144 154 - GrowthModels Ver1 - tcm28 36191
144 154 - GrowthModels Ver1 - tcm28 36191
net/publication/242636096
CITATIONS READS
19 1,917
1 author:
Mladen Sokele
Zagreb university of applied sciences
27 PUBLICATIONS 118 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Determination of time criteria for assessment in Learning Management Systems View project
Research on the importance of blended learning components for student exam success View project
All content following this page was uploaded by Mladen Sokele on 30 October 2015.
Appropriate forecasting of product market adoption enables optimal planning of resources, investments,
revenue, marketing and sales. Quantitative forecasting methods for the new product adoption rely on
the S-shaped (sigmoidal) growth model such as the logistic, the Richards and Bass growth models. This
paper presents adaptations of these models by introducing explanatory and marketing variables which
are suitable for the forecasting prior to product launch or in the early phases of the product life cycle.
Mladen Sokele
is Assistant
1 Introduction N(ti) – f (ti; α1, α2, ..., αk) = 0, i = 1, ..., k (1)
Director in
Croatian Tele- Growth models are widely used in quantitative
com Inc, Zagreb research in order to understand the forces that influ- System (1) is usually a nonlinear system, so iterative
ence growth in the sense of its dynamics, market numerical methods for its solution are needed to be
capacities as well as forecasting of growth in the performed (eg. Newton method).
future. Particularly, diffusion of innovation and new
technology, market adoption of consumer durables, In cases when k or more data points are available, the
number of users of subscription services (for exam- weighted least squares method can be used for
ple: telecommunications services) and allocation of parameter estimation to adjust the parameters of a
restricted resources at the beginning of a product life model so as to best fit a data set. The objective is to
cycle (PLC) have sigmoidal growth. Such products/ minimize the sum of squared difference between data
services that do not include repeat sales are denoted points and model evaluated points:
products, and the cumulative volume of adopted n
2
products by users, number of users. S= wi · [N (ti ) − f (ti ; α1 , α2 , ..., αk )] (2)
i=1
Growth models based on the logistic, Richards and where wi are weights. When weights are equal to 1
Bass models involving explanatory and marketing (wi = 1), the method is called ordinary least squares
variables will be presented and examined for the fore- method (OLS).
casting application in cases of new product market
adoption. Minimisation of (2) can be done by software tools
such as Microsoft Excel solver. Analytically, values
of parameters are resulting from solution of the sys-
2 Growth Forecasting tem of equations (3):
Growth forecasting relies on the basic principle: The
∂S
growth model will be valid in the perceivable future, = 0, j = 1, ..., k (3)
∂αj
and the forecasting result could be obtainable by
extrapolation of observed values sequentially through By use of least squares method, obtained values for
time and supplementary information. In general, this parameters are statistically smoothed; ie. the influ-
principle is valid only for stable markets where inter- ence on parameter values due to particular measure-
nal forces remain the same (eg. same market segment ment errors (such as unanticipated seasonal variation,
boundaries, competition, cause-and-effect among uncertain measure, etc) is reduced.
