just one sequence variable to 1
for each facility in each iteration. The variable to be rounded was selected either by the
lowest starting times of the campaigns or by the highest value of the binary sequence
indicators. The results from these approaches were not promising. Other solution
approaches, e.g. priority rule based top down scheduling, have not been evaluated
although they might lead to good results.
4.3 Computation analysis and results
The main target of the computation analysis was to ensure the applicability of the
optimization method to real-world instances of the problem and to customize it to
the specific requirements of the key users. A thorough numerical investigation by
comparing the proposed method with other optimization approaches based on this
specific problem or even other (academic) benchmarking problems has not been in
focus. Nevertheless, the efficiency of the proposed method has been evaluated and
empirical evidence of its applicability to real-world problem instances is given.
Within the computation analysis, 18 small instances comprising only a limited
number of demand elements and five large instances consisting of a number of demand
Fig. 3 Iterative approach for schedule construction

elements comparable to real-world requirements were evaluated. Characteristics of

the instances are shown in Table 2. The method was performed to all instances with
different parameter settings for the objective function of MILP 2—four settings each
focusing on only one of the optimization objectives (inventory reduction, due date
achievement, timespan minimization, set up optimization), and a fifth setting focusing
on all objectives in a balanced way.
An analysis of the results for the small problem instances achieved by solving MILP
2 exactly allows to estimate the number of iterations of MILP 2 required to solve real-
world problem instances: On the one hand, MILP 2 can be solved exactly for small
problem instances comprising up to 8 campaigns scheduled on the same facility, but
on the other hand, MILP 2 cannot be solved exactly for some instances with not more
than seven campaigns on the same facility. This observation indicates that MILP 2 can
be solved exactly if not more than 6–7 campaigns are scheduled on one facility, and
real-world instances comprise max. 20–30 campaigns scheduled on the same facility.
Therefore it should be sufficient to iterate the heuristics for MILP 2 only 4–5 times
with equally large partitions of the demand element set E to determine an optimized
Table 2 Characteristics of evaluated problem instances
Instance Inst. # Camp. |E| # Due |C| # Binaries MILP 2

size no. per fac. dates solved exactly
Small 1 5 14 12 31 207 Yes

2 5 14 12 32 185 Yes
3 6 15 11 33 236 Yes
4 6 15 13 33 279 Yes
5 6 21 17 54 490 Yes
6 6 23 15 55 824 Yes
7 7 15 13 39 540 Yes
8 7 20 16 48 629 Yes
9 7 23 14 48 607 Yes
10 8 14 11 24 206 Yes
11 7 24 17 57 952 No
12 7 23 17 55 824 No
13 7 24 13 57 952 No
14 7 31 19 64 1,000 No
15 7 22 15 61 1,095 No
16 9 15 17 39 540 No
17 10 18 14 47 825 No
18 14 26 17 64 1,733 No
Large 1 21 87 27 196 10,059 No
2 23 95 28 227 12,571 No
3 26 94 28 215 12,816 No
4 27 99 26 221 13,204 No
5 29 85 28 213 12,117 No
solution for real-world problem instances. Similar conclusions based on the number of
order elements instead of campaigns on one facility indicate 6–7 required iterations.
Tests on different large problem instances have shown that 4–6 iterations were always
sufficient to determine optimized schedules within 350–540 s. A total calculation time
of 30 min (incl. data transfer to the APS and visualization) was considered acceptable.
A cost comparison between the schedules that were generated manually by the
SC Planners and the ones that were calculated by the optimization method was not
performed. Nevertheless, all schedules obtained by optimization were highly accepted
by the key users who were enabled to reduce the work spent for routine planning and
to focus more strongly on planning exceptions. Furthermore, the APS provides a set
of standard algorithms for schedule construction and optimization. Different ways of
applying these algorithms were also tested, but in most cases it was not possible to
generate a feasible schedule of comparable quality within an acceptable time. Some
key figures characterizing size, properties and complexity of optimized schedules are
given in Table 3.
To evaluate the efficiency of the proposed method, the obtained calculation results
for both exact solution of MILP 2 and the iteration heuristics are analyzed and com-
pared with lower bounds obtained by the LP relaxation of MILP 2. For the timespan
variant of the problem, i.e. the parameter setting of the objective function which only
focuses on timespan minimization, a second lower bound is given by the maximum
workload calculated by MILP 1 within the campaign selection step. The results of
these comparisons are shown in Table 4.
The comparison with the optimal solutions of MILP 2 shows that for balanced
objectives the iteration heuristics has an optimality gap of 4% on average and 12%
in the worst case. Focusing only on set up optimization or timespan minimization
results in higher optimality gaps. Focusing only on inventory reduction results
in optimal solutions, which is easy to explain because scheduling the campaign with
optimal synchronization is simple if no other objectives have to be considered. An
optimal schedule for the inventory reduction variant of the problem can be obtained
as follows: start with an arbitrary campaign chain and schedule the campaigns of this
chain by decreasing manufacturing level in such a way that minimum delay times are
achieved. Perform a right shift of campaigns to obtain an optimal synchronization and
continue with the next chain until all campaigns are scheduled.
The comparison between the iterative approach and the LP relaxation of MILP
2 does neither confirm nor contradict the observations made so far. A comparison
Table 3 Characteristics of optimized schedules for real-world problem instances
Min. Avg. Max.
# of campaigns in total 196 214 227

# of campaigns on one facility 1 8.2 29
# batches in total approx. 1,800
# batches in one campaign 3 9 45
Timespan approx. 250 days
# free binaries 10,059 12,103 13,204
Table 4 Gaps of the iterative heuristics approach
Instance size Bound Gap Optimization objective
Balanced Timespan Inventory Due date Set up

(%) (%) (%) viol. (%) (%)
Small MILP 2 opt. Avg. 4 15 0 3 10

Max. 12 39 0 24 28
LP relax. Avg. 11 28 1 6 26
Max. 19 60 3 34 93
MILP 1 opt. Avg. − 31 − − −
Max. − 88 − − −
Large LP relax. Avg. 62 84 1 2, 640 48
Max. 217 116 2 4, 915 184
MILP 1 opt. Avg. − 50 − − −
Max. − 128 − − −
with optimal solutions for MILP 2 has shown that especially for the minimization of
timespan or due date violations the LP relaxation does not deliver a good lower bound.
This might explain why the approach of rounding based on LP relaxation does not
lead to good results. These effects can already be observed at the small instances and
they are stronger visible at the large instances.
Comparing the results of the iterative approach for the timespan problem with the
workload calculated by MILP 1 shows a gap of 31% on average and 88% in the worst
case. For the large instances, the gap is 50% on average and 128% in the worst case.
At a first glance, these observations do not seem to be promising, but a comparison
with the optimal solutions for MILP 2 has shown that the workload obtained by MILP
1 unfortunately is not a good bound: even the optimal solution has high gaps to the
workload bound—27% on average and 88% in the worst case for small problem
instances.
The facility workload obtained by MILP 1 or the solutions of the LP relaxation
of MILP 2 are not appropriate bounds to estimate the optimality gap of the iteration
heuristics. Nevertheless, a comparison of results obtained by the iterative approach
and optimal solutions of MILP 2 has shown that the partition approach does not
result in a large loss of optimality and solution quality. An explanation is that different
demand elements for the same finished product are often put into different partitions,
and therefore the chronological sequence of demand elements induces a chronological
sorting of the corresponding production campaigns. Furthermore, each partition still
comprises approximately 20–30 demand elements, and the resulting (smaller) decision
problem still has resource conflicts with a considerable optimization potential. Within
reasonable time trial runs with problem instances of real-world size resulted in solu-
tions of good quality—as shown especially by key user acceptance and comparison
to standard algorithms of the APS—and therefore give evidence of the applicability
of the solution approach in daily business.
4.4 Integration in APS
For a variety of reasons, it has to be possible to integrate the optimization method into
an APS to gain maximum benefit in business applications:
– APS offer possibilities for graphical depiction of the resulting production schedules
and give quick access to additional information on the schedule elements.
– APS facilitate manual modification of existing production schedules, especially
for management by exceptions.
– APS enable integration with ERP systems, other planning modules or process
control systems.
A more detailed description of benefits of APS is given by e.g. Kolisch et al. (2000).
Siletti and Petrides (2003) show the necessity for the interaction between batch process
scheduling and systems for production planning and process control. Experiences and
expectations of the application of SAP R/3 in process industry are described by e.g.
Schumann (1997).
The described solution method is implemented in C code and bases on a CPLEX
Optimizer for solving the MILP steps. This method is integrated into a standard APS
containing master and transactional data as well as additional parameters (e.g. cost
parameters for objective functions). All parameters can be exported quickly and easily
by Tcl routines via an open interface.
An example of such an optimized schedule depicting single batch process orders
of the scheduled campaign objects in an APS is shown in Fig. 4. The APS enabled
the key user to get detailed information on the elements and inventory levels of an
optimized schedule or to modify it by mouse clicks. The schedule itself has a timespan
Fig. 4 Gantt chart of an optimized schedule

of more than 8 months and shows high variances in resource occupation although a
balanced workload was targeted in MILP 1.
5 Conclusion
In this paper, we present a real-world problem of campaign scheduling for a specialty

chemicals plant. The plant comprises multi-mode production facilities to operate batch
processes with joint production, refinement steps, sequence dependencies and other
industry-specific characteristics. Multi-level production formulations include all types
of material flows (incl. recycling) across more than 200 products. We have described
a solution methodology on an MILP-based time-continuous model formulation to
solve problem instances of real-world size. The feasibility of this approach is proved
empirically, the results obtained exhibit reasonable computation time and solution
quality. The possibility of integration into APS is outlined and demonstrated using a
standard tool.
The paper leaves room for further research activities, for instance, applying other
solution techniques to the considered problem. The possibilities of embedding all
decisions in one mathematical model and the analysis of results achieved by such
approaches in comparison to the solutions obtained by the method described in this
paper are of special interest. The adaptation of the described solution approach to
other industry specific problem characteristics—frequency-dependent set ups, cyclic
downtimes of production facilities, variable batch sizes or storage restrictions just to
name a few—could be relevant for the industrial practice. Such an analysis could be
based on different well-known benchmark problems which have already been used
for evaluations of other solution approaches.
Acknowledgment The authors would like to thank Hans-Heinrich Böther, Heinrich Schuchard and the
referees for their constructive comments and suggestions that contributed to the improvement of this paper.
References
Alle A, Papageorgiou LG, Pinto JM (2003) Economic lot scheduling under performance decay. In: Gross-
mann IE, McDonald CM (eds) Proceedings of FOCAPO 2003, pp 307–310, CACHE
Altrichter F, Caillet T (2005) SCM in a pharmaceutical company. In: Stadtler H, Kilger C (eds) Supply
chain management and advanced planning—concepts, models, software and case studies. Springer,
Berlin Heidelberg New York, pp 355–370
Bartsch H, Bickenbach P (2001) Supply chain management mit SAP APO. Galileo Press, Bonn (in German)
Berning G, Brandenburg M, Gürsoy K, Mehta V, Tölle FJ (2003) An integrated system solution for supply
chain optimization in the chemical process industry. In: Günther HO, van Beeck P (eds) Advan-
ced planning and scheduling solutions in process industry. Springer, Berlin Heidelberg New York,
pp 229–260
Blömer F, Günther HO (1998) Scheduling of a multi-product batch process in the chemical industry. Comput
Ind 36:245–259
Blömer F, Günther HO (2000) LP-based heuristics for scheduling chemical batch processes. International
J Prod Res 38:1029–1051
Burkard RE, Hatzl J (2005) Review, extensions and computational comparison of MILP formulations for
scheduling of batch processes. Comput Chem Eng 29:1752–1769
Burkard RE, Hatzl J (2006) A complex time based construction heuristics for batch scheduling problems
in the chemical industry. Eur J Oper Res 174:1162–1183
Burkard RE, Hujter M, Klinz B, Rudolf R, Wenning M (1998a) A process scheduling problem arising from
chemical production planning. Optim Methods Softw 10:175–196
Burkard RE, Kocher M, Rudolf R (1998b) Rounding strategies for mixed integer programs arising from
chemical production planning. Yugoslav J Oper Res 8:9–23
Castro P, Mendez C, Grossmann IE, Harjunkoski I, Fahl M (2005) Efficient MILP-based solution strategies
for large-scale industrial batch scheduling problems. In: Marquard W, Pantelides C (eds) In: 16th
European symposium on computer aided process engineering and 9th international symposium on
process systems engineering, pp 2231–2236
Dimitriadis AD, Shah N, Pantelides CC (1998) Efficient modelling of partial resource equivalence in
Resource-Task Networks. Comput Chem Eng 22:563–570
Drexl A, Kimms A (1997) Lotsizing and scheduling: survey and extensions. Eur J Oper Res 99:221–235
Drexl A, Fleischmann B, Günther HO, Stadtler H, Tempelmeier H (1994) Konzeptionelle Grundlagen
kapazitätsorientierter PPS-Systeme. Z Betriebswirtschaftliche Forsch 46:1022–1045 (in German)
Floudas CA, Lin X (2004) Continuous-time versus discrete-time approaches for scheduling of chemical
processes: a review. Comput Chem Eng 28:2109–2129
Floudas CA, Lin X (2005) Mixed integer linear programming in process scheduling: modeling, algorithms,
and approaches. Ann Oper Res 139:131–162
Giannelos NF, Georgiadis MC (2002) A simple new continuous-time formulation for short-term scheduling
of multipurpose batch processes. Ind Eng Chem Res 41:2178–2184
Grossmann I (2005) Enterprise-wide optimization: a new frontier in process systems engineering. AIChE
J 51:1846–1857
Grossmann I, Heever SA, Harjunkoski I (2002) Discrete optimization methods and their role in the inte-
gration of planning and scheduling. AIChE Symp 98(326):150–168
Grunow M, Günther HO, Lehmann M (2003a) Campaign planning for multi-stage batch processes in the
chemical industry. In: Günther HO, van Beeck P (eds) Advanced planning and scheduling solutions
in process industry. Springer, Berlin Heidelberg New York, pp 73–106
Grunow M, Günther HO, Yang G (2003b) Plant co-ordination in pharmaceutics supply networks. In:
Günther HO, van Beeck P (eds) Advanced planning and scheduling solutions in process industry
Springer, Berlin Heidelberg New York, pp 261–294
Günther HO (2005) Supply chain management and advanced planning systems. In: Günther HO, Mattfeld
D, Suhl L (eds) Supply chain management und logistik: optimierung, simulation, decision support.
Physica, Heidelberg, pp 3–40
Gupta S, Karimi IA (2003) An improved MILP formulation for scheduling multiproduct. multistage batch
plants. Ind Eng Chem Res 42:2365–2380
Harjunkoski, Grossmann IE (2002) Decomposition techniques for multistage production scheduling pro-
blems using mixed-integer and constraint programming methods. Comput Chem Eng 26:1533–1552
Hui CW, Avaneesh G, Meulen HAJ (2000) A novel MILP formulation for short-term scheduling of
multi-stage multi-product batch plants with sequence-dependent constraints. Comput Chem Eng 24:
2705–2717
Kallrath J (2003a) Planning and scheduling in the process industry. In: Günther HO, van Beeck P (eds)
Advanced planning and scheduling solutions in process industry. Springer, Berlin Heidelberg New
York, pp 11–42
Kallrath J (2003b) Combined strategic and operational planning—a MILP success story in chemical industry.
In: Günther HO, van Beeck P (eds) Advanced planning and scheduling solutions in process industry.
Kallrath J, Maindl TI (2006) Real optimization with SAP APO. Springer, Berlin Heidelberg New York
Karimi IA, McDonald CM (1997a) Planning and scheduling of parallel semicontinuous processes. 1. Pro-
duction planning. Ind Eng Chem Res 36:2691–2700
Karimi IA, McDonald CM (1997b) Planning and scheduling of parallel semicontinuous processes. 2. Short-
term scheduling. Ind Eng Chem Res 36:2701–2714
Knolmayer G, Mertens P, Zeier A (2002) Supply chain management based on SAP systems. Springer, Berlin
Heidelberg New York
Kolisch R, Brandenburg M, Krüger C (2000) OR-Software: Numetrix/3 Production Scheduling. OR Spec-
trum 22:307–312
Kolisch R, Padman R (2001) An integrated survey of project scheduling—models. algorithms, problems,

and applications. Omega 29:249–272
Kondili E, Pantelides CC, Sargent RWH (1993) A general algorithm of short-term scheduling of batch
operations—I. MILP formulation. Comput Chem Eng 17:211–227
Maravelias CT (2006) A decomposition framework for the scheduling of single- and multi-stage processes.
Comput Chem Eng 30:407–420
Maravelias CT, Grossmann IE (2004a) A hybrid MILP/CP decomposition approach for the continuous time
scheduling of multipurpose batch plants. Comput Chem Eng 28:1921–1949
Maravelias CT, Grossmann IE (2004b) Using MILP and CP for the scheduling of batch chemical processes.
Lect Notes Comput Sci 3011:1–20
Mendez CA, Cerda J, Grossmann I, Harjunkoski I, Fahl M (2006) State-of-the-art review of optimization
methods for short-term scheduling of batch processes. Comput Chem Eng 30:913–946
Meyr H (2004) Simultane Losgrößen—und Reihenfolgeplanung bei mehrstufiger kontinuierlicher Ferti-
gung. Zeitsch Betriebswirtschaft 74(6):585–610 (in German)
Meyr H, Rohde J, Wagner M, Wetterauer U (2005a) Architecture of selected APS. In: Stadtler H, Kilger C
(eds) Supply chain management and advanced planning: concepts, models, software and case studies.
Meyr H, Wagner M, Rohde J (2005b) Structure of Advanced Planning Systems. In: Stadtler H, Kilger
C (eds) Supply Chain Management and Advanced Planning: concepts, models, software and case
studies. Springer, Berlin Heidelberg New York, pp 109–115
Monma CL, Potts CN (1989) On the complexity of scheduling with batch setup times. Oper Res 37:798–804
Neumann K, Schwindt C, Trautmann N (2003) Advanced production scheduling for batch plants in process
industries. In: Günther HO, van Beeck P (eds) Advanced planning and scheduling solutions in process
industry. Springer, Berlin Heidelberg New York, pp 43–72
Oh HO, Karimi IA (2001a) Planning production on a single processor with sequence-dependent setups—
Part I: determination of campaigns. Comput Chem Eng 25:1021–1030
Oh HO, Karimi IA (2001b) Planning production on a single processor with sequence-dependent setups—
Part II: campaign sequencing and scheduling. Comput Chem Eng 25:1031–1043
Pantelides CC (1994) Unified frameworks for optimal process planning and scheduling. In: Rippin DWT,
Hale JC, Davis J (eds) Proceedings of FOCAPO 1994, pp 253–274, CACHE
Papageorgiou LG, Pantelides CC (1993) A hierarchical approach for campaign planning of multipurpose
batch plants. Comput Chem Eng 17:27–32
Papageorgiou LG, Pantelides CC (1996a) Optimal campaign planning/scheduling of muiltipurpose
batch/semiconductors plants—1. Mathematical formulation. Ind Eng Chem Res 35:488–509
Papageorgiou LG, Pantelides CC (1996b) Optimal campaign planning/scheduling of muiltipurpose
batch/semiconductors plants—2. A mathematical decomposition approach. Ind Eng Chem Res
35:510–529
Pekny JF, Reklaitis GV (1998) Towards the convergence of theory and practice: A technology guide for
scheduling/planning methodology. In: Pekny JF, Blau GE (eds) Proceedings of FOCAPO 1998, pp
91–111, CACHE
Pinto JM, Grossmann IE (1998) Assignment and sequencing models for the scheduling of process systems.
Ann Oper Res 81:433–466
Reklaitis GV (1996) Overview of scheduling and planning of batch process operations. In: Reklaitis GV,
Sunol AK, Rippin DWT, Hortascu O (eds) Batch processing system engineering. Springer, Berlin
Heidelberg New York, pp 660–705
Reuter B (2005) Demand planning of styrene plastics. In: Stadtler H, Kilger C (eds) Supply chain ma-
nagement and advanced planning—concepts, models, software and case studies. Springer, Berlin
Heidelberg New York, pp 407–420
Richter M, Stockrahm V (2005) Scheduling of synthetic granulate. In: Stadtler H, Kilger C (eds) Supply
Chain Management and Advanced Planning—concepts, models, software and case studies. Springer,
Berlin Heidelberg New York, pp 435–452
Schilling G, Pantelides CC (1996) A simple continuous-time process scheduling formulation and a novel
solution algorithm. Comput Chem Eng 20:S1221–S1226
Schulz C, Engell S, Rudolf R (1998) Scheduling of a multi-product polymer batch plant. In: Pekny JF, Blau
GE (eds) Proceedings of FOCAPO 1998, pp 224–230, CACHE
Schumann A (1997) SAP R/3 in process industries: expectations, experiences and outlooks. ISA Trans
36(3):161–166
Shah N (1998) Single- and multisite planning and scheduling: current status and future challenges. In:
Pekny JF, Blau GE (eds) Proceedings of FOCAPO 1998, pp 91–111, CACHE
Shah N, Pantelides CC, Sargent RWH (1993) A general algorithm of short-term scheduling of batch ope-
rations - II. Computational issues. Comput Chem Eng 17:229–244
Siletti CA, Petrides D (2003) Overcoming the barriers to batch process scheduling. In: Grossmann IE,
McDonald CM (eds) Proceedings of FOCAPO 2003, pp 375–379, CACHE
Stadtler H, Kilger C (eds) (2005) Supply chain management and advanced planning,—concepts, models,
software and case studies. Springer, Berlin Heidelberg New York
Supply Chain Council (2004) Supply-chain operations reference-model SCOR Version 6.1. http://www.
supply-chain.org
Tempelmeier H (2001) Supply chain planning with advanced planning systems. 3rd Aegean International
Conference on “Design and Analysis of Manufacturing Systems”, Timos Island, Greece, May 19–22,
2001
Timpe C (2003) Solving planning and scheduling problems with combined integer and constraint program-
ming. In: Günther HO, van Beeck P (eds) Advanced planning and scheduling solutions in process
industry, pp 107–124. Springer, Berlin Heidelberg New York
Yi G, Reklaitis GV (2003) Optimal scheduling of multi-stage polymer plants. In: Grossmann IE, McDonald
CM (eds) Proceedings of FOCAPO 2003, pp 331–334, CACHE
Supply chain planning
in the German automotive industry∗
Herbert Meyr

OR Spectrum (2004) 26:447–470
DOI 10.1007/s00291-004-0168-4
Abstract Following the evolution in the computer industry, quite a lot of car man-
ufacturers currently intend to move from a built-to-stock oriented production of
standardized cars towards a customized built-to-order (BTO) production. In the pre-
mium segment of Germany’s automotive industry, the share of customized BTO cars
traditionally is comparatively high. Nevertheless, German car manufacturers have
spent a lot of efforts in recent years to further increase this share in order to realize
short delivery times, high delivery reliability and a fast responsiveness.
Surprisingly, comprehensive overviews of the short- and mid-term planning land-
scape of car manufacturers cannot be found in the scientific literature. Thus, the first
part of the paper discusses supply chain planning, as traditionally established in
the premium segment of the German automotive industry, and reviews methods of
Operations Research (OR) that are able to support the various planning tasks involved.
In the second part, the major change in strategy, currently to be observed in the Ger-
man automotive industry, is briefly summarized in order to derive its impacts for the
planning system and for the respective planning methods. In this way, challenges for
a future application of OR methods in the automotive industry can be identified.
Keywords Supply chain planning · Operations research · Automotive industry
1 Introduction
Mass customization [42] that aims at offering customized products in a high variety
but for still low prices and within short delivery times gains increasing importance
in various branches of business and, in the meantime, also captivates the automotive
industry. The BMW Group, for example, spent $55 million on its new Euro-
pean online-ordering system [24] to cut order-to-delivery times by 20 days on the
H. Meyr (B)
Department of Production and Supply Chain Management, Technical University of Darmstadt,
∗ This article was originally published in OR Spectrum 26/4 (2004), pp. 447–470, and reflects the
situation in the automotive industry during the years prior to this publication date.

344 H. Meyr
average. At the same time, BMW offers up to 1032 variants (at least theoretically),
several thousands of them actually being demanded [51, p. 42]. Other manufacturers
also declared their intention to decrease order-to-delivery times from an average of
40 days to about 15 days [22] and try to make the transition from “build-to-stock“
(BTS) to “build-to-order” (BTO) that has successfully been demonstrated by the
computer industry, and first and foremost by its paragon Dell.
The transition to BTO in the computer industry caused a reorganization of plan-
ning processes and led to an increased use of “Advanced Planning Systems” (APS,
[29]), i.e. of computer-based decision support systems, which – at least partly – rely
on sophisticated methods of Operations Research (OR). Thus the questions arise,
whether and how the transition of the automotive industry changes their respective
planning tasks and planning processes, and to what extent planning and OR meth-
ods are and will be affected. Since mutual interrelations are particularly important
for operational planning tasks, the discussion will concentrate on mid- and short-
term supply chain planning, and here especially focusing on the car manufacturers’
point of view. But before discussing changes it has to be shown what the planning
landscape of automotive industries traditionally looks like. There are, of course,
discussions of various individual planning tasks (see Sect. 3) and some overviews
of the order-to-delivery process (see e.g. [23, 51]). However, to the author’s knowl-
edge, in scientific literature no comprehensive overviews of the short- and mid-term
planning landscape of car manufacturers can be found.
Due to this lack of literature and since common scientific approaches like ques-
tionnaires and structured interviews did not seem to be very promising because quite
a lot of confidence is needed to get such a sensitive information, the following char-
acterization of the planning system of car manufacturers mainly builds on various
joint projects with German car manufacturers and communication with their respon-
sible planners and with employees of automotive consultancies. In order to verify
the conclusions drawn, a working paper has been written, sent to skilled people in
these companies and they have been asked for statements about its validity. The re-
sults of this process are presented in the following. To sum up, the contribution of
this paper is
– first, that the planning systems of German car manufacturers are analyzed, de-
scribed and thus made available to the academic literature,
– secondly, that OR methods suitable for planning within the automotive industries
are reviewed, categorized with respect to the planning tasks of (German) car
manufacturers and that insufficiently supported planning tasks are disclosed, and
– thirdly, that the challenges of the managerial changes from BTS to BTO are
outlined that arise for the planning tasks, the planning systems and for the OR
models/methods involved.
Due to this broad scope of the paper, a review of OR methods – even though
restricted to short- and mid-term planning – cannot be comprehensive. This paper
rather intends to give an idea where (i.e. at which subsection within the overall
planning system of a car manufacturer) OR methods already contribute or may con-
tribute in the future.
Supply chain planning in the German automotive industry 345
Long-term, strategic planning provides potentials, which mid-term planning has

to further develop and short-term planning has to implement. Of course, also long-
term planning tasks are supported by OR methods. Concerning the product design,
for example, the optimal commonality of automotive components (e.g. wire har-
nesses) is determined [53] or the impact of product variety on the performance of
mixed model assembly lines is analyzed [6, 16]. It is even worth to include assem-
bly sequencing issues into product design decisions [52]. Analytical and simulation
models provide general hints (“chaining strategies’’) how to assign products to man-
ufacturing plants so that high process flexibility is achieved for both single stage
[28] and multi stage [5, 19] automotive supply chains. Linear Programming (LP) or
Mixed-Integer linear Programming (MIP) models are, for instance, used by APS to
design the inbound system of assembly plants [20] or car distribution networks [2]
(see also [36] without use of APS). Concerning the inside of assembly plants, the
planning of the physical layout and of buffer sizes of assembly shops, in general,
and of body shops [41, 49, p. 20 ff. and 73 ff.], in particular, can be supported by
simulative, analytical and combinatorial optimization methods. A comprehensive
overview of OR methods for the well-known assembly line balancing, which is a
rather strategic than mid-term task in the automotive industry, is given by [47]. A
recent survey of heuristic methods for cost-oriented assembly line balancing can be
found in [1].
In order to understand why automotive planning systems are organized the way
they are, Sect. 2 describes the characteristics of automotive supply chains. These
vary substantially for car manufacturers in different parts of the world (North
America, Japan/Korea, Europe/Germany), operating on different market segments.
This paper mainly concentrates on premium brands (like BMW, Mercedes, Audi)
but not luxury cars (like Rolls-Royce, Maybach, Bentley) and on the German au-
tomotive industry. Nevertheless, quite a lot of the statements and findings of this
paper can be transferred to car manufacturers in other parts of the world tackling
related product segments because (even though beginning with different starting
points) many of them similarly intend to change to a BTO production. Section 3 then
presents the traditional short- and mid-term planning system and – after introduc-
ing the respective planning tasks – points to appropriate planning and OR methods.
After briefly summarizing the measures to improve BTO assembly currently being
implemented in the German automotive industry (Sect. 4), their impact on the plan-
ning system is discussed in Sect. 5. Thus changed requirements for planning
methods can be derived and challenges for future research can finally be identified
(Sect. 6).
2 Automotive supply chains
Cars are sold to final customers either directly via sales subsidiaries of the car man-
ufacturer or indirectly via legally separate retailers. The bill-of-material (BOM) is
strictly convergent, i.e. assembly processes are dominant. Cars often are thought to
be standard products. However, in the premium segment of this line of business,
there is a high degree of customization. This allows the customer to specify oblig-
atory features like the color of the car and type of upholstery or optional features
346 H. Meyr
like air conditioning or a navigation system, to name only a few. In the follow-
ing both obligatory and optional features are just referred to as “options”. A car
manufacturer usually offers several types of cars (e.g. the E-class or C-class of
DaimlerChrysler), which again differ in several body-in-white variants (coupe, con-
vertibles, etc.). Not every customer needs his car immediately. According to [51,
p. 38] the order lead time desired by a final customer is normally distributed with a
mean value of 4–6 weeks.
The sales organization and distribution network of a car manufacturer have a di-
vergent structure, which comprises several stages like the central sales department
of the manufacturer, sales persons responsible for different world regions (also at the
headquarter), sales companies in different countries or local areas and a rather high
number of further retailers and sales subsidiaries. This type of customized premium
cars can only be assembled “to order”, i.e. there has to be an “order” available –
either by the final customer, a retailer or a sales department of the manufacturer –
that specifies the options of the car. Current SCM initiatives in the automotive
industry try to increase the share of final customers’ orders and to decrease the share
of retailers’ and sales departments’ orders (see Sect. 4.1).
Commonly, manufacturer and retailers communicate in two types of interaction
rounds: In the first one, a retailer sends his mid-term requests for cars to the manu-
facturer. Both “negotiate” the number of cars (so-called “quota”) the retailer will get
during the next year. Usually, this “negotiation” process is clearly dominated by the
manufacturer so that – due to the preferences of the manufacturer – the agreed quota
may be less or even higher than the original requests. Since these quotas are, for ex-
ample, defined for the next year on a monthly basis, only body-in-white variants
and the type of engines (referred to as “models” in the following) are considered,
but the options are not specified at this point in time.
In a second round, about three to five weeks before planned production, the re-
tailer has to specify the options for all cars of his quota, which are due and have
not been assigned to final customer orders that had arrived in the meantime. From
a retailer’s point of view, these cars are “built to stock” (BTS-cars), based on a sort
of forecasting process for options. From the manufacturer’s point of view, an order
of the retailer exists, thus justifying the term “built to order”.
Figure 1 illustrates the different states of demand information that are implied
by these two interaction rounds. The curve (I) shows the cumulated share of fully
specified orders of final customers with respect to the overall number of orders
(incl. forecasts) considered by planning. It can be computed by calculating the distri-
bution function of the order lead times, that are desired by final customers (see [51,
p. 38]). This distribution function is drawn backward in time, starting with the de-
livery of cars to final customers.
For the section above the curve, no information about the preferences of final
customers is available. This lack of information has to be replaced with forecasts.
Concerning the options of BTS-cars, this is done by retailers with a lead time of
3–5 weeks before production (II). Beforehand, with a lead time of one year at a
maximum, only the retailers’ requests for models are known (III), which also are
the result of a forecasting process of the retailer. For yet earlier planning tasks, a car
manufacturer has to rely on his own pure forecasts for models (IV).
share of
final customers‘ orders
100 %
pure requests options

forecast for models specified
for models by retailers by
(IV) (III) retailers
(II) 50 %
options
„built-to-stock“ specified by
of retailers final customer
(I)
time
sending requests specification of options delivery to

for models by retailers final customer
Fig. 1 Demand information available to a car manufacturer
The production system in a car assembly plant usually comprises the four stages
pressing of metal or aluminium sheets, welding the body-in-white from the moulded
sheets in the body shop, painting it in the paint shop and final assembly, where
painted body, engine, transmission and the further equipment are brought together
or built in. For the final assembly one or several production lines are used. A pro-
duction line consists of quite a lot of serially arranged assembly stations, between
which cars are conveyed with a fixed belt rate. The processing time at an assembly
station depends on the option chosen for the car to be assembled. Therefore, the
overall utilization of a station is determined by the sequence in which cars/orders
are assembled on a line (the so-called “model mix”). If too many cars requiring the
same options are following one another, some of the stations may be overloaded
whereas others are underloaded. Thus a “balanced” model mix has to be found,
almost equally utilizing the various stations of an assembly line.
Because of the convergent BOM and ten thousands of components to be pur-
chased, a procurement network with several hundreds direct and an enormous
number of indirect suppliers has to be coordinated. For the delivery of incoming
goods normally several transport modes are applied. Voluminous and expensive
components are – as far as possible – delivered “just in time” (JIT) at the day of as-
sembly, partly even directly to the assembly line and thus arranged in the sequence
of planned assembly (“sequence-in-line supply”, SILS). The remaining incoming
goods are collected by regional carriers, consolidated and brought to intermediate
warehouses of the car manufacturers, which are close to their assembly sites.
The structure of an automotive supply chain is characterized by a convergent
flow of material upstream of the assembly plants of the car manufacturer and a
divergent flow of finished cars downstream. An automotive SC is difficult to co-
ordinate, because not only production capacity and manpower may turn out to be
bottlenecks, but also incoming goods. For reasons of flexibility, high-volume car
models can sometimes be produced at several assembly plants. Because of these
constraints, the promised delivery dates cannot always satisfy the expectations of
the final customers. Furthermore, final customers need a reliable delivery date because
348 H. Meyr
important further activities (like selling the old car, making money available) have
to be synchronized with the arrival of the new car. Thus, “order promising” not
only has to aim at setting a delivery date close to the customer’s wishes, but also at
promising a reliable date considering as many of the above constraints as possible.
Besides a significant intra-organizational information flow between different
planning units/departments of the car manufacturer itself (as will be discussed in
Sect. 3), there is also a vital inter-organizational exchange of information between
the different members of the SC. Commonly, car manufacturers prepare a rough
mid-term supply plan of the next year for their (first-tier) suppliers in order to draw
early attention to potential capacity bottlenecks. In the short term, daily supply plans
are sent to the suppliers. These include binding orders for the next day, but also quite
reliable “forecasts” for the next days/weeks and even rough forecasts for the next
months.
3 Traditional planning processes
To cope with the various planning tasks of automotive supply chains, quite a lot of
planning units/departments have to be involved. These planning tasks and the re-
spective decisions can be assigned to several planning levels (e.g. strategic, tactical,
operational) comprising different planning horizons (e.g. long-, mid-, short-term).
Depending on the planning horizon and the lead time necessary to make a certain
decision, different phases of the time axis of Fig. 1 are relevant and thus a differ-
ent state of knowledge about actual customer demand is available. Therefore, from
a manufacturer’s point of view, one may distinguish between forecast-driven long-
and mid-term planning (phases (III) and (IV) of Fig. 1) and order-driven short-term
planning (phases (I) and (II)). In Sect. 3.1 forecast-driven mid-term planning tasks
and their information flows, which are more or less common for the German au-
tomotive industry, will be discussed first. Order-driven planning will then be the
concern of Sect. 3.2. Within both sections organizational issues are left aside. This
will be covered in Sect. 3.3.
3.1 Forecast-driven planning
Figure 2 summarizes the forecast-driven planning activities. Planning tasks are