products, etc) and without change of external influ-
ences (eg. technology, macroeconomics, purchasing For forecasting purposes, parameter estimation
power, regulatory, etc changes). This type of fore- is usually focused on the time interval near the last
casting belongs to the quantitative time series meth- observed data point. Thus, weights in equation (2)
ods [1, 11]. can be set to higher value for the most recent data
points than for data points in far history. For example,
For time series growth model f (ti; α1, α2, ..., αk) the geometric series for weights
based on a set of k parameters {αi}, at least k known
1
data points (ti; N(ti)) are needed for the complete wi = , q>1 (4)
q n−i
parameter estimation. In cases when exactly k data
points are available, parameters αi are solutions of
a system of equations (1):
100 100
% %
a=A a=A
75 a = -A 75 a = 0.5•A
50 50
25 25
0 0
B-15
B-10
B-5
B+5
B+10
B+15
B-15
B-10
B-5
B+5
B+10
B+15
100 100
% %
b=B M= MC
75 b = B-5 75 M = 0.7•MC
50 50
25 25
0 0
B-15
B-10
B-5
B+5
B+10
B+15
B-15
B-10
B-5
B+5
B+10
B+15
t d2 L(t)
(eaΔ -1)/2 L (t) = =
dt2
aM/4
−a2 M · e−a(t−b) · 1 − e−a(t−b)
= 3 (10)
0 b
Time 1 + e−a(t−b)
From (10) follows that L(t) has inflexion for tI = b,
Figure 2 Characteristic values and points of the which is for a > 0 the maximum of L'(t), too (see
logistic growth model Figure 2):
aM
max L (t) = ⇔ a > 0; t = b (11)
4
3 Logistic Growth Model
The logistic model L(t) describes growth in the num- The value of the logistic model at point of inflexion is
ber of users observed over time in a closed market, (see Figure 2) L(b) = M / 2.
without the impact of any other product. The model
is defined by three parameters: M – market capacity, Accordingly, maximum of sales occurs at t = b = tI
a – growth rate parameter and b – time shift parame- when penetration is 50 %. For t1 and t2 near b from
ter. To emphasise the model’s dependence on its (9) follows that sales in time interval [t1, t2] can be
parameters, it is convenient to indicate the model as approximated by:
L(t; M, a, b) [11]:
aM
max Sales(t1 , t2 ) ≈ (t2 − t1 ) · (12)
M 4
L(t; M, a, b) = L(t) = (6)
1 + e−a(t−b) Asymptotes of logistic growth for positive and nega-
Figure 1 shows the effects of change of parameters a, tive parameter a can be summarized in (13):
b and M on the form of S-curve. L(t) = 0 ⇔ a > 0
lim
t→−∞ L(t) = M ⇔ a < 0
3.1 Characteristics of the Logistic Growth
L(t) = M ⇔ a < 0
Model lim (13)
t→+∞ L(t) = 0 ⇔ a > 0
The logistic model is a widely used growth model
with many useful properties for technological and The growth rate GR for time interval Δ t is:
market development forecasting. Model (6) is the
L(t) − L(t − Δt)
solution of differential equation (7) which consists of GRΔt = =
L(t − Δt)
exponential growth term and negative feedback term
[5]. In the beginning, the growth of the logistic model 1 + e−a(t−Δt−b) (14)
= −1
is identical to exponential growth, but later negative 1 + e−a(t−b)
feedback slows the gradient of growth as L(t) app- For positive a, the growth rate is always positive and
roaches the market capacity limit M: the maximum of growth rate is when t –> – ∞ (see
Figure 2, grey columns):
dL(t) L(t)
= aL(t) · 1 −
dt M (7) max GRΔt = eaΔt − 1 ⇔ a > 0; t → −∞
←→ ←→
Exponential Negative
growth feedback When t = b = tI , a sales peak occurs. The growth rate
First derivative of L(t) is given in (8): at this point in time is half of its maximum value:
Therefore model L(t; M, ts, Δ t, u) has only one auxil- Local logistic model is useful for forecasting from the
iary parameter u: last observed point t > tp. The idea is that it is better
to start forecasting from a known base rather than to
M
L(t; M, ts , Δt, u) = 1 1−2(t−ts )/Δt (20) rely on an anticipated but un-modelled reversion to a
1+ u −1 historical trend. [4]
u = 5 %, v = 95 % u = 10 %, v = 90 %
M M
Δt = 2 years N (t) = N (t) =
1+ e−2.944(t−ts −1) 1+ e−2.197(t−ts −1)
M M
Δt = 5 years N (t) = N (t) =
1+ e−1.178(t−ts −2.5) 1+ e−0.879(t−ts −2.5)
M M
Δt = 10 years N (t) = N (t) =
1 + e−0.589(t−ts −5) 1 + e−0.439(t−ts −5)
M M
Δt = 15 years N (t) = N (t) =
1 + e−0.393(t−ts −7.5) 1+ e−0.293(t−ts −7.5)
Note: Characteristic duration Δt according to [6] can be assumed as follows: products consist of units sold that have a typical life cycle
of 6 to 10 quarters; product families consist of related products that have a typical business cycle of 5 years, and basic technologies
consist of a set of related product families that have a typical cycle of 10 to 15 years.