marked by rectangles, arcs illustrate the information flows in between. From the
bottom to the top, the level of aggregation and the planning horizon are increasing,
the frequency of planning is decreasing, however. The planning tasks are roughly
assigned to the logistical functions procurement, production, distribution and sales,
again. Of course, not all of the mid-term planning tasks of a car manufacturer will
be discussed. Only the most important ones which show a close interrelation have
been selected.
The annual budget planning determines the overall monetary budgets of the car
manufacturer’s departments and assembly plants for the next year. For this, produc-
tion plans for the respective plants and the sales plans for the respective sales regions
have to be calculated, too. This is done once per year, for the next year, by deciding
procurement production distribution sales
budget planning
take rates
pro-
sales
(monthly)
duction aggregate
quotas
volume goals, (monthly)
earning goals, production
forecasts, demand
production annual & sales
MRP take rates planning
plans working time plans
master production pl. requests of
regions
supply pro- allocation
sales
plans duction
planning
(weekly)
(weekly) aggregate quotas
production
requests
plans, detailed quotas for models
overtime
plants
suppliers retailers
Fig. 2 Overview of (mainly) forecast-driven planning
about production and sales quantities of car models (per plant and world region,
for example) on a monthly basis. The overall yearly quantities can be considered
as “volume goals” of the next year for both sales and production. From these, the
expected production costs and earnings can be derived (“earning goals”).
A further result of the annual budget planning is the usage or reservation of addi-
tional capacities, as far as these can still be influenced on a mid-term basis. Because
of the long lead times (e.g. two years or more to install an assembly line or a plant),
usually capacities of production resources are adapted to customer demand in the
long term and thus are a concern of strategic planning. However, agreements about
the extent and flexibility of the yearly working time, for example, are also a task of
mid-term planning. A lot of further constraints have to be respected like potential
bottlenecks of suppliers, model mix restrictions (capacities of crucial options, min-
imum utilization) and upper or lower bounds of the sales in certain markets. Lower
bounds, for example, result from strategic directives about the presence in impor-
tant markets, upper bounds may be due to marketing analyses about final customer
demand.
Input data for the annual budget planning mainly are forecasts for final cus-
tomers’ demand (see also Fig. 1), which result from the demand planning. These
are made on basis of historical sales data, of the few already known and fully spec-
ified orders from final customers (e.g. car rentals), of the retailers’ annual requests
for models, of the sales companies’ decentral knowledge about the local prefer-
ences of their customers (“requests of regions”) and on basis of information about
marketing capabilities to influence final customer demand.
Since budget planning has to decide about car models on the one hand and to an-
ticipate potential bottlenecks of suppliers on the other hand, the component demand
needs to be estimated, too. One way to do this is to forecast take rates directly, i.e. to
calculate the probability that a certain option or even component is demanded (in
350 H. Meyr
a specific customer region) and to multiply it with the total number of car models
planned (for this region).
The task master production planning is similar to the annual budget planning.
Again, production and sales plans have to be determined and coordinated. However,
both now require a higher level of detail (e.g. weekly instead of monthly quanti-
ties) and they are not used to derive budget goals any further. The planning horizon
of a monthly rolling horizon planning varies between three months and one year.
Nevertheless, only the weekly quantities of the first month or the first two months
(depending on the lead times of planning) are put into practice.
Input data (see Fig. 2) are the already mentioned sales forecasts for models and
forecasts for take rates. Because of the high share of final customers’ orders, that
is available for this shorter planning period (see Fig. 1), these monthly forecasts
are more reliable than the annual forecasts used for budget planning. Further input
data are the production and sales quantities per month that have been agreed upon
in the budget planning, or the respective volume and earning goals (e.g. per year).
One objective of the master production planning is to meet these targets as close as
possible in the short term. Constraints to be respected are quite the same as were
relevant for the budget planning. However, again a higher level of detail is necessary.
Results of the master production planning are the updated and more detailed
(e.g. weekly) production plans of the assembly plants and sales plans. The lat-
ter ones include the quotas for the different sales regions. Because of the above
mentioned constraints, these quotas may exceed or fall below the requests for car
models, originally demanded by the regions. A similar setting of (monthly instead
of weekly) quotas for sales regions may possibly also be part of the annual bud-
get planning. For both budget and master production planning LP or MIP models
seem to be appropriate. However, for reasons to be explained in Sect. 3.3, they are
not used in practice at the moment. Planning usually is only supported by simple
spreadsheet modeling.
The production plans for car models, which are a result of the annual budget and
master production planning, are the basis to derive the component demand in a fur-
ther material requirements planning (MRP) procedure. The component demand is
communicated to the first-tier suppliers as a preview of the quantities to be deli-
vered within the next months. As the options of the cars are just specified for the
3–5 weeks before production (see Fig. 1, phases (I) and (II)), and since the share of
final customers’ orders decreases rapidly for longer lead times (phases (III) and
(IV)), this component demand becomes more and more unreliable, the longer the
forecast horizon is.
On the sales side, the allocation planning has to allocate the aggregate quotas,
which are known as a result of the budget planning on a monthly basis and as a
result of the master production planning on a weekly basis, to the lower levels of
the sales system. Depending on the organizational structure of the car manufacturer,
this planning task may occur on several hierarchical levels, e.g. first an allocation of
quotas of world regions to different countries, and afterwards an allocation of these
more detailed quotas to the countries’ respective retailers and sales subsidiaries.
As an example, in the following only the relation “world region → countries” is
considered: After the annual budget planning, the respective monthly quotas (sales
plans) of the world regions have to be allocated to the countries with respect to
their original requests. If the requests cannot all be satisfied, it has to be decided,
whose demand will only be fulfilled partly. This “shortage planning” may follow
some predefined rules (so-called “fair share rules”, see e.g. [30, p. 169 f.]), which,
for example, might reflect the purchase behavior of a country in the past, or more
or less be based on “negotiations” between representatives of the world regions and
of the respective countries. Furthermore, a region has to balance the deviations of
the countries’ actual demands from their former requests between all the different
countries assigned to the region. For this purpose, the region may also (call for and)
hold a “regional” pool of cars, originally not having been requested by one of the
countries.
3.2 Order-driven planning
Until now planning tasks have been discussed, which mainly build on forecasts for
options. In other words, only a few fully specified orders are known at the time of
planning. In this section planning tasks will be considered which are exclusively
triggered by fully specified orders, either of final customers or of sales subsidiaries
and retailers. Figure 3 gives an overview of these order-driven planning tasks and
their interrelations.
Direct buying of cars via the Internet is not (yet) worth mentioning. Normally
private customers order their cars via the sales subsidiaries or retailers of the car
manufacturer. The respective sales personnel tells the final customers the expected
delivery dates of their desired cars. Usually, a granularity of weeks is sufficient
for the customer, who e.g. has to provide the money on time and to synchronize
the delivery with the selling of his used car. Thus order promising, i.e. promising
procurement production distribution sales
(weekly) aggregate quotas

production plans,
overtime
plant specified orders allocation
assignment with due dates planning
(weekly)
production orders detailed customer &
per plant quotas retailer orders
MRP & daily buckets line assignment &
lot-sizing order promising
model mix planning specification
changes
procurement promised customer
lot-sizes daily buckets dates orders
JIT-
final
calls MRP sequencing
suppliers distribution customers
SILS car
sequence cars
Fig. 3 Overview of order-driven planning

352 H. Meyr
reliable delivery dates to the customer, is an important task. If a free quota of the
sales subsidiary or retailer is available, the final customer gets his desired delivery
date promised. Otherwise, the next free quota is recommended or a standard delivery
time is proposed (if quotas are not available in sufficient detail). The customer may
accept the promised date, change the options of his desired car or even the model
type (in order to get an earlier delivery date), or try his luck with another retailer.
Furthermore, retailers and sales subsidiaries have to specify the options for that
part of their quotas that has not been filled up with final customers’ orders until the
agreed date of specification (see phase (II) of Fig. 1). In order to reduce inventories
at the retailers’ sites, the desired options of potential customers have to be antici-
pated as precisely as possible. Because of the rather small number of customers and
large number of options, this is an almost unsolvable problem for a single retailer.
Thus, Stautner [51] suggests central support of the manufacturer for these decen-
tral forecasts of the retailers (see Sect. 5.1) and Holweg and Pil [24] even propose a
central pool of BTS cars.
Traditionally, these fully specified orders are collected by the respective sales
organization, responsible for a certain retailer, and sent in bulk (e.g. all orders of a
week) to the next higher level of the sales hierarchy. A central order management
department of the car manufacturer finally has to select an assembly plant, able to
produce the car model requested by a certain order. This plant assignment has to
consider the production quantities and capacities per plant, that have been agreed
upon in the master production planning. If the actually requested car options signifi-
cantly deviate from the ones assumed within master production planning (e.g. when
anticipating bottlenecks of components or model mix constraints), some orders have
to be fulfilled earlier and others have to be delayed, thus resulting in a re-assignment
of orders to weeks.
The selection of an assembly plant was not a big problem so far because tradi-
tionally car manufacturers had little flexibility in assigning cars to plants and thus
this task has (up to the author’s knowledge) not directly been addressed in the OR
literature. However, recently body shop and assembly have become flexible enough
to allow model swap and thus the degrees of freedom and the need for intelligent
planning methods grow. In [17], the more important assignment of customer orders
to discrete time buckets with respect to promised due dates and to material/capacity
constraints is introduced as a planning task called “demand supply matching” and
corresponding LP/MIP models are formulated. However, the specific requirements
of the automotive industry (e.g. several assembly plants, model mix constraints) are
not considered. Lovgren and Racer [33] make a first step towards mixed model as-
sembly line sequencing with respect to given due dates of orders. They calculate
detailed sequences of cars for a single assembly line. Thus, their model is rather
designed for the short-term line sequencing (see below) than the more aggregate
plant assignment. However, the problem of early or late demand fulfillment in the
automotive industry is at least generally addressed.
After this assignment, the decentral short-term production planning departments
of the assembly plants have production orders available, that ought to be assembled
within (or up to) their pre-defined week of production (ideally still the promised
week minus a standard lead time for delivery to the respective customer). The
shorter the planning horizon is, the more restrictive the model mix constraints are.
Thus, the line assignment & model mix planning have to distribute the production
orders among the possibly parallel assembly lines and to assign days of produc-
tion to the orders. Doing this, the most important model mix constraints (e.g. “at a
maximum 300 air conditionings per day”) have to be considered, but the assembly
sequence of a day is not yet determined. Scholl [47, p. 108 f.] denotes this task as
“Master Sequencing” and suggests, for reasons of complexity, a further aggregation
of individual orders to families of cars. Again, this planning task has not adequately
been tackled in the literature. Only Mergenthaler et al. [35] and Ding and Tolani
[11] address the single line (sub-)problem directly. The former ones try to smooth
the daily workload of a week by modifying a bin packing algorithm in order to
minimize the quadratic model mix deviation in a greedy manner, whereas the latter
ones apply simple neighborhood operations like “switching models of differently
utilized days” in a two-phase greedy algorithm.
As compared to mid-term planning, car options are now known with a high re-
liability. Since the daily assembly buckets are also known as a result of the line
assignment & model mix planning, the daily demand of components can directly
be derived. For components and material, that are collected by regional carriers and
temporarily stored in an intermediate warehouse (see p. 347), an MRP & lot-sizing
procedure is appropriate that balances the trade off between inventory holding costs
and degressive transportation costs of the regional carriers and determines adequate
supply frequencies.
The daily buckets of the line assignment & model mix planning are also guide-
lines for the daily sequencing of the assembly lines. Here, the sequences of the
production orders on the final assembly lines are determined on a rolling horizon
basis with a planning horizon of one to two weeks. The level of detail again is higher
than in model mix planning. Now all potential bottlenecks have to be considered,
for example, the availability of all of the components and “distance” restrictions of
the lines like “no two cars with air-conditioning are allowed to follow each other ”.
For this reason, sometimes the earlier assignment to days of production cannot be
maintained. However, it should be avoided to postpone an order to another week
than the planned (and promised) one. To use flexible workforce or to work during
lunch breaks are short-term measures to extend capacity.
Undoubtedly, most scientific research on planning aspects of automotive supply
chains has been done in the fields of balancing and sequencing mixed-model assem-
bly lines. In the sequencing literature, usually it is assumed that orders have already
been assigned to a certain period (e.g. a day) of production, so that promised due
dates need not to be considered any further. The various sequencing approaches
differ with respect to their different objectives. Besides cost-oriented objectives,
mainly time related or JIT-objectives and combinations thereof are pursued (see
e.g. [32, p. 44 ff.] and [47, p. 98 ff.]). A comprehensive literature review of models
and exact/heuristic solution methods is given by Scholl [47]. Summing up, priority
based (greedy) heuristics [47, p. 205 ff.] are – for reasons of complexity – clearly
favored over exact (mainly branch and bound) methods [47, p. 199 ff.]. Newer
heuristic approaches also apply multi agent systems [9] or local search methods like
simulated annealing or genetic algorithms (see e.g. [25, 35, 43]; [46, p. 40]).
354 H. Meyr
For a review of models and methods with respect to the different objectives,
the reader is referred to Lochmann [32]. Models with time related objectives
[32, p. 58 ff.] try to smooth the work load and minimize the overload of the var-
ious stations of a line. For this, usually MIP models are formulated. JIT-objectives
attempt to smooth the material supply at the stations in order to keep the inventory
of components constantly low. The usage rates of components are either leveled
directly [32, p. 81 ff.] or, in case the cars require a similar number and mix of compo-
nents, the mix of cars is leveled instead [32, p. 86 ff.]. The latter “level scheduling”
was introduced by Miltenburg [37] and commonly pursues nonlinear goals. Thus
both time related objectives and JIT-objectives directly address the model mix con-
straints discussed so far.
The car sequencing problem (CSP), originally introduced by Parretto et al. [40],
allows to model the above mentioned minimum distances between orders with the
same options and further separation rules like a “maximum number of identical
options within a car sequence of predefined length”. The CSP in not widely known
within the OR community, but one of the classical problems in the literature on
constraint satisfaction problems [7]. Brailsford et al. [7] review this kind of literature,
showing that these “soft” constraints also pursue time-related objectives and that
JIT- and some further objectives can also be modeled as soft constraints of a CSP.
They report that – by using a hybrid approach combining simulated annealing and
constraint logic programming – David and Chew [10] are able to obtain good solutions
for a practical problem at Renault involving 7,500 cars with 50–100 options each.
Recent approaches of Drexl et al. combine the classical CSP with level schedul-
ing [12] and solve it in a two stage approach [13]. Also Monden [38, Chap. 17]
extends his JIT-oriented “goal-chasing” heuristics in order to respect CSP distance
objectives (denoted as “continuation control” and “interval control”). Zeramdini et
al. [55] propose a two-stage approach, smoothing the components’ usage first and
the workload secondly, to optimize the bicriteria sequencing problem. Korkmazel
and Meral [31] reformulate the same combined problem as an assignment prob-
lem with weighted objectives and develop heuristics for it. Hyun et al. [25] and
later on Ponnambalam et al. [43] consider “minimization of setup costs” as a third
striving objective and find (near-) Pareto optimal solutions for the multi-objective
problem by using a genetic algorithm. A further overview of models and methods
for combined objectives is given in [32, p. 92 ff.]. Concluding this brief discussion
of sequencing, it can be stated that there is a trend in recent literature on mixed-
model assembly line sequencing to consider several objectives, simultaneously.
The frozen car sequence is then the basis to derive the component demand for
JIT calls and SIL supply. This short-term material requirements planning (MRP) is
not a “real” planning task because there is nothing left to be decided about. Just the
BOM has to be exploded as late as possible before the scheduled delivery (usually
several times per day). It is just mentioned to provide a complete picture of supplier
relationships.
If final customers do not pick up their cars at the assembly sites directly, the fin-
ished cars have to be brought to the customers or their respective retailers and sales
subsidiaries. There again are some decisions to be made concerning the distribution
of the finished cars. For example, the actual carrier has to be chosen, and transport
frequencies (how often to deliver to a retailer) and vehicle routes (sequence of