Table 1 Logistic model framework for forecasting of new product adoption prior to launch
50 50
25 25
0 0
ts ts+Δt ts ts+Δt
(27)
100 100
% %
75 75
50 50
25 25 p=0.053, q=0.473
p=0.026, q=0.236
p=0.139, q=0.015 p=0.279, q=0.031
0 0
ts-5
ts+5
ts+10
ts+15
ts
ts+20
ts+25
ts-5
ts+5
ts+10
ts+15
ts
ts+20
ts+25
100 100
% %
75 75
50 50
25 25 p=0.008, q=0.747
p=0.004, q=0.374
p=0.298, q=0.003
p=0.149, q=0.002
0 0
ts-5
ts+5
ts+10
ts+15
ts+20
ts+25
ts+5
ts+10
ts+15
ts+20
ts+25
ts
ts-5
ts
Figure 5 Effects of different values of parameters p and q. Chosen values are explained in Section 4.2
ts+10
tI
ts
ts+20
ts+10
tI
ts+20
ts
(q-p)
M 2q
I
q≤p q>p
B(t) B(t)
dB(t)/dt dB(t)/dt
In cases when q < p, the inflexion point and maxi- dependent while shape Bass model S-curve. Namely,
mum of B'(t) occurs before the product starts (tI < ts), the value of a characteristic duration of a product is
and the value of the Bass model at that point is nega- provided only indirectly through the values of p and
tive according to (34), therefore interior maximum of q parameters (see Figure 5).
B'(t) occurs at t = ts . Similarly, in cases when q = p,
inflexion point and maximum of B'(t) occurs when The idea is to replace p and q with two independent
the product starts (tI = ts). For q > p, the sales peak explanatory parameters: a parameter that describes
occurs in conventional sense of a PLC (tI > ts). The the vertical shape of the S-curve s and characteristic
above mentioned is summarized in (35): duration (time to reach certain penetration level mea-
2 sured from ts) Δt. Expected penetration level at time
M (p+q) ⇔ q > p, t = tI
max B (t) = 4q (35) point ts + Δ t is v (see explanation for Δt and v in Sec-
M p ⇔ q ≤ p, t = ts
tion 3.2).
Accordingly, a maximum of sales occurs when pene-
tration is (q – p) / 2q in cases when q > p (at t = tI), The shape parameter s is chosen in order to encom-
and in cases when q≤ p, maximum of sales occurs at pass the relation between the amplitude of the posi-
t = ts when penetration is 0, which is summarized in tive S-curve part and the amplitude of the negative
(36): S-curve part. Asymptotes of the Bass model are:
max Sales(t1 , t2 ) ≈ p
lim B(t) = − M lim B(t) = M
2 t→−∞ q t→+∞
(t2 − t1 ) · M (p+q) ⇔ q > p; t1 and t2 are near tI
≈ 4q
(t2 − t1 ) · M p ⇔ q ≤ p, t1 and t2 are near ts (36) The ratio between negative asymptote and the dis-
tance of these asymptotes lays in range (0,1] which is
The growth rate GR for time interval Δt is always convenient to choose as the shape parameter s, and
positive: which can be measured in percent. In fact, according
to the value of s, the S-curve is stretched in the verti-
B(t) − B(t − Δt)
GRΔt = cal direction (on the y-axis) preserving the total mar-
B(t − Δt)
ket capacity M. [9]
Due to the fact that the Bass model starts from ts,
B(ts) = 0, the growth rate for t –> ts goes to infinity. The distance between these asymptotes is M . (1 + p /
q), so the shape parameter s is:
Above described characteristics of the Bass model
pM/q p
with its explanatory attributes can be used as helpful s= = , p > 0, q ≥ 0 (37)
M + pM/q q+p
input for estimation or assessment of model parame-
ters for forecasting purposes. Characteristic values of s are:
4.2 The Bass Model with Explanatory s –> 0 negative asymptote –> 0, imitation prevails,
Parameters curve is similar to a simple logistic growth
Parameters M and ts are descriptive and can easily be model, (q >> p > 0),
linked with market conditions. Although p as a coef-
ficient of innovation and q as a coefficient of imita- s = 0.5 sales peak occurs at time when product starts
tion have explanatory features, they are mutually (q = p > 0)
Top-left 10 % 90 % 20 years
From (37) follows:
Top-right 10 % 90 % 10 years
v
1 − 1 + s(1−v)
Δt
1+ · s s(1−v)
=M − t−ts (40) where parameters of logistic growth model a and b
Δt
v
1 + (1/s − 1) · 1 + s(1−v) are:
Expression (40) is the reparameterized Bass model 1 v ln s
a= ln , b = ts −
with explanatory parameters (instead of p and q) Δt s(1 − v) a
where: M – market capacity; ts – time when product • For s = 0.5 the Bass model gets a form:
is introduced, B(ts) = 0, ts ≤ t, Δ t – characteristic
− t−ts
duration of product, Δ t > 0, s – shape parameter, 1− 1+v Δt
1−v
0 < s ≤ 1; and v – penetration at time point ts + Δ t, B(t; M, ts , Δt, s = 0.5, v) = M − t−ts
=
Δt
1+v
0 ≤ v < 1. The model from (40), B(t; M, ts , Δ t, s, v), 1+ 1−v
needs four parameters; M, ts , Δ t and s to be deter- 2M
= − t−ts
− M = L(t; 2M, a, b) − M
mined. The value of the auxiliary parameter v does 1+v Δt
1+ 1−v
not need to be determined, it just allows the forecast-
ing practitioner to choose which level of penetration This curve has a shape of logistic model with dou-
he/she wants to deal with (ie. 90 %, 95 %, etc). ble market capacity M but vertically shifted down
by M. Parameters a and b of this ‘halved’ logistic
Special cases of (40): model are:
• For v = 0, the value of model B(t) is zero: 1 1+v
a= ln , b = ts
Δt 1−v
B(t; M, ts , Δ t, s, v = 0) = 0 • For s = 1 the Bass model degrades into an exponen-
tial saturation growth model:
• For s –> 0, the Bass model degrades into a simple
t−ts
logistic model: B(t; M, ts , Δt, s = 1, v) = M 1 − (1 − v) Δt
s = 20 % s = 50 % s = 80 %
Table 3 Reparameterized Bass model framework for forecasting new products adoption prior to launch.
s – shape parameter, Δ t – characteristic duration (time to reach penetration level vM measured from ts ),
value for v is chosen for 95 % penetration (v = 95 %)
Table 4 Framework for parameter estimation in cases of a larger set of known data than minimal
Mladen Sokele is Assistant Director of the Operative Planning and Project Management Department in Croatian Telecom Inc.,
Zagreb, Croatia. He graduated in 1983 and received his Master’s degree in 1987 at the Faculty of Electrical Engineering and
Computing of the University of Zagreb. After graduation he joined Ericsson Nikola Tesla Company and from 1989 he was
lecturer at the Faculty of Electrical Engineering in Osijek. He has been with Croatian Telecom Inc. since 1992. Mr. Sokele is
author or co-author of more than fifty published scientific and professional papers in the field of telecommunications and
related indicators on modelling and forecasting. He is also author of the LOgistic Spline Trend (LOST) method and program
tool for the prediction of service penetration over time.