retailers within a tour) have to be determined. Some of these tasks are in the planning
domain of logistic service providers [8].
3.3 Organizational issues
One has to be aware that in the preceding sections only “abstract” planning tasks
of car manufacturers have been described, but organizational issues have been left
aside. In reality, often several different planning departments are involved in a single
planning task. Then there are several “coordination rounds“ whose result is a com-
mon plan. Within each coordination round, a single department has to contribute its
own (locally “optimal”) partial plan until some predefined date. Such a (temporar-
ily valid) partial plan is a sort of self-commitment of the respective department and
provides input for the next planning activity of another department. This procedure
iterates until the common plan hopefully respects all relevant constraints and fulfills
the various and sometimes conflicting objectives of the different departments to an
acceptable level.
The mutual arcs between production and sales in the budget planning and master
production planning boxes of Fig. 2 ought to indicate that in practice the respective
planning task usually is not tackled in a single, simultaneous planning procedure,
but in the above mentioned coordination rounds. This is one reason why LP and
MIP models are not used for a simultaneous budget planning or a simultaneous
master production planning as it is common practice in other types of industries
like consumer goods manufacturing, for example [45]. Wahl [54] proposes appro-
priate models to – at least individually – optimize the planning decisions of the sales
department in this way. But even such a local application of LP and MIP has not
been implemented in practice for reasons like missing (IT) infrastructures, inappro-
priate forms of organization or mostly a lack of acceptance and understanding of
OR methods.
4 Current trends in the German automotive industry
As Fig. 1 shows, “to move from BTS to BTO” is a somewhat imprecise formulation.
The task is rather to increase the share of final customers’ orders. Further strategic
goals, currently pursued in the German automotive industry, are to shorten customer
order delivery times of customized cars, to keep promised delivery dates with a
high reliability and to allow customers to change their car options also in the very
short term [51, p. 31 ff.]. In order to reach these goals in addition to supply chain
collaboration (see e.g. [18]) two major bundles of measures, online ordering and
late order assignment, have been and are still being implemented.
4.1 Online ordering
The total order-to-delivery lead time (OTD) can be shortened by reducing the lead
times of all individual processes (like order entry and processing, manufacturing,
356 H. Meyr
100 %
accumulated
(IV) (III) (II) customer
options share
decreasing lead times specified by
of order entry increases final customers
share of final customers (I)
30 %
10 %
time
average lead times of

pro- inven-
order
duc- distribution tory
entry
tion
arrival handing over
retailer to customer
Fig. 4 Example of lead time reduction by online ordering
distribution) involved. Since manufacturing and distribution only comprise a very

small percentage of the OTD (about 16% according to Holweg and Jones [22,
Fig. 3]), the highest potential can be found in order entry and processing. “On-
line ordering” initiatives aim at simplifying and accelerating the circumstantial and
timely collecting and (weekly) bulk processing of orders within the multi-stage sales
hierarchy. Thus retailers send fully specified ordering requests of final customers via
the Extranet or Internet directly to a central order processing system, where the re-
quests are online (i.e. within seconds or minutes) checked for technical feasibility
and provided with a promised delivery date. In case of final customer’s acceptance
of the promised date, the final order is processed with the same speed on the same
route. By implementing such a system, the car manufacturer BMW tries to reduce
the lead time of order entry from 13–17 days to a single day [44], for instance.
Figure 4 graphically illustrates how online ordering reduces demand uncertainty.
In this (fictitious) example, cutting the lead times of order entry in half triples the
share of final customers’ orders known. Thus the forecast-based BTS inventory of
retailers (phase (II), see also Fig. 1) can be reduced significantly.
4.2 Late order assignment
Traditionally each body-in-white, physically processed within the body shop, is al-
ready assigned to a customer order (“order assignment”) and a re-assignment to
another order is only rarely practicable. Following the pull-principles of the just-in-
time philosophy the final assembly as the last production stage has to be planned
first and synchronizes all direct suppliers and upstream production stages, especially
the paint shop and the body shop. In the light of “lean thinking” the work-in-process
buffers (body store and painted body store) should be small and thus body and paint
shop ideally produce in the same sequence of customer orders as is planned for the
final assembly. However, these buffers are still necessary because process failures in
the body and paint shops occur frequently [49, p. 29 f.]. According to Holweg [21]
the rework rate is even up to 40–50%. For this reason, a planned assembly sequence
can only be considered to be reliable, when the respective orders’ painted bodies
have left the paint shop. Thus the sequence can only be transmitted to the SIL sup-
pliers a few hours before planned assembly, depending on the assembly station and
the respective component.
In order to guarantee more reliable assembly plans, which can be fixed for a
longer time interval (about 4–6 days), the order assignment nowadays is postponed
to the final assembly stage (“late order assignment” or “late order tagging”, see
[21]), i.e. the bodies in the body and paint shops are no longer identified by customer
orders. Body and paint shops still get the information about the customer orders to
be assembled, but are free to deviate from the planned assembly sequence. Although
there is no demand uncertainty, safety stocks have to be installed for each body-in-
white variant and paint color. These safety stocks exclusively hedge against the
process failures in the body and paint shops. In order to limit the total amount of
safety stock required and to restrict buffer sizes, the number of body-in-white vari-
ations and paints (the so-called “internal complexity”, [21]) should be low. For this
reason, BMW reduced the number of body-in-white variations from 40,000 to 16
for their new three series when introducing late order assignment [21]. The higher
stability of assembly plans is expected to increase the radius of JIT/SIL delivery and
the share of JIT/SIL-suppliers significantly.
5 Impacts on planning
Online ordering and late order assignment have been and still are being introduced
by BMW (project title “Kundenorientierter Vertriebs- und Produktionsprozess”
[44]) and DaimlerChrysler (project titles “Global Ordering” and “Perlenkette”
[18]). Further car manufacturers intend to follow. These two types of measures con-
siderably influence the traditional planning landscape as discussed in Sect. 3. Thus
it is necessary to check how planning requirements and information flows change
(some planning tasks may loose importance whereas others win) and which new
planning tasks arise.
5.1 Impacts of online ordering
Online ordering and online order promising require extremely short response times
for incoming customer requests. If highly reliable promised delivery dates shall
be achieved, the capacities of all potential bottlenecks (material or production re-
sources) have to be checked. Thus the formerly decentral order promising has to
be automated and centralized. The changes in the planning landscape depend on
the level of delivery reliability aspired. In the following only two extreme scenar-
ios, denoted as quota-available-to-promise (QATP) and capable-to-promise (CTP)
scenario, are discussed as examples. Of course, there are various intermediates con-
ceivable between these extremes.
358 H. Meyr
production sales
aggregate quotas
(weekly)
production plans,
overtime plant allocation
assignment planning
specified orders
with due dates detailed
production orders
per plant quotas
(QATP)
daily line assignment & specification online
buckets model mix planning changes order promising promised
dates
requests,
promised orders,
daily buckets dates specification
retailers
changes
final
customers promised
dates
Fig. 5 QATP scenario
5.1.1 QATP scenario
The QATP scenario is more or less an automation of already existing processes. As

Fig. 5 shows, the general logic of planning stays the same. The quotas for retail-
ers and sales subsidiaries, which have been determined on a weekly basis anyway
and have been synchronized with capacities in the medium term, are (as far as they
have not yet been assigned to final customers’ orders) considered to be “available
to be promised”. Incoming customer requests and customer orders, respectively,
are checked for technical feasibility [26], first, and according to simple precedence
rules [30] for free quotas, secondly. Such a proceeding is known from material con-
strained industries like the computer industry and successfully applied there [29]. In
contrast, however, material availability is not yet checked in the simple QATP sce-
nario. The installation of an online ordering system (OOS) is technically lavish and
costly, but hardly changes the planning logic. When comparing Fig. 5 with Fig. 3,
the major differences are that specified orders (and their due dates) are directly
transmitted to the plant assignment instead of using the multi-stage sales hierarchy
and that specification changes can be sent faster (and thus later) to the model mix
planning.
However, because the mid-term capacity check, on which free quotas (QATP)
are based, had no detailed information about the customers’ choice of car options,
there is a high probability that the promised delivery dates do not fit the model mix
constraints and thus cannot be kept on the short-term.
5.1.2 CTP scenario
In order to achieve a higher delivery reliability, a shorter-term and more detailed ca-
pacity check is necessary, which motivates the other, more challenging extreme,
the capable-to-promise (CTP) scenario. When accepting orders and confirming
delivery dates, the customer orders are directly booked [22] to a day of production or
week of production (if a late delivery is desired by the final customer) of an adequate
assembly plant. In contrary to the QATP scenario, all or at least the most crucial
constraints, relevant for model mix planning (like options of the orders, material
required, production capacity, quotas of the respective sales hierarchy), are consid-
ered. The order promising is extended such that production orders can automatically
be generated. Thus the online order promising takes on planning tasks of the short-
term production planning or – at least – limits its scope. Furthermore, also the plant
assignment has to be integrated into such a comprehensive online ordering.
Questions, which have to be answered online, are for example: Is a BTS car
physically available somewhere in the supply chain, which fits the requirements of
the new customer order to a very high degree? Is a similar BTS car planned and can
its options be changed so that the order still can be assigned to it? Which plant has
to be chosen if a new production order has to be generated? Should be produced
earlier or later than the desired date, if this is already (over)booked? If model mix
constraints are limiting, which car specifications should a customer change in order
to still get his desired delivery date promised?
However, one has to keep in mind that the computational burden to update all
the necessary data and the desired response times of the OOS are conflicting. The
major problem is to find the right trade off between modeling capacities as detailed
as necessary (increases delivery reliability) and updating as few data as possible (in
order to guarantee short response times).
Figure 6 shows the embedding of a CTP online order promising into the plan-
ning landscape. The online order promising needs free quotas (QATP) and not yet
assigned net capacities of material (“material-available-to-promise”, MATP) and
assembly resources (CTP), e.g. expressed by a maximum number of cars with a spe-
cific (combination of) option(s) per day, as inputs. The results of the order promising
are weekly delivery dates, which are promised to the customers, and “production
production sales
(weekly) production aggregate quotas

plans, overtime
netting allocation
planning
MATP,
CTP detailed quotas
daily
buckets (QATP)
production orders
per plant online
line assignment & and day/week order promising promised
model mix planning specification (incl. plant assignm.) dates
changes
daily requests,
promised orders,
buckets
dates specification retailers
changes
sequencing final
customers promised
dates
Fig. 6 CTP scenario

360 H. Meyr
orders” with a promised delivery date and a planned day (or week) of production,
which are sent to the line assignment & model mix planning of the respective plants.
A decentral model mix planning is still necessary for several reasons. For exam-
ple, the preliminary production plans of order promising have to be updated with
respect to (for complexity reasons) still unconsidered capacity constraints and a line
assignment has to be made. Production orders, which have only been allocated to
a week of production because of rather long customer order lead times being de-
sired, have to be assigned to a day of production. The more detailed the capacity
constraints of order promising are, the less changes of its plans should be necessary
later on in the model mix planning because the most crucial potential bottlenecks
have already been anticipated. However, short-term failures in supply and produc-
tion can never be avoided and thus make a re-planning necessary.
The results of the model mix planning are daily buckets, which again are sent to
the sequencing, but are – in a further netting procedure – also used to calculate the
(net) MATP and (net) CTP for the online order promising. Further input for the net-
ting is up-to-date information about plant capacities and projected material supply,
which have been synchronized in the master production planning in the medium
term (see Fig. 2). Fleischmann and Meyr [17] illustrate the interaction between
“order entry” (online order promising) and “MATP/CTP (re)calculation” (netting
and model mix planning) by means of two more detailed workflows and discuss
the planning tasks of demand fulfillment for various positions of decoupling points.
They also propose LP and MIP models for order (re-) promising, which are useful
if several customer requests/orders can be processed in a batch. However, if each
customer request has to be answered immediately, the degree of freedom is rather
low. Consequently, the importance and impact of previous planning tasks, like mas-
ter production and allocation planning, grows. The APS vendor SAP [46] offers a
software module called Realtime-Positioning, which has especially been designed
for the online order promising in the CTP-scenario, and Ohl [39, p. 207 ff.] dis-
cusses the advantages of “code rules”, describing the interrelations between various
car options, for a capacitated BOM-explosion of online queries. However, formal
models for the MATP/CTP calculation and search are not presented. Similarly to
the approach of Ohl, Bertrand et al. [4] propose to use a hierarchical pseudobill of
material for the MATP check in case of strong interdependencies between different
options (so-called “non-modular products”).
Besides newly arriving customer requests/orders also changes of the specifica-
tion of already accepted orders can be processed and checked for capacity using
the online ordering system. Furthermore, the specification of not yet fulfilled quo-
tas by retailers (see phase (II) of Fig. 1) can be checked with respect to model mix
constraints. Online ordering accelerates order processing and increases the share of
final customers’ orders (see Fig. 4), but BTS cars cannot completely be avoided [51,
p. 6]. In order to decrease the times in inventory of the remaining BTS cars, final
customers’ desired options should be anticipated more precisely. Central statistics
about the final customers’ preferences and about frequently purchased options can
comfortably be made available to retailers by means of the OOS. They widen the
local view of the retailers and promise a higher quality of forecasts for BTS spec-
ifications [51, p. 176 ff.]. These proposals for BTS options and the more detailed
MATP/CTP capacity check can be seen as new potentials that arise due to the cen-
tralization of order promising and the online connection to retailers.
5.2 Impacts of late order assignment
Late order assignment undoubtedly has its major impacts on strategic planning.
Products have to be re-designed so that a high number of options (high external
variety) can be kept up while simultaneously reducing the number of body-in-white
variations (low internal complexity [21]). There is a rich OR literature on design
for postponement and modularization (see Sect. 1 and [3], for instance), which tries
to support such issues. Furthermore, the re-dimensioning of the (body store and)
painted body store is a strategic planning task.
But also for the operational planning of the body and paint shops and their re-
spective stores new challenges arise because of the higher degrees of freedom. The
safety stocks of the body store and the painted body store have to be refilled with re-
spect to the failure probability of the respective production processes. Because of the
rather loose coupling to the assembly sequence and because of the increased buffer
sizes, lot-sizing issues can now be considered easier in the paint shop. Although
changeover times are negligible, batching lots is economically desirable because a
change of the paints incurs costs between e10 and e30 [49, p. 30]. Taking cars out
of the body store is a Sequential Ordering Problem [15], a special variant of the
Traveling Salesman Problem. For paint shops as a practical application, Spiecker-
mann [49, p. 126 ff.] proposes a branch-and-bound approach which takes advantage
of special knowledge about common structures of body stores in the automotive
industry (see also [50] for earlier approaches to the same problem). Engel et al.
[14] propose a heuristic for workload leveling which can be extended for the batch
sequencing of paints of the same color.
Inman and Schmeling [27] prove the operational advantages of late order assign-
ment by means of simulation. They compare the traditional irreversible coupling of
orders and physical vehicles at the body shop with a flexible assignment procedure
(“Agile Assemble-to-Order” (AAO) system) that is able to assign and re-assign or-
ders to vehicles before the body shop, paint shop and the final assembly are entered.
The objective of the AAO system is a weighted function comprising penalty terms
for violating lead time, paint color, spacing and levelness constraints. Orders are
selected by the AAO system in a greedy manner with the weights varying according
to preferences of the production stage under consideration.
6 Conclusions and outlook
Concerning forecast-driven planning it can be stated that the quotas of the tra-
ditional master production and allocation planning had a detrimental effect on
meeting final customers’ demand on time. This gets even worse if the same quotas
are directly taken over to an OOS with automatic booking and without the pos-
sibility of human intervention (see Sect. 5.1.1). Thus, if still necessary to smooth
the workload in the medium term, one has to think about more flexible allocation
362 H. Meyr
mechanisms incorporating the increased knowledge (see Sect. 4.1) about final
customers’ demand. Virtual, central car pools, accessible for several retailers, are a
first step in this direction. The choice of adequate aggregation levels, allowing to
postpone decisions as long as possible, is crucial.
In other lines of business it has been shown that LP and MIP models can support
planning tasks like budget, master production and allocation planning. Wahl [54]
has proven that this would also be true for (at least the sales side of) automotive
industries. The reasons, why the proposals of Wahl have not been put into practice,
should have diminished or even vanished in the meantime. Information technology
has improved dramatically in recent years and there seems to be a higher willingness
to make use of OR tools. APS, for example, are a comfortable and user friendly way
to apply LP and MIP methods in practice. In addition, simultaneous optimization
covering several departments like production, procurement and sales in a single
model could exploit further potentials and – at least simulatively – support and
accelerate the lengthy coordination rounds (see Sect. 3.3).
Regarding the traditional order-driven planning it has been shown that OR
support for the planning tasks plant assignment and line assignment & model mix
planning was very poor. However, these tasks will change their character anyway
when online ordering and the CTP scenario are installed. On the other hand, there is
rich literature on assembly line sequencing and research in this field is an ongoing
process. Recent OR-related papers tend to pursue several objectives simultaneously,
thus becoming more attractive for practical application in the automotive industry.
However, scalability of sophisticated methods is still a problem and should be a
topic of future research.
As we have seen, the measures to move from BTS to BTO also have signifi-
cant impact on planning. The consequences for forecast-driven planning have been
sketched above. Further challenges can be identified for the future order-driven
planning. Due to late order assignment the close coupling of body, paint and assem-
bly shops has been decreased now. Thus there remains supplementary freedom for
paint shop sequencing and batching of paints of the same color. However, because
of still limited buffer sizes, OR models have to take care that paint shop sequences
may not deviate too far from assembly sequences.
Online ordering is most challenging in the CTP scenario when incoming orders
have to be booked directly into a (capacitated) production plan of a plant. In this
case, online order promising takes over functionalities of the traditional plant as-
signment and the traditional line assignment & model mix planning. The three most
crucial problems are
– how to model quotas and model mix restrictions as constraints for the online
order promising (within the netting procedure, respecting the results of the pre-
vious master production and allocation planning),
– which fast algorithms or search rules to use for allocating free QATP, MATP and
CTP (within online order promising) and
– how to revise the resulting preliminary production plans in case of still unconsid-
ered constraints and unforeseen short-term events (new line assignment & model
mix planning, respecting the already promised due dates).
Research has to be done on both OR models/methods for the different planning tasks
involved and – since responsibilities change – also on the (hierarchical) interrelation
of these planning tasks within the overall planning framework. If, above all, car
manufacturers think about customized sales prices, which may vary according to the
delivery times desired by final customers, the relationship to revenue management
(see e.g. [34, 48]), as common in airline industries, has to be further investigated.
References
1. Amen M (2000) Heuristic methods for cost-oriented assembly line balancing: a survey. Int J Prod
Econ 68: 1–14
2. Ashayeri J, van Eekelen B, Vossebeld R (2001) Lean and clean distribution in Europe: how to
get an efficient European distribution network. Technical Report, Department of Econometrics &
Operations Research, Tilburg University, The Netherlands; Presented at EURO 2001 Conference,
Rotterdam, The Netherlands, July 9–11
3. Aviv Y, Federgruen A (1999) The benefits of design for postponement. In: Tayur SR, Ganeshan
R, Magazine MJ (eds) Quantitative models for supply chain management. Kluwer, Amsterdam,
pp 553–584
4. Bertrand JWM, Zuijderwijk M, Hegge HMH (2000) Using hierarchical pseudo bills of material for
customer order acceptance and optimal material replenishment in assemble to order manufacturing
of non-modular products. Int J Prod Econ 66: 171–184
5. Biller S, Bish BK, Muriel A (2002) Impact of manufacturing flexibility on supply chain performance
in the automotive industry. In: Song J-S, Yao DY (eds) Supply chain structures: coordination,
information and optimization. Kluwer, Boston, MA, pp 73–118
6. Bradley JR, Blossom AP (2002) Using product-mix flexibility to implement a made-to-order as-
sembly line. Technical Report, Cornell University, Ithaca NY; under revision for IIE Transactions
7. Brailsford SC, Potts CN, Smith BM (1999) Constraint satisfaction problems: algorithms and appli-
cations. Eur J Oper Res 119: 557–581
8. Carbone V, De Martino M (2003) LSPs in the automotive supply chain: a case study. In: Juga J (ed)
NOFOMA 2003: striving for leading edge logistics. Proceedings of the 15th Annual Conference for
Nordic Researchers in Logistics, Oulu, Finland, 12–13 June, pp 416–431
9. Caridi M, Sianesi A (2000) Multi-agent systems in production planning and control: an application
to the scheduling of mixed-model assembly lines. Int J Prod Econ 68: 29–42
10. David J-M, Chew TL (1995) Constraint-based applications in production planning: examples from
the automotive industry. In: Practical Applications Company (ed) Proceedings of Practical Appli-
cations of Constraint Technology (PACT’95), Blackpool, UK, pp 37–51
11. Ding F-Y, Tolani R (2003) Production planning to support mixed-model assembly. Comput Ind Eng
45: 375–392
12. Drexl A, Kimms A (2001) Sequencing JIT mixed-model assembly lines under station-load and
part-usage constraints. Manag Sci 47(3): 480–491
13. Drexl A, Kimms A, Matthießen L (2002) Algorithms for the car sequencing and level scheduling
problem. Technical Report, University of Kiel, Germany
14. Engel C, Zimmermann J, Steinhoff A (1998) Objectives for order-sequencing in automobile pro-
duction. In: Drexl A, Kimms A (eds) Beyond manufacturing resource planning (MRP II). Springer,
Berlin, pp 307–331
15. Escudero LF (1988) An inexact algorithm for the sequential ordering problem. Eur J Oper Res 37:
236–253
16. Fisher ML, Ittner CD (1999) The impact of product variety on automobile assembly operations:
empirical evidence and simulation analysis. Manag Sci 45(6): 771–786
17. Fleischmann B, Meyr H (2003) Customer orientation in advanced planning systems. In: Dyckhoff
H, Lackes R, Reese J (eds) Supply chain management and reverse logistics. Springer, Berlin,
pp 297–321
18. Graf H, Putzlocher S (2002) DaimlerChrysler: Integrierte Beschaffungsnetzwerke. In: Corsten D,
Gabriel C (eds) Supply Chain Management erfolgreich umsetzen. Springer, Berlin, pp 47–61
364 H. Meyr
19. Graves SC, Tomlin BT (2003) Process flexibility in supply chains. Manag Sci 49(7): 907–919
20. Hoffman KC (2000) Ford develops different kind of engine – one that powers the supply chain.
Global Logistics Supply Chain Strategies, pp 42–50
21. Holweg M (2002) Product variety – friend or foe? Technical Report, Lean Enterprise Research
Centre, Cardiff Business School, United Kingdom; to appear in Int J Prod Econ
22. Holweg M, Jones DT (2001a) The build-to-order challenge. Automotive World Jan–Feb, pp 40–45
23. Holweg M, Jones DT (2001b) The challenge: building cars to order – can current automotive supply
systems cope? In: Tyler D, Brunt D (eds) Manufacturing operations and supply chain management
– the Lean approach. Thomson Learning, London, pp 362–374
24. Holweg M, Pil FK (2001) Successful build-to-order strategies – start with the customer. Sloan
Management Review Fall, pp 74–83
25. Hyun CJ, Kim Y, Kim YK (1998) A genetic algorithm for multiple objective sequencing problems
in mixed model assembly lines. Comput Oper Res 25(7/8): 675–690
26. ILOG (2001) ILOG configurator. Technical Report, ILOG Worldwide Information Center,
http://www.ilog.com/products/configurator/whitepaper/in dex.cfm?filename=wp configurator.pdf –
last check of address: March 31, 2004
27. Inman RR, Schmeling DM (2003) Algorithm for agile assembling-to-order in the automotive in-
dustry. Int J Prod Res 41(16): 3831–3848
28. Jordan WC, Graves SC (1995) Principles on the benefits of manufacturing process flexibility. Manag
Sci 41(4): 577–594
29. Kilger C, Schneeweiss L (2002a) Computer assembly. In: Stadtler H, Kilger C (eds) Supply chain
management and advanced planning, 2nd ed. Springer, Berlin, pp 335–352
30. Kilger C, Schneeweiss L (2002b) Demand fulfilment and ATP. In: Stadtler H, Kilger C (eds) Supply
chain management and advanced planning, 2nd ed. Springer, Berlin, pp 161–175
31. Korkmazel T, Meral S (2001) Bicriteria sequencing methods for the mixed-model assembly line in
just-in-time production systems. Eur J Oper Res 131: 188–207
32. Lochmann M (1999) Operative Planung und Steuerung von Mehrprodukt-Fließlinien. Deutscher-
Universitäts-Verlag, Wiesbaden
33. Lovgren RH, Racer MJ (2000) Algorithms for mixed-model sequencing with due date restrictions.
34. McGill JI, Van Ryzin GJ (1999) Revenue management: research overview and prospects. Transport
Sci 33(2): 233–256
35. Mergenthaler W, Stadler W, Wilbertz H, Zimmer N (1995) Optimizing automotive manufacturing
sequences using simulated annealing and genetic algorithms. Control Eng Pract 3(4): 569–573
36. Miller T, Wise D, Clair L (1996) Transport network design and mode choice modeling for automo-
bile distribution: a case study. Location Sci 4(3/4): 37–48
37. Miltenburg GJ (1989) Level schedules for mixed-model assembly lines in just-in-time production
systems. Manag Sci 35: 192–207
38. Monden Y (1998) Toyota production system, 3rd ed. Engineering & Management Press, Norcross
39. Ohl S (2000) Prognose und Planung variantenreicher Produkte am Beispiel der Automobilindustrie.
Fortschritt-Berichte VDI 16/120, VDI Verlag, Düsseldorf
40. Parretto BD, Kabat W, Wos L (1986) Jobshop scheduling using automated reasoning: a case study
of the car sequencing problem. J Automat Reas 2: 1–42
41. Patchong A, Lemoine T, Kern G (2003) Improving car body production at PSA Peugeot Citroen.
Interfaces 33(1): 36–49
42. Pine BJ (1993) Mass customization: the new frontier in business competition. Harvard Business
School Press, Boston
43. Ponnambalam SG, Aravindan P, Subba Rao M (2003) Genetic algorithms for sequencing problems
in mixed model assembly lines. Comput Ind Eng 45: 669–690
44. Renner P (2000) Kundenorientierung in der Automobilindustrie. In: Innovationen in Logistik-
strukturen der Automobilindustrie, Tagung 28.9.00 / VDI-Gesellschaft Fördertechnik, Materialfluß,
Logistik, VDI Bericht 1571. VDI Verlag, Düsseldorf, pp 21–33
45. Rohde J, Wagner M (2002) Master planning. In: Stadtler H, Kilger C (eds) Supply chain manage-
ment and advanced planning, 2nd ed. Springer, Berlin, pp 143–160
46. SAP AG (2001) “Managing the e-supply chain” in der Automobilindustrie mit mySAP automotive.
SAP White Paper (50 047 263), 05/2001.
47. Scholl A (1995) Balancing and sequencing of assembly lines. Physica-Verlag, Heidelberg
48. Secomandi N, Abbott K, Atan T, Boyd EA (2002) From revenue management concepts to software
systems. Interfaces 32(2): 1–11
49. Spieckermann S (2002) Neue Lösungsansätze für ausgewählte Planungsprobleme in Automobilroh-
bau und -lackiererei. Shaker Verlag, Aachen
50. Spieckermann S, Voß S (1996) Paint shop simulation in the automotive industry. ASIM Mitteilungen
54: 367–380
51. Stautner U (2001) Kundenorientierte Lagerfertigung im Automobilvertrieb. Deutscher-Universitäts-
Verlag, Wiesbaden
52. Swaminathan JM, Tayur SR (1999) Managing design of assembly sequences for product lines that
delay product differentiation. IIE Trans 31: 1015–1026
53. Thonemann UW, Brandeau ML (2000) Optimal commonality in component design. Oper Res
48(1): 1–19
54. Wahl AH (1994) Simultane Produktions- und Distributionsplanung. Peter Lang, Frankfurt a.M.
55. Zeramdini W, Aigbedo H, Monden Y (2000) Bicriteria sequencing for just-in-time mixed-model
assembly lines. Int J Prod Res 38(15): 3451–3470
Modeling and optimizing of strategic and tactical
production planning in the automotive industry
under uncertainty
Ralf Bihlmaier · Achim Koberstein · René Obst

OR Spectrum (2009) 31:311–336
DOI 10.1007/s00291-008-0147-2
Abstract This work considers the strategic flexibility and capacity planning under
uncertain demands in production networks of automobile manufacturers. We present
a deterministic and a stochastic model, which extend existing approaches, especially
by an anticipation scheme for tactical workforce planning. This scheme is compared
to an extended formulation of the deterministic model, which incorporates workforce
planning via detailed shift models. The stochastic model is efficiently solved by an
accelerated decomposition approach. The solution approach is integrated into a deci-
sion support system, which calculates minimum-cost product allocations and capacity
plans. Our numerical results show that, in spite of the considerably increased com-
plexity, our approach can efficiently handle hundreds of scenarios. Finally, we present
an industrial case study.
Keywords Strategic network design · Anticipation of tactical planning · Stochastic

programming · Decomposition approach
1 Introduction
Today companies of the automobile industry face a market situation which is par-
ticularly characterized by dynamic change and uncertainty. Due to decreasing sales
figures and stagnating market prices especially the companies in the premium
R. Bihlmaier (B)
Daimler AG, Group Research and Advanced Engineering, 89073 Ulm, Germany
A. Koberstein (B) · R. Obst

DS & OR Lab, University of Paderborn, Warburger Str. 100, 33098 Paderborn, Germany

368 R. Bihlmaier et al.
segment increasingly concentrate on new niche markets. To get higher market-shares

in the oligopolistic cutthroat competition, they try to constantly reduce product life
cycles and raise the diversity of their products. One of the consequences is the increase
of over-capacities, since installed capacities of the production facilities are considered
to be short- or medium-term unchangeable (cf. Becker 2005; Jordan and Graves 1995;
Friese et al. 2005).
Since the production facilities in the automobile industries require high investments,
a low degree of utilization yields to diminishing profit margins in terms of a high fixed
costs risk (see e.g. Roscher 2008, p.39). Therefore highly dynamic and uncertain
markets pose a high risk to the companies. The adequate answer to these challenges
is to make production facilities more flexible within the whole network (Bruynesteyn
2003). In general the meaning of flexibility is the ability to react purposefully on unex-
pected changes (Schneeweiss 1999). In the research literature, there exist a multiplicity
of different flexibility terms. A detailed theoretical taxonomy of the term flexibility is
given by Sethi and Sethi (1990). More applicable definitions are discussed for example
by Chandra et al. (2005), Gerwin (1982) and Pibernik (2001). Regarding the strategic
scope of our work, we only use the terms product flexibility, volume flexibility and
successor flexibility in the following. Product flexibility measures the ability of a pro-
duction line to manufacture several different products, volume flexibility denotes its
ability to adapt to different output rates in a cost efficient way (see Roscher 2008, p.28).
A line is characterized as successor flexible if it can be enabled to produce future prod-
uct variants. In a flexible production network a low degree of utilization and high profit
risk can be avoided by exploiting the opportunity to produce one product in different
plants or to produce different products in one plant respectively. Thus market demands
can effectively be assigned to several capacities. Hence, the assignment of products to
potential plants and the installation of flexible production capacity are the fundamental
tasks for the strategic network planning. However, the degree of flexibility is fairly
restricted by high investments, fixed costs or restrictions like the applied manufacturing
technology or different materials for the varying products (Friese et al. 2005).
Nowadays, deterministic network design problems of considerable complexity can
be solved. However, it was early recognized that deterministic models are not suit-
able to represent planning problems in highly dynamic and uncertain environments.
Therefore stochastic versions of the deterministic models and specially tailored solu-
tion methods were proposed. However, only very recent models, e.g. as proposed by
Santoso et al. (2005), are able to solve a stochastic network design problem of realistic
complexity for a large number of scenarios in acceptable time. In this paper we extend
and customize the general model of Santoso et al. to comply with special requirements
of production network planning in the automobile industry. In particular, we incor-
porate an anticipation scheme for tactical workforce planning and compare it to an
extended formulation of the deterministic model, in which we use detailed shift mod-
els to consider the task of workforce planning in a more realistic way. Furthermore we
consider multi-period combined product allocation and capacity initialization with
regard to combined investment and cost parameters. These enhancements lead to a
considerably higher complexity compared to existing models. However, our numer-
ical results indicate that it is still possible to solve instances of realistic size, which
involve hundreds of scenarios. We incorporated our solution approach into a decision
Strategic and tactical production planning under uncertainty 369
support system which is being deployed in the strategic production planning of a major
German car manufacturer. Finally we present a real world case study with the goal to
evaluate and leverage benefits of flexibility in the manufacturer’s production network.
The rest of the paper is structured as follows. In the next section we describe the
challenges of strategic network planning in the automobile industry, clarify the goals of
this study and give a literature survey. In Sect. 3 the problem is modeled as a determin-
istic and a two-stage stochastic mixed integer program with fixed complete recourse.
The deterministic model is extended by a detailed model for tactical workforce plan-
ning. Section 4 deals with a brief description of the implemented solution approach
based on an extended version of Benders’ decomposition algorithm. The application
of model and solution method to real world problems is discussed in Sect. 5 by solving
a complex planning problem of realistic size stemming from the European automobile
industry. Finally, Sect. 6 reflects the conclusions of this work and gives an outlook to
future studies.
2 Statement of the problem
2.1 Strategic network planning in the automobile industry
In the order-oriented production of the automobile industry, strategic decisions must

mainly concentrate on future markets. In this respect a main difficulty of global network
planning lies in the consideration of uncertain magnitudes. Santoso (2003) identifies
a variety of such uncertain magnitudes in global production networks, for example,
product demands, product life cycles, market prices or production costs and trans-
port costs which all are discussed by Vidal and Goetschalckx in detail (2000). Also
exchange rates can be regarded as such a critical risk factor of uncertainty (Meyer
2004), but corporate hedging strategies limit their influence. We therefore concentrate
on uncertain demand quantities and represent their magnitudes by forecasted discrete
probability distributions. However, this is just a mild limitation since continuous dis-
tributions can be discretely approximated sufficiently close (Boettcher 1989).
The main decisions in strategic network planning include where to close existing
or open new production sites, where to manufacture which product and from where
to satisfy the customers demand (Goetschalckx 2002) in every time period. Further-
more the capacities of the manufacturing lines have to be fixed. Due to complexity
avoidance, decisions related to suppliers or supplier structure are not included in the
decision model in this paper. The planning horizon should cover at least one product
life cycle for each product or two life cycles if successor flexibility potentials could be
utilized. The current life cycle duration of automotive products is 5–7 years Barnett
et al. (1995), so that the planning horizon of the stated problem will end up in five
to maximal twenty years. Regarding the long-term character of the involved strategic
decisions, 1 year is a proper representation for one time period in our model.
In the following we distinguish between technical and organizational capacity
in order to distinguish the strategic planning task of initialization from the tactical
planning task of adaptation. Technical capacity is defined as the maximal quantity a
manufacturing facility is able to produce, whereas organizational capacity determines
the actual utilization of a manufacturing facility, e.g., depending on a chosen working

time model. Regarding the process of capacity dimensioning, it has to be considered
that installed capacities can be adjusted to changing market circumstances by applying
either technical or organizational planning options. Technical options are e.g. adding
equipment to the production lines, changing production technology or altering the
cycle time of a production line. But these technical options are typically linked to
high costs. A more appropriate approach is to take advantage of the organizational
opportunity of workforce flexibility. Organizational options based on workforce flex-
ibility are the variation of shift length, percentage of temporary staff, Saturday shifts
etc. Using these midterm instruments, the management is able to react properly to
demand variations and therefore to lower the fixed costs risk. In order to fully exploit
the impact of these additional degrees of freedom of the tactical planning, they have
to be already anticipated in the strategic planning. This anticipation should be done at
least approximately to ensure no or just a sub-proportional increase in planning com-
plexity. Jordan and Graves (1995) clarify that flexibility and capacity can represent
substitutable magnitudes. Therefore the modeling of strategic network planning has
to involve decisions for both aspects simultaneously.
In order to determine the primary result of the decision model, the opening and
closing of facilities and the product assignments, even the optimal decisions of the
operational level should be rudimentarily anticipated. The objective of this integrated
optimization is to determine the production and transportation potentials as well as
the shortfalls by minimum costs and under consideration of the uncertain demands
and a given corporate strategy and policy. Hereby several decisions like using existing
facilities and given product allocations for reasons like securing existing job banks or
lowering additional investments as well as local content conditions have to be regarded
in our model formulation. The secondary result of the model gives information on the
expected utilization of the flexibility and capacity defaults. This result can only be
determined by anticipation of demand. For each type of product there are degrees of
freedom in the outputs of assigned production lines, in the transportation quantities
to markets and in the shortfalls by unfulfilled demands. Duties or import taxes, which
could also have an impact on planning global production networks (Arntzen et al.
1995), are not considered in detail. We assume that duties can be seen as a static vari-
able cost parameter which we include into the term of transportation costs in order to
shorten the model formulation later on.
2.2 Literature review
Since the ground breaking work of Geoffrion and Graves (1974) in the mid-seventies,
a large variety of literature work has been published in the field of strategic network
planning (see Beamon 1998 for a survey). The examples given are organized in the
order of increasing complexity and applicability to real world problems.
Geoffrion and Graves (1974) develop one of the first models for strategic network
planning. The scope of the mixed-integer model is to create a cost-reducing design
of a multi-commodity production and distribution network. The solution’s approach
is based upon a Benders’ Decomposition which separates the binary decisions (e.g.
location of distribution centers) from the continuous decisions (e.g. transport flows).
However, the model is not a multi-period stochastic formulation.
Arntzen et al. (1995) present a multi-period, mixed-integer model for global supply
chain planning. The model includes both a detailed production, inventory and trans-
portation planning and strategic decisions as product allocation with related fixed
costs. Investment requirements are not considered. The objective function includes
the minimization of costs as well as weighted production and shipping times. A par-
ticular feature of the model is the focus on international aspects such as duties, import
taxes or duty drawbacks. The model is applied to a real-world problem of a computer
manufacturer.
Jordan and Graves (1995) concentrate on evaluating product flexibility in order to
hedge against uncertain demands. The main target of this work is to find an optimal
degree of product flexibility. They document through numerical studies that a special
partly-flexible strategy exhibits nearly the same advantages as a fully-flexible strat-
egy. This special strategy is called chaining, because products and plants are linked
together alternately in a closed chain. Their approach could be seen as trend-setting
in the field of flexibility evaluation. Monetary analysis as well as the consideration
of capacity decisions are not considered in this approach. Boyer and Leong (1996)
expand the model of Jordan and Graves by including diverse setup costs which are
incurred by simultaneous production of several products in a flexible plant.
Following the work of Jordan and Graves, Francas et al. (2007) evaluate the impact
of demand dynamics caused by product life-cycles. Using a stochastic programming
model the authors show that benefits of flexible configurations might be substantially
misjudged if product life-cycles are not considered. However, their results also indi-
cate that prominent flexibility strategies like chaining plants remain robust even when
life-cycles are included in the analysis.
Chandra et al. (2005) formulate a model for flexibility planning specifically for the
automotive industry. The goal of their work is to investigate the dependencies between
product allocation, commonality of parts and capacity planning. They develop an
algorithmic strategy which assumes product allocation and commonality of parts to
be fixed configurations. They then calculate expected demand quantities using Monte
Carlo simulation and, lastly, optimize the capacity planning using a combination of a
genetic algorithm and linear optimization.
MirHassani et al. (2000) consider a multi-period, mixed-integer, two-stage stochas-
tic program to determine an optimal solution of capacity planning for supply chain
design problems. The first stage comprises of the opening and closing of plants, and
sets capacity levels. In the second stage, optimized decisions about production and
distribution costs are made. Additionally, the authors demonstrate how to use previ-
ous Wait-And-See analysis results in the solution method. The approach extends the
first stage problem by involving the second stage decision of one chosen scenario. By
solving the extended problem using Benders’ Decomposition, they show that when
choosing a “good” scenario, the solution time is greatly reduced. Using numerical
studies, the authors demonstrate how well this approach can handle hundreds of sce-
narios.
Alonso-Ayuso et al. (2003) present a two-stage stochastic program and a cor-
responding solution method for supply chain design problems. The implemented
strategic decisions cover plant openings, product allocation, product selection and
raw material supplier selection—all of which to maximize the net profit margin. The
net profit margin consists of sales, operating costs and depreciation of investments
figures. They model uncertain values in the costs of raw materials and production, and
incorporate them into a scenario-based approach. In the second stage, the strategic
decisions are evaluated by making tactical decisions on discrete capacity expansion as
well as production, stock and transport volumes. The authors develop a heuristic based
on a branch-and-fix coordination scheme to solve their model efficiently. Numerical
studies are only presented for a very small number of scenarios, thus, for this detailed
approach, the uncertainty is only considered rudimentarily.
Fleischmann et al. (2006) present a detailed multi-period mixed-integer model
based on experiences at BMW (for detail see Ferber 2005). While it is also based on
classic theoretical modeling approaches, the following issues are pointed out to have
a positive impact on acceptance in practice. Firstly, the choice of a cash flow based
objective, the net present value, allows to compare an optimized solution with man-
ually computed strategies. Secondly, the simultaneous optimization of capacity and
flexibility strategies fulfills the claim for integrated planning by Jordan and Graves.
Furthermore, discrete technical capacity stages are extended with a linear overload to
represent workforce planning instruments.
Santoso et al. (2005) develop a mixed-integer, two-stage stochastic program for
planning realistically scaled supply chain design networks. The first stage of the pro-
gram includes decisions on the opening and closing of facilities, capacity levels and
product allocation among the plants. The second stage includes tactical decisions to
determine optimal production and transportation volumes. The authors consider uncer-
tainty in transportation costs, demand and supply quantities as well as discrete steps
of plant capacities. In order to minimize computing time, they integrate an acceler-
ated Benders’ Decomposition method utilizing a sampling strategy to handle a great
number of scenarios.
All the approaches above show models and methods for special variants of strategic
network design problems. Some concentrate on the stochastic character of strategic
flexibility planning. Few include realistic capacity optimization. A very important and
crucial aspect of capacity planning is missing in almost all of these approaches: the
division of capacity planning tasks into technical capacity initialization and organiza-
tional capacity adaptation. To cope with complex real-world problems we present a
two-stage stochastic, mixed-integer program (including the anticipation of organiza-
tional capacity adaptation) for strategic flexibility and capacity planning of production
networks in the automobile industry in the next section. This model deals with the main
strategic decisions in the first stage and the tactical and operational decisions in the
second stage by optimizing a net present value of the profits over an extended planning
horizon.
3 Model formulation
Let P be the set of products being produced, transformed or transported in a supply

chain network within the planning horizon described by the set of time periods T .
Table 1 List of indexes
Symbol Definition
P Set of products
F Set of facilities (plants/production lines)
Sf Set of capacity-initializing stages for line f ∈ F
M Set of markets
T Set of time-periods
N Set of demand scenarios
W Set of shift models
Table 2 Cost parameters
Symbol Definition Unit
rt Interest rate for the calculation of the capital value in period t (%)
k pPfI Amount of product specific investment, (MU)
if product p is allocated to facility f
ksKf I Amount of capacity based investment, (MU)
if technical capacity stage s is initialized in facility f
PV
k ps Variable production costs of product p (MU/QU)
ft
in capacity stage s, facility f and period t
k pPfFt Production based fixed costs of product p in facility f and period t (MU)
ksKf Ft Capacity based fix costs of the initialized capacity stage s, that occur, (MU)
if it is actually deployed in facility f and period t
k Tp fI f t Cost rate for internal transport of one unit of product p (MU/QU)
from facility f to facility f in period t
k Tp fEmt Cost rate for external transport of one unit of product p (MU/QU)
from facility f to market m in period t
k Spmt
F Opportunity costs for shortfall of one unit of product p (MU/QU)
in market m and period t
ksMf tI N Cost to reduce the capacity of stage s in facility f (MU/CU)
and period t using organizational instruments by one unit (linear approximation)
ksMf tAX Cost to increase the capacity of stage s in facility f (MU/CU)
and period t using organizational instruments by one unit (linear approximation)
Every element p ∈ P may represent a raw material, an intermediate good or a final

good. Furthermore let F be the set of production lines and plants. Consequently, an
element f ∈ F characterizes a facility transforming a product p ∈ P into another
product p ∈ P. The set of markets M and the set of logistic connections between
production lines or production lines and markets form the production network as a
directed non-cyclic graph. In Table 1 we give a complete list of the sets used in our
model. Parameters of the model are distinguished in cost-based and quantity-based
Table 3 Quantity based parameters
d pmt Demand of product p in market m and period t (QU)

csKf Capacity of stage s in facility f per period in regular working time (CU)
c Ep fF F Factor that reflects the loss of efficiency induced by flexible (%)
production of product p in facility f
c pAV
f Factor that reflects the loss of capacity in the first period (%)
of production of product p in facility f
cK B
ps f Amount of capacity units of stage s needed to produce one unit (CU/QU)
of product p in facility f
cM
f
K Technical capacity per period in facility f in maximal working time (CU)
cB OM
p p
Number of units of product p to produce one unit (QU/QU)
of product p (bill of material)
Table 4 Miscellaneous parameters
ρn Probability of scenario n
dsKf M I N Minimal relative capacity reduction by organizational instruments (%)
in capacity stage s and facility f
dsKf M AX Maximal relative capacity increase by organizational instruments (%)
in capacity stage s and facility f (%)
d pPfMt I N Lower bound on the amount of product p produced in facility f and period t
d pPfMt AX Upper bound on the amount of product p produced in facility f and period t
d pL fEt Upper bound on product allocation variable y pPfAt ,
which indicates the allocation of product p to facility f in period t
(if set to 0, y pPfAt is fixed to 0)
d pL fFt Lower bound on product allocation variable y pPfAt ,
which indicates the allocation of product p to facility f in period t
(if set to 1, y pPfAt is fixed to 1)
parameters (see Tables 2, 3). Additionally, several parameters exist to represent cor-
porate-policy settings (see Table 4). There are cost based parameters regarding both
single period payment flows—like investments or investment-based one time costs—
and continuously payment flows—like fixed and variable costs. MU represents the unit
of measurement for cost parameters and refers to capacity units (CU), quantity units
(QU), time-periods (t) or a combination of several units of measurement. Following
the hierarchical planning process of the considered problem, decision variables of the
corresponding model can be ordered in strategic and tactical variables. Strategic deci-
sions involve all ‘Yes/No’-decisions and are characterized by a binary code. In order
Table 5 Decision variables
Symbol Definition
y pPfIt Indicator variable: 1, if the allocation of product p

to facility f is initialized in period t, 0 otherwise
y pPfAt Indicator variable: 1, if the product p is produced
in facility f period t, 0 otherwise
ysKf It Indicator variable: 1, if the technical capacity stage s
in facility f is initialized in period t, 0 otherwise
ysKf tA Indicator variable: 1, if the technical capacity stage s
in facility f is deployed in period t, 0 otherwise
P
x ps Real nonnegative variable: amount of product p produced
ft
in facility f and period t using capacity stage s
x Tp fI f t Real nonnegative variable: amount of product p transported
from facility f to facility f in period t (internal transport)
x Tp fEmt Real nonnegative variable: amount of product p transported
from facility f to market m in period m (external transport)
xsMf tI N Real nonnegative variable: amount of which the capacity of stage s in facility f
is reduced by organizational instruments
x sMf tAX Real nonnegative variable: amount of which the capacity of stage s in facility f
is increased by organizational instruments
z Spmt
F Real nonnegative variable: shortfall of product p on market m in period t
to approximate the real problem, all tactical decisions are represented by continuous
real-valued variables (see Table 5).
3.1 Deterministic model
Below we present the deterministic formulation of the optimization model. As an

objective value of the described model (1)–(17), a monetary ratio in terms of the
present value of period-based payment flows is determined in formula (1). Referring
to Goetschalckx (2001), the net present value represents an adequate objective for stra-
tegic network design problems as it reflects both an efficiency principle and temporary
advantages. The impact of the strategic and tactical decisions on the payment flows
are formulated separately in the functions Z F S (t) and Z F T (t), respectively. Strate-
gic decisions involve investments and induce fixed costs. Tactical decisions depend
on variable costs and cause running expenses and profits.
Z F S (t) + Z F T (t)
min Z F = (1)
(1 + rt )t
T

with Z F S (t) = k pPfI y pPfIt + k pPfFt y pPfAt
P F

+ ksKf I ysKf It + ksKf Ft ysKf tA (2)
S F

and Z F T (t) = ksMf tI N xsMf tI N + ksMf tAX xsMf tAX
S F

+ PV P
k ps f t x ps f t
P S F

+ k Tp fI f t x Tp fI f t
P F F

+ k Tp fEmt x Tp fEmt
P F M

+ k Spmt
F SF
z pmt (3)
P M
subject to (4)–(17).
Constraints (4) and (5) enforce the indispensable dependencies of the strategic deci-
sions. In particular, link and technical capacity decisions can only be used after the
corresponding initialization. Due to the character of the capacity decision only one of
the given set of options can be utilized for each production facility (6).

y pPfAt ≤ y pPfIt ∀ p, f, t (4)
t ≤t

ysKf tA ≤ ysKf It ∀s, f, t (5)
t ≤t

ysKf It ≤ 1 ∀ f (6)
S T
The decisions on the strategic level have a direct influence on the tactical utilization
of the production network. On the one hand, the decision about the links determines
the disposition of production feasibilities to lines or locations (7). On the other hand,
the capacity is already dimensioned on this decision level by construction of buildings
and technical facilities (8–10). Thereby the frame of action for cost-efficient capacity
adaptation via organizational instruments is given by a linear approximation scheme
for every capacity initialization option. The practicability of the approximation is
shown in Sect. 5.
c KpsBf x ps ft ≤ cf y p f t ∀ p, s, f, t
P MK PA
(7)

c ps f x ps f t ≤
KB P
cs f ys f t + xsMf tAX − xsMf tI N ∀ f, t
K KA
(8)
P S S
xsMf tAX ≤ dsKf M AX csKf ysKf tA ∀s, f, t (9)

xsMf tI N ≤ 1 − dsKf M I N csKf ysKf tA ∀s, f, t (10)
The forecasted demand for final products on the customer markets is playing a
decisive role regarding the design and use of optional network structures. Since the
satisfaction of all possible realizations of demand is an economic goal by its own
right, and the objective value only involves expected pay-offs, the relative difference
between possibly requested demand and provided supply quantities must be compen-
sated by unfulfilled demand, the so-called shortfall. To simplify the modeling, we
assume that every unfulfilled demand quantity results in lost sales.

z Spmt
F
+ x Tp fEmt ≥ d pmt ∀ p, m, t (11)
F
Since our planning problem is based on a multilevel value-added process, two

additional equations (12 and 13) have to be included to ensure a closed system of the
material flow. For every node of our network exactly one incoming and one outgoing
equation is implemented to achieve a material balance.

x Tp fI f t = c BppO M x pP s f t ∀ p, f, t (12)
F S P

ft = x Tp fI f t + x Tp fEmt ∀ p, f, t
P
x ps (13)
S F M
Additional constraints are established in order to incorporate project expert knowl-

edge or political decisions into the generic model. Constraint (14) shows the fact
that some product allocation decisions might be fixed, prohibited or technologically
impossible. Furthermore, constraint (15) enforces, if given, a frame for the feasible
output of several manufacturing facilities. For example, this output frame could be set
to secure the economic future of a production site or to align the output assignment
between different sites.
d pLfFt ≤ y pPfAt ≤ d pLfEt ∀ p, f, t (14)

d pPfMt I N ≤ P
x ps f t ≤ dpf t
P M AX
∀ p, f, t (15)
S
An aspect that cannot be neglected in the long-term capacity planning is, that not the
whole capacity in years of product launches is available for each affected production
line. Therefore the capacity will be reduced by a percentage rate c pAVf in constraint (16).
Furthermore, we include potential disadvantages like efficiency losses when producing
different products on the same line by a rate c Ep fF F in constraint (17).

csKf ysKf tA + xsMf tAX − xsMf tI N ≤ 1 − c pAVf y pPfIt csMf K ∀s, f, t (16)
P

csKf ysKf tA + xsMf tAX − xsMf tI N ≤ 1− c Ep fF F y pPfAt csMf K ∀s, f, t (17)
P
By constraints (4) to (17), the solution space for the deterministic model is described
sufficiently. But for a potential reduction of the required solution time, additional valid
inequalities (18) and (19) are applied to the model. Cordeau et al. (2006) describes
such additional inequalities for logistic network design problems. An analysis of the
influence to solution time will also be conducted later on in Section 5.

y pPfIt ≤ 1 ∀ p, f (18)
T

ysKf It ≤ y pPfIt ∀s, f (19)
T T P
3.2 Extension to tactical workforce planning
As mentioned in Sect. 2, tactical workforce planning should already be anticipated

at the strategic level (for a detailed workforce planning model see e.g. Askar and
Zimmermann 2006). These decisions have an essential impact on the optimal design
of the network structure by enabling the production systems to adapt capacities over
time. However, the integration—especially the level of detail—of the tactical work-
force planning has to be done carefully because of the crucial influence on required
solution time. Hence, the model formulated above integrates the tactical decisions via
a linearized approximation scheme. This guarantees acceptable solution time without
highly affecting strategic decisions. However, this approximation scheme is not suffi-
cient to determine total life-cycle costs. Therefore we extend the above model which
supports the identification of optimal capacity adaptation paths on the one hand, and
the calculation of life-cycle costs on the other hand for a given network structure and
a given demand realization. The additional parameters and decision variables used in
the extended model are described in Table 6.
The original model will be adjusted by substituting the capacity adaptation cost
terms in the objective function and the capacity adaptation constraints. In the new
objective function, the function Z F T (t) is redefined as shown in Eq. (20).

Z F T (t) = SM W F W F
rws f t k f x ws f t
F W S

+ k Hf W F x Hf tW F + k Ff W F x Ff tW F

+ PV P
k ps f t x ps f t
P S F

+ k Tp fI f t x TppI f f t
P P F F

+ k Tp fEmt x Tp fEmt
P F M

+ k Spmt
F SF
z pmt (20)
P M
Table 6 Additional parameters and decision variables in workforce planning extension
SM
yws Indicator variable: 1, if shift model w is chosen
ft
in capacity stage s, facility f and period t, 0 otherwise
WF
x ws Real nonnegative variable: number of employees deployed in shift model w,
ft
capacity stage s, facility f and period t
xH
ft
WF Real nonnegative variable: number of employees hired in facility f in period t
x Ff tW F Real nonnegative variable: number of employees dismissed in facility f in period t
SM
rws Cost parameter: factor for the shift model bonus of shift model w, capacity stage s, (%)
ft
facility f in period t (it is multiplied with employees’s wage
kW
f
F to obtain time, shift model, and capacity stage dependent costs)
kW f
F Cost parameter: wage per employee in facility f (MU)
kf F
H W Cost parameter: hiring costs per employee in facility f (MU)
k Ff W F Cost parameter: dismissal costs per employee in facility f (MU)
SM
cws Capacity parameter: amount of capacity available in shift model w and stage s, (CU)
ft
facility f and period t
WF
dws Workforce parameter: minimal number of employees required to deploy shift model w
ft
in capacity stage s, facility f and period t
Constraints (8), (9), (10), (16) and (17) in the original model are replaced by con-
straints (21) to (26). The suitable shift model for a capacity stage s in a facility f
and in period t is determined in constraint (21). Constraint (22) ensures, that at most
one shift model is chosen for a capacity stage s deployed in time period t and facil-
W F from falling below the
ity f . Constraint (23) prevents the number of employers xws ft
required workforce for a chosen shift model. Constraint (24) determines, how many
employees have to be hired or dismissed to meet the required workforce in period t
and facility f based on the previous time period. Constraints (25) and (26) are the
equivalents to constraints (16) and (17) in the original model.

c KpsBf x ps ft ≤ ∀s, f, t
P SM SM
cws f t yws f t (21)
P W

f t ≤ ys f t ∀s, f, t
SM KA
yws (22)
W
f t ≥ dws f t yws f t ∀w, s, f, t

WF W F SM
xws (23)

ft = xft − x Ff tW F + ∀ f, t
WF HW F WF
xws xws f (t−1) (24)
W S W S

f t yws f t ≤ 1 − ∀s, f, t
SM SM
cws c pAVf y pPfIt csMf K (25)
W P

SM
cws SM
f t yws f t ≤ 1− c Ep fF F y pPfAt csMf K ∀s, f, t (26)
W P
Note, that the extended model is of much greater computational complexity than the
original model. This is mainly due to the additional binary decision variables ywsSM .
ft
Because these variables are situated on the tactical stage of the model, the stochastic
version of the extended model is not amenable to the Benders’ decomposition approach
presented in Sect. 4. Therefore, it is not solvable for practical problem dimensions.
Nevertheless, the extended deterministic model is solvable and can be deployed, e.g.,
to evaluate strategic planning solutions found by the stochastic model, which is pre-
sented in the next Section. A typical scenario of how the different model variants can
be combined in practise is presented in the case study in Sect. 5.
3.3 Stochastic model
To extend the deterministic model of the planning problem by stochastic influences,

we presume that only demand quantities are uncertain with known probability distribu-
tions. Basics of stochastic programming are discussed by Birge and Louveaux (1997)
and Kall and Wallace (1994). The extended model, more precisely the determinis-
tic equivalent model, is formulated as a two-stage stochastic, mixed-integer program
with fixed complete recourse. The first stage decisions on the strategic level remain
unchanged, the second stage decisions are included in the objective function via the
expected value of each scenario’s optimal value Q n . The two stages of the determinis-
tic equivalent model are specified below. Problem (27) represents the strategic (first)
stage, which is formulated as a pure 0–1 problem, containing the decisions about
product allocation and capacity dimensioning. The tactical (second) stage is shown in
problem (28). Only constraint set (29) differs slightly from the demand constraint (11)
by considering a scenario-dependent demand dnpmt . Also, constraints (16) and (17)
are not considered in the stochastic formulation, since they were not needed in our
practical case studies. They can easily be added, if necessary.
Z F S (t)
min Z F Stoch = + ρn Q n y PA
, y KA
(27)
(1 + rt )t
T N
s.t. Constraints (4)–(6), (14), (18)–(19)
y pPfIt , y pPfAt , ysKf It , ysKf tA ∈ {0, 1} ∀ p, f, s, t
with
Z F T (t)
Q n y P A , y K A = min (28)
(1 + rt )t
T

s.t. z Spmt
F
+ x Tp fEmt ≥ dnpmt ∀n, p, m, t (29)
F
Constraints (7)–(10), (12), (13), (15)
xsMf tI N , xsMf tAX , x ps
P , xT I , xT E , zSF
ft p f f t ≥ 0 ∀ f, s, p, m, t
p f mt pmt
Note, that, in order to keep the model computationally tractable, the linear approx-
imation of the organizational capacity adaption is used in the stochastic model instead
of the extension described in Sect. 3.2. Nevertheless, the computation of the formu-
lated stochastic model via common MIP solving algorithms will not lead to acceptable
solution times when applying the model to real world cases in the automotive industry.
However, since the second stage does not contain any binary variables, the model is
amenable to the highly specialized benders decomposition approach, which will be
presented in the next section.
4 Solution method
This section describes an adequate solution method for the formulated two-stage sto-
chastic program. A common algorithm for the optimization of two-stage stochastic
programs is the decomposition scheme proposed by Benders in (1962). This decom-
position scheme is the basis of the proposed algorithm strategy. Hence, the formulated
model is split into a master problem, including all strategic decisions, and into n
sub problems, each of which includes the tactical decisions for one demand scenario.
The anticipated tactical decisions—regarding organizational capacity adaptation—are
incorporated following the linear approximation scheme. This is done, because the lin-
ear approximation of the capacity–cost relationship has just or even no influence on the
strategic decisions compared to the detailed mixed integer formulation of the tactical
planning problem. Still, the detailed calculation of costs and utilization has to be done.
Therefore, after solving the main problem, the shift model planning task is performed
for fixed strategic decisions and every demand scenario.
4.1 Abstract design of the solution algorithm
The first set of the algorithmic strategy is the generation of different demand scenarios.
The user can influence this process by determining a set of generation parameters like
mean value, variance, correlation and life-cycle curves for the products’ demand quan-
tities. The different demand scenarios are computed by a Monte Carlo Method which
involves the calculation of log-normal distributions for each product’s life-cycle and
considers also correlations among the different products. The second step starts the
initialization of the strategic variables by running a heuristic procedure based on a sim-
ple estimation of distribution algorithm for combinatorial problems (for more details
see Larranaga and Lozano 2001). The dependencies between the parameter values are
estimated via the computation of probabilistic distribution models like Bayesian net-
works. The fitness evaluation of each strategic variable realization is done by solving
small linear programs which model the production plan and demand fulfillment in
an approximated way. Subsequently, the applied Benders’ Decomposition approach
operates in usual iterative way by solving of the master problem, solving the subprob-
lems and adding n optimality cuts to the master problem until a stopping criteria is
reached. Since the formulated stochastic model is characterized by a complete fixed
recourse, no feasibility cuts have to be added to the master problem. Algorithm 1
shows the framework of the implemented algorithmic strategy.
Algorithm 1: Abstract algorithmic strategy

Input: strategic network planning task, predicted demand quantities
1 Generating n demand scenarios via MonteCarlo Simulation;
2 Initialization of strategic variables;
3 while Stopping criteria is not met do
4 Solve Master problem;
5 Solve n Subproblems including linear approximation scheme;
6 Add n Optimality Cuts to Master;
7 Store optimal values of strategic decisions;
8 Solve mixed integer program of tactical shift planning for fixed strategic decisions;
Output: optimal flexibility strategy and capacity plan
4.2 Applied Benders’ decomposition
For the present formulation of the deterministic equivalent model, the two-stage sto-
chastic program is separated into two models: the strategic and tactical level. In order
to reduce processing time, the model is implemented in the multi-cut version (Birge
and Louveaux 1988). Following the approach of MirHassani et al. (2000) the master
model includes the tactical decisions of one specifically selected scenario to increase
the content of information—especially in the first iteration steps of the decomposition
scheme. Thereby, the master or strategic level model is extended only by |N | − 1
continuous variables Θn that represent a lower bound to the tactical level objective
for each scenario representation |N | − 1. The |N | − 1 tactical decisions are sepa-
rated into |N | − 1 subproblems. The decomposition scheme iterates the solving of
the extended master model, the solving of the |N | − 1 subproblems, the calculation
of the cut coefficients and the addition of the resulting optimality cuts until a defined
stopping criteria is met. The stopping criteria applied to the present approach depends
on the difference between lower bound and upper bound of the computed objective
value in each iteration. The master model in iteration i is given in 30 and 31.
Z F S (t)
min Z F Stoch = + (30)
(1 + rt )t
T

Q n̂ y P A , y K A + ρn Θn
N \{n̂}
s.t. Constraints (4)–(6), (14), (18)–(19)
y pPfIt , y pPfAt , ysKf It , ysKf tA ∈ {0, 1} ∀ p, f, s, t

α n,k
pf t ypf t +
PA
βsn,k KA
f t ys f t
T F P T F S
≤ Θn − γ n,k ∀n ∈ (N \n̂) ∀k = 1, . . . , i (31)
In each iteration the current best solution of the strategic variables is stored and is
passed to the subproblems. On this basis the dual solutions of the remaining |N | − 1
subproblems are computed. Equations (32) to (34) show the calculation of the coef-
ficients α n,k n,k
p f t , βs f t and γ
n,k for the optimality cuts applied to the extended master
problem for the n-th scenario after the k-th iteration. Before the first iteration the coef-
ficients are initialized with 0. The calculation makes use of vectors of dual variables δ,
, ζ , η, κ, λ and µ associated (in this ordering) with the subproblem restrictions (29),
(7)–(10) and (15). Constraint (15) is associated with two vectors of dual variables λ
and µ. The dual variables of the constraints (12) and (13) are not used.

α n,k
pf t = cM K n,k−1
f ps f t (32)
S

βsn,k
ft = 1 − dsKf M I N csKf κsn,k−1
ft + dsKf M AX csKf ηsn,k−1
ft + csKf ζ n,k−1
ft (33)

n,k−1 P M I N n,k−1 P M AX n,k−1
γ n,k = dnpmt δ pmt + dpf t λpf t + dpf t µpf t (34)
T P M F
As proposed by Santoso et al. (2005) and Wentges (1996), acceleration techniques

are applied to the decomposition scheme. Two of the presented techniques, the con-
cept of trust region and the upper bounding heuristic, are added because they showed
the greatest positive influence on processing time for our model formulation. In the
first iterations of the Benders’ Decomposition a bad convergence behavior can appear,
because the calculated solutions of the master problem are oscillating wildly in the
solution space. By avoiding this oscillation, a more effective convergence behav-
ior can be achieved. The main idea of the trust region method is to limit the gap
between sequenced solutions explicitly, so that randomness of the search direction
can be reduced (Ruszczyński and Shapiro 2003). In our case the trust region method
can be further enhanced. This is due to the fact, that on the one hand the strategic
decisions are explicitly split into product allocation and capacity initialization. On
the other hand, due to the multi-period nature of the problem
the trust region con-

cept has to be applied in a different way. Let YiP I := ( p, f )| T y pPfIt = 1 and

YiK I := (s, f )| T ysKf It = 1 be sets of indices of the i-th iteration. By introducing
the constraint

1 − ypf t
PI
+ PI
ypf t ≤ ∆iP I (35)
( p, f )∈YiP I T ( p, f )∈Y
/ iP I T

1− ysKf It + ysKf It ≤ ∆iK I (36)
(s, f )∈YiK I T (s, f )∈Y
/ iK I T
in the master problem of the (i + 1)-th iteration, it is guaranteed, that the computed
solution is bounded within a Hamming distance of ∆iK I and ∆iP I , respectively, to the
previous solution. To apply a limitation within the first iterations the decomposition
algorithm is started with ∆iP I < |F| × |P| and ∆iK I < |F| × |S|. Furthermore, in the
process of the method the ∆’s are raised with a given rate R∆ to
∆i+1
PI
= ∆iP I + (|F| × |P|)R∆ (37)
∆i+1
KI
= ∆iK I + (|F| × |S|)R∆ (38)
until the constraints (35) and (36) redundant. Thereby a raise of ∆ is performed when
the current objective function value lbi of the master problem is near to the current
maximum objective function value lbmax so that for a given rate Rlb
|lbi − lbmax |
< Rlb (39)
lbmax
is fulfilled. If (39) is fulfilled for the ∆i ’s, the method is defined as stable and the
trust region constraint is enlarged. During the operation of Benders’ decomposition
the value of the upper bound ub decreases with an ever smaller rate. When the opti-
mality gap is small, the method performs numerous iterations, in which the upper
bound improves only insignificantly. This performance problem can be attributed to
inc
the calculated optimal flexibility decisions y pPfIt of the master problem, which differ
inc
only slightly. Corresponding capacity decisions ysKf It variate more but have an even
smaller influence on the resulting objective function value. To avoid these inefficient
iterations the following heuristic procedure by Santoso et al. (2005) can be deployed,
whose problem-specific adaption is described in algorithm (2).The heuristic considers
Algorithm 2: Upper bounding heuristic

Input: parameter N sub where N sub << N
P I inc P Ainc
1 Fix best solution found so far y p f t and y p f t ;
2 Solve Deterministic equivalent model with N sub scenarios;
K I uh and y K Auh ;
3 Store optimal solutions ys f t
s f tinc
= PI P Ainc K I uh K Auh ;
4 Compute objective value z f N y p f t , y p f t , ys f t , ys f t

P I cur P Acur K I cur , y K Acur then
5 if z < f N y p f t , y p f t , ys f t sf t
cur inc cur inc
6 update y pPfIt ← y pPfIt and y pPfAt ← y pPfAt ;
cur uh cur uh
7 update ysKf It ← ysKf It and ysKf tA ← ysKf tA ;
8 if z < ub then
9 update ub ← z;
inc uh
10 update ysKf tA ← ysKf tA ;
a small number N sub of the N scenarios. In the first step the best product allocation
found up to this point is fixed. In the second step the reduced deterministic equivalent
model will be solved for N sub scenarios, resulting in presumably suboptimal capac-
uh uh
ity decisions ysKf It and ysKf tA . For the combined fixation of these flexibility- and

capacity decisions, the objective function value
inc inc uh uh
f N y pPfIt , y pPfAt , ysKf It , ysKf tA of
the problem for N scenarios is computed in the third step. Furthermore, the objective
Table 7 Main characteristics

Property Quantity
Product variants 16
Facilities 8
periods 5
Technical capacity stages 2
Binary variables 1,440
Continuous variables 1,520
Equalities 1,480
Inequalities 13,520

function value
cur cur cur cur
f N y pPfIt , y pPfAt , ysKf It , ysKf tA of the current solution of the mas-
inc inc uh uh
ter problem is determined. If the solution y pPfIt , y pPfAt , ysKf It , ysKf tA is better, then
this one will be fixed as current solution of the decomposition algorithm. In case that
this solution is additionally a lowest upper bound, ub is updated and the solution is
stored.
5 Numerical results and case study
In this section, we evaluate our modeling and solution approach regarding computa-
tional efficiency and applicability to real-world problems. First, the performance of
the acceleration techniques is measured by applying it to an extended version of an
abstract problem formulated by Jordan and Graves (1995). Secondly, we compare the
two implemented capacity adaptation schemes, the linear approximation scheme and
the detailed workforce planning, for a small near-real problem instance. Finally, the
results of a pilot application to a real-world problem in the automotive industry are
discussed. The outcomes of the solution approach are compared to a flexibility and
capacity strategy developed by planners.
5.1 Performance tests
The performance tests are conducted on an extension of the abstract strategic net-
work design problem proposed by (Jordan and Graves 1995, p. 585). It is extended to
multiple time periods with distinctive life-cycles and different cost parameters for pro-
duction, transport and shortfall at each facility. In addition, the flexibility options are
not restricted, but associated with different investment levels. Table 7 shows the main
characteristics of the deterministic formulation of the test problem. Demand quantities
are computed on the basis of a logarithmic normal distribution with given parameters,
i.e., the expected value and deviation ranges.Our solution approach is implemented
in JAVA using the concert technology libraries of CPLEX 10.0 for solving the lin-
ear and mixed integer programs within the decomposition scheme. To compare the
performance with common solution approaches, we implemented a complete formu-
lation of the deterministic equivalent model within this framework. The experiment
35.000
standard
decomposition
30.000
solution approach
CPU Time in seconds
25.000
20.000
15.000
10.000
5.000
0
0 50 100 150 200 250
Scenarios
Fig. 1 Solution time
computations were run on a Pentium M 1.7 GHz PC with 1.0 GB RAM running Win-
dows XP. To analyse the performance of our approach, we compared the results of
different sample sizes to the results of the complete deterministic equivalent formula-
tion. A solution is regarded as optimal if the gap between lower and upper bound is less
than 0.1%. Figure 1 shows the results of our solution approach (solution approach)
compared with a standard Benders’ Decomposition scheme (decomposition) as well as
with the complete deterministic equivalent formulation (standard) with sample sizes
from 5 to 200 scenarios. The standard Benders’ Decomposition scheme is already
advantageous compared to the deterministic equivalent formulation for a sample size
greater than 20 scenarios. However, the standard approach is impractical for more
than 50 scenarios, the solution time would be greater than 10 h. In total our solution
approach shows a significant superiority under the given circumstances.Regarding the
implemented techniques, different combinations were examined to achieve an optimal
setting which performs best on different problem instances. The results in Fig. 2 show,
that there is no technique which dominates all of the other techniques with respect
to performance. Single deployments of the various acceleration techniques are sig-
nificantly outperformed by combinations of them. The best combination for both the
abstract problem and the real-world instance seems to be the combination of multi-
cutting, trust region and expansion of the master model by a specially chosen subset
of the subproblem.
5.2 Comparison of capacity adaptation schemes
The first case study deals with the anticipation scheme for the tactical workforce plan-
ning. It is based on a small real world problem instance. The target is to determine
a cost-efficient network structure of two products, three possible plants each with
25000
benders decomposition
22500 Expanded Master(EM)
Trust Region(TR)
20000 Upper Bounding Heuristic(UB)
CPU Time in seconds
StartValues(SV)
17500 Multi Cut (MC)
EM+TR+MC
15000
12500
10000
7500
5000
20 40 60 80 100 120 140 160 180 200 220
Scenarios
Fig. 2 Solution techniques
Demand Quantities
1.600
Demand per Period
1.400
1.200
1.000
800
600
400
200 Product A
Product B
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Period
Plant 1 Plant 2 Plant 3

Stage 1 Stage 2 Stage 3 Stage 1 Stage 2 Stage 3 Stage 1 Stage 2 Stage 3
Capacity in regular working time (c^K) 700 725 750 400 425 450 700 1000 1400
Capacity based invest (k^KI) 4100 4200 4400 4800 4900 5000 3200 3600 6400
Capacity based fixed costs (k^KF] 385 390 395 430 440 450 520 630 1000
Input Data for linear
Minimal relative capacity reduction (c^KMIN) 33% 33% 33% 33% 33% 33% 33% 33% 33%
approximation Maximal relative capacity increase (c^KMAX) 104% 104% 104% 104% 104% 104% 104% 104% 104%
Capacity reduction costs (k^MIN) 0,10 0,10 0,10 0,45 0,43 0,42 0,25 0,23 0,17
Capacity increase costs (k^MAX) 0,07 0,07 0,07 0,60 0,60 0,60 0,10 0,09 0,07
Number of employees per shift (d^WF) 250 260 270 200 225 250 200 250 400
Wage per employee (c^WF) 0,55
Wage per temporary worker 0,60
Input Data for detailed
Hiring cost per employee (c^HWF) 0,01
Workforce planning Dissmissial Cost per employee (c^FWF) 0,55
Shiftmodel bonus (r^SM) late shift (15%), night shift (20%), Saturday shift (30%), Saturday late shift (45%)
maximal amount of temporary workers 30%
Fig. 3 Input data for workforce planning
three possible capacity expansion stages for one life-cycle with half-year periods. We
only consider investments and labor fixed costs. Figure 3 shows the possible networks
structure and the related input data, for confidentiality reasons the problem data has
been abstracted.
While the linear approximation model needs an ex ante calculation of capacity fixed
costs and adaptation costs, in terms of labor costs, the detailed workforce planning
model explicitly incorporates decisions of hiring and dismissing workers, considers
Capacity Load Plant 1 (linear Approximation) Capacity Load Plant 1 (Detailed Workforce Planning)
800 800
700 700
Production Product B
Production Product A
600 600 Capacity (Regular WorkTime) Plant 1
Capacity (Organizational) Plant 1
500 500
400 400
300 300
200 200
100 100
0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Period Period
Fig. 4 Comparison of Capacity Adaptation (Example: Plant 1)
Table 8 Comparison of workforce approximation schemes
Linear approximation Detailed workforce planning
Product Allocation Plant 1 A, B A, B

Product Allocation Plant 2 – –
Product Allocation Plant 3 A A
Capacity Plant 1 700 715
Capacity Plant 2 0 0
Capacity Plant 3 1,000 1,020
Product specific invests 3,550 3,550
Capacity specific invests 7,700 8,800
Labor fixed costs 14,211 13,671
different shift models and the option of employing workers just temporarily. The results
show that the strategic decisions are mostly identical. Applying the detailed workforce
planning just leads to a minor increase in technical capacity initialization in case of
the higher shift premium for Saturday. Figure 4 presents the difference in capacity
load for the product-flexible Plant No. 1. While the linear approximation model is try-
ing to utilize the technical capacity, the detailed workforce planning is visualizing an
almost realistic utilization of the employed workforce.If the whole network structure
is considered, however, these different planning objectives, as implemented within the
different approximation schemes, lead to the same strategic decisions with nearly the
same costs (Table 8).
5.3 Application to real-world problem
The real-world example consists of a network of three existing manufacturing plants.

Regarding the abstract problem of the former subsection, it could be seen as a more
detailed planning task on the same network structure. To be more realistic we consider
additional planning restrictions and parameters. In this case, there are no decisions
about shutdown of existing or opening of new facilities. The production process is
divided into several stages which have to be almost completely carried out at each
location due to “just in sequence” concepts and high internal logistics rates. Never-
theless, flexibility strategies have to be incorporated since every location consists of
Table 9 Concept comparison based on certain demand
Performance indicator Planners’ strategy Computed strategy
Additional links in BIW 0 6

Technical capacity utilization (%) 82 87
Demand fullfilment (%) 100 100
Investment costs (GE) 1,080 1,109
Operating costs (GE) 1,210 1,180
Net present value (GE) 988 990
several production lines which can be configured in their degree of flexibility and tech-
nical capacity. On the other hand pooling concepts for body in white components like
hoods or doors could be implemented. The importance of successor flexibility is due
to the fact that the planning horizon for three car models with up to three main variants
each covers two product life-cycles. The emphasis of the study is on the body shop,
since the potential of flexibility in this section is strongly linked to high investment
requirements and a high degree of automation. In the sections paint shop and final
assembly, the decisions are solely concentrated on the installation of suitable capaci-
ties as the flexibility strategies are supposed to be already dignified by the decisions in
the body shop. Reasons for this are the system immanent flexibility (paint shop) and
the high share of manual processes (final assembly). Five possible technical capacity
stages with associated investment requirements, as well as cost rates for the estimated
demand quantities, were determined for each production line. The adaptation of capac-
ity via organizational actions results from the opportunity to work in ten to seventeen
shifts per week. The fixed costs related to varying shift models were approximated by
assuming a proportional relationship between organizational capacities and cost rates.
Shift model changing costs were neglected due to their minor influence on total cost
which was a result of preceding numerical studies. Finally up to 100 scenarios were
computed out of the estimated demand quantities. These scenarios were determined by
a plain monte carlo scheme. We will compare the solution of the optimization model
to a configuration designed by the company’s planning experts. This configuration
results by calculating the technical capacities on the basis of the estimated demand
quantities. The flexibility strategy resulting form this procedure is mainly characterized
by solitary production lines. The fixed costs related to the choice of the shift models
are computed by a small sized linear program, since this planning exercise itself is
too complex for manual calculation. Comparing the flexibility and capacity strategy
of this manual calculation (further on called planners’ strategy) with the solution of
a deterministic optimization approach (computed strategy) based on certain demand
scenarios, a slight advantage of the solitary strategy is recognizable (cf. Table 9). This
advantage is due to two effects. Firstly, just a slightly higher optimal system capacity
is installed for the solitary strategy. Secondly, a minor degree of flexibility to fulfill the
estimated demand is implemented, only the several variants for each product model
are manufactured on flexible production lines. These two aspects result in a better
net present value by lower investment requirements. Considering uncertain demand
Table 10 Concept comparison based on demand deviation
Performance indicator Planners’ strategy Computed strategy
Average technical capacity utilization (%) 82 87

Average demand fullfilment (%) 99.23 99.64
Average shortfall per year 1,473 422
Investment costs (GE) 1,080 1,109
Average operating costs (GE) 1,310 1,241
Average shortfall costs (GE) 85 21
Average net present value (GE) 1,021 1,006
Standard deviation of npv (GE) 2.7 2.3
quantities with deviations up to 40 percent, the computed strategy is getting more

and more advantageous (see Table 10). The computed strategy is able to cover higher
demand quantities due to a higher ‘free’ system capacity. Therefore shortfall costs are
quite low. Scenarios where volumes are shifted from one product model to another,
lower fixed costs can be achieved by usage of the flexibility potentials. On the other
hand the planners’ strategy is linked to higher shortfall costs in the high volume sce-
narios. In the product mix shifting scenarios the planners’ strategy is characterized by
low utilization and therefore by higher fixed costs.
Including uncertainty into the comparison of the two strategies, different approaches
can be chosen. Well-tried indicators for the analysis considering uncertainty are, e.g.,
the mean value for the estimated returns or the variance for risk measurement. Regard-
ing the real-world-problem the average net present value is taken as an equivalent to
the mean value. Now the cost advantage is turning round towards the computed strat-
egy. Nevertheless indicators like the mean value or standard deviation are mostly too
aggregated to achieve a perfect reliability in the advantage of one strategy. New ways
to visualize the advantages of flexible strategies have to be found.
6 Conclusions
In this paper, we have proposed mathematical formulations of strategic network design

problems under uncertain demands for the automotive industry from a capacity and
production planning perspective. To emphasize the necessity of anticipating consec-
utive stages in a hierarchical planning process, the formulations also include tactical
aspects such as workforce planning. Additionally, customized solution approaches
have been implemented based on Benders’ Decomposition. The numerical results
show that the solution approach greatly decreases the solution time in comparison to
standard methods. Furthermore, we have demonstrated that these methods can handle
large-scale, real-world problems and lead to better decisions than widely accepted
methods for actual planning problems in the automotive industry. Continued research
is necessary to study in detail the effect of tactical decisions in strategic network design
problems. Moreover, the models must be extended to support calculation of strategic
supply and transport problems which occur in flexible production networks.
References
Alonso-Ayuso A, Escudero LF, Garfín A, Ortuño MT, Pérez G (2003) An approach for strategic supply
chain planning under uncertainty based on stochastic 0–1 programming. J Global Optim 26(1):97–124
Arntzen BC, Brown GG, Harrison TP, Traffton LL (1995) Global supply chain management at digital
equipment corporation. Interfaces 25(1):69–93
Askar G, Zimmermann J (2006) Optimal usage of flexibility instruments in automotive plants. Oper Res
Proc, pp 479–484
Barnett L, Leaney PG, Matke J (1995) Striving for manufacturing flexibility in body construction. In:
Proceedings of the IMechE AutoTech/95 Conference
Beamon BM (1998) Supply chain design and analysis: Models and methods. Int J Prod Econ 55(3):281–294
Becker H (2005) Auf Crashkurs. Automobilindustrie im globalen Verdrängungswettbewerb. Springer,
Berlin
Benders JF (1962) Partitioning prodedures for solving mixed-variables programming problems. Numer
Math 4:238–252
Birge JR, Louveaux F (1988) A multicut algorithm for two-stage stochastic linear programs. Euro J Oper
Res 34(3):384–392
Birge JR, Louveaux F (1997) Introduction to stochastic programming. Springer Series in Operations
Research. Springer, New York
Boettcher J (1989) Stochastische lineare Programme mit Kompensation, Mathematical systems in econom-
ics. vol 115. Athenäum, Frankfurt am Main
Boyer KK, Leong GK (1996) Manufacturing flexibility at the plant level. Omega Int J Manage Sci
24(5):495–510
Bruynesteyn M (2003) Flex appeal. Prudential Equity Group, LLC
Chandra C, Everson M, Grabis J (2005) Evaluation of enterprise-level benefits of manufacturing flexibility.
Omega Int J Manage Sci 33(1):17–31
Cordeau SF, Pasin F, Solomon MM (2006) An integrated model for logistics network design. Ann Oper
Res 144(1):59–82
Ferber S (2005) Strategische Kapazitäts- und Investitionsplanung in der globalen Supply Chain eines Auto-
mobilherstellers. Shaker, Indianapolis
Fleischmann B, Ferber S, Henrich P (2006) Strategic planning of BMW’s global production network. Inter-
faces 36:194–208
Francas D, Kremer M, Minner S, Friese M (2007) Strategic process flexibility under lifecycle demand. Int
J Prod Econ. doi:10.1016/j.ijpe.2006.12.062
Friese M, Bihlmaier R, Buerkner S (2005) Planning of flexible production networks in the automotive
industry. In: International conference on changeable, agile, reconfigurable and virtual production,
Muenchen, 22–23, September
Geoffrion AM, Graves GW (1974) Multicommodity distribution system design by benders decomposition.
Manage Sci 20(5):822–844
Gerwin D (1982) Do’s and don’ts of computerized manufacturing. Harvard Bus Rev 60(2):107–116
Goetschalckx Strategic Network Planning. BT and Centre for Automotive Industry Research, Cardiff (2001)
Goetschalckx M (2002) Strategic network planning. In: Stadtler H, Kilger C (eds) Supply chain management
and advanced planning. Springer, Heidelberg, pp 105–122
ILOG Inc. ILOG CPLEX 10.0 Advanced reference manual. ILOG CPLEX Division, France, 2006
Jordan WC, Graves SC (1995) Principles on the benefits of manufacturing process flexibility. Manage Sci
41(4):577–594
Kall P, Wallace SW (1994) Stochastic programming. Wiley, Chichester
Larranaga P, Lozano JA (2001) Estimation of distribution algorithms: a new tool for evolutionary compu-
tation. Springer, Heidelberg
Meyer B (2004) Value-adding logistics for a world assembly line. Bonifatius Verlag, Paderborn
MirHassani SA, Lucas C, Mitra G, Messina E, Poojari CA (2000) Computational solution of capacity
planning models under uncertainty. Parallel Comput 26(5):511–538
Pibernik R (2001) Flexibilitätsplanung in Wertschöpfungsnetzwerken, 1 edn. Dt. Universitäts-Verlag,
Wiesbaden
Roscher J (2008) Bewertung von Flexibilitätsstrategien fuer die Endmontage in der Automobilindustrie.
PhD Thesis, Institut für Industrielle Fertigung und Fabrikbetrieb, Universität Stuttgart
Ruszczyński A, Shapiro A (2003) Stochastic Programming, Handbooks in Operations Research and Man-
agement Science. vol 10. Academic Press, Australia
Santoso T (2003) A comprehensive model and efficient solution algorithm for the design of global supply
chains under uncertainty. PhD Thesis, Georgia Institute of Technology
Santoso T, Ahmed S, Goetschalckx M, Shapiro A (2005) A stochastic programming approach for supply
chain network design under uncertainty. Euro J Oper Res 167(1):96–115
Schneeweiss C (1999) Hierarchies in distributed decision making. Springer, Berlin
Sethi A, Sethi S (1990) Flexibility in manufacturing: a survey. Int J Flexible Manufact Syst 2(4):289–328
Vidal CJ, Goetschalckx M (2000) Modeling the impact of uncertainties on global logistics systems. J Bus
Logistics 29(1):95–120
Wentges P (1996) Accelerating bender’s decomposition for the capacitated facility location problem. Math
Methods Oper Res 44:267–290

Supply Chain Planning

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Supply Chain Planning

Uploaded by

Copyright:

Available Formats

Supply Chain Planning

Hans-Otto Günther • Herbert Meyr

Supply Chain Planning

ISBN 978-3-540-93774-6 e-ISBN 978-3-540-93775-3

Library of Congress Control Number: 2008943995

© 2009 Springer-Verlag Berlin Heidelberg

Cover design: WMXDesign GmbH, Heidelberg, Germany

Printed on acid-free paper

special emphasis on model-based decision support at the operational level of inter-

multi-item multi-stage inventory system. As a case study, the supply chain of a

Berlin and Darmstadt, Hans-Otto Günther

Part I: Decision Making in Supply Chains

Simulation and optimization of supply chains: alternative

Part II: Demand Management

Active demand management for substitute products through price

Customer segmentation, allocation planning and order promising

Managing product availability in an assemble-to-order supply

Inventory reservation and real-time order promising

Part III: Inventory Management

Supplier managed inventory in the OEM supply chain: the impact

Vendor-managed inventory and the effect of channel power ........................ 247

Part IV: Applications in the Chemical Industry

MILP-based campaign scheduling in a specialty chemicals plant:

Part V: Applications in the Automotive Industry

Modeling and optimizing of strategic and tactical production planning

Originally published in:

Abstract Inter-organizational supply chain management incurs the challenge to

Keywords Collaborative planning · Supply chain management · State-of-the-art

1 Introduction and definitions

Supply chain management (SCM) is concerned with the coordination of material,

H.O. Günther, H. Meyr, Supply Chain Planning 3

of collaborative planning (CP) together with a state-of-the-art overview of concepts

2 Related research areas

While in the original model collaboration is restricted to mere information

3 SC structure and relationships among SC members

3.1 Criteria (1.1): Structural elements of the SC

For defining the structure of a SC we consider

business function at different echelons in an organization this is termed multi-plant

3.2 Criteria (1.2): Relationships among SC members

The relationships among SC members comprise the following issues:

• when designing a SC and fixing the conditions of a collaboration strategy (note,

3.3 Criteria (1.3): Requirements for a SC solution

4 Characteristics of the decision situation of each SC member

4.1 Criteria (2.1): Decision models—which decisions take place?

level supply menu

skeleton negotiations adaptations renegotiate or

4.2 Criteria (2.2): Phases of collaboration—when does collaboration take place?

4.3 Criteria (2.3): Information status—what is the information status of each SC

The information status in an inter-organizational collaboration is assumed to be asym-

5 Characteristics of collaborative planning schemes

3.1. Mediator – – yes yes – yes – – – –

6 The notion of fairness

– minimizing (the sum of) unfairness across members,

7 Summary and outlook

Fig. 2 Summary of characteristics considered in CP approaches analysed

Table 5 Criteria for discriminating SC structures, relationships and decision situations

1.1 Structural elements

Christian Almeder · Margaretha Preusser ·

Originally published in:

Abstract Discrete-event simulation and (mixed-integer) linear programming are

Keywords Supply chain management · Optimization · Discrete-event simulation ·

C. Almeder (B) · M. Preusser · R. F. Hartl

H.O. Günther, H. Meyr, Supply Chain Planning 29