The Duckworth-Lewis Method and Twenty20 Cricket
The Duckworth-Lewis Method and Twenty20 Cricket
CRICKET
by
Rianka Bhattacharya
c Rianka Bhattacharya 2010
SIMON FRASER UNIVERSITY
Spring 2010
Date Approved:
ii
Abstract
Cricket has been a very popular sport around the world. But since most versions of cricket
take longer to play than other popular sports, cricket is likely to be affected by unfavourable
weather conditions. In 1998, Duckworth and Lewis developed a method for resetting the
target scores for the team batting second in interrupted one-day cricket.
Twenty20 is the latest form of cricket. Currently, the Duckworth-Lewis method is also
used for resetting targets in interrupted Twenty20 matches. However, this may be less than
ideal since the scoring pattern in Twenty20 is much more aggressive than that in one-day
cricket.
In this project, we consider the use of the Duckworth-Lewis method as an approach to
reset target scores in interrupted Twenty20 matches. The construction of the Duckworth-
Lewis table is reviewed and alternate resource tables are presented for Twenty20. Alternative
resource tables are constructed in a nonparametric fashion using two different approaches.
iii
To my dear parents, Mr. Arun Prokas Bhattacharya and Mrs. Bani Bhattacharya.
iv
“T20 is a fun game. It is like practising fours and sixes.”
— Kamran Akmal, 2009 Twenty20 World Cup (after being judged the “Man Of The
Match” against Netherlands), 2009
v
Acknowledgments
I would like to begin by thanking the faculty in the Department of Statistics and Actuarial
Science for their great commitment to their students. Their kind attention to all my queries
encouraged me a lot in continuing my studies.
I am truly indebted to my senior supervisor, Dr. Tim Swartz. Without his expert
guidance it would not be possible to for me to continue with my studies in Statistics. His
support on academic, professional and personal issues have been undoubtably priceless. He
was always there to help me to get on to the right direction and to correct my chapters.
He taught me how to write scientific papers and how to be persistent in accomplishing the
project. It is impossible to thank him enough for his support throughout my studies at SFU.
It has really been a great honour for me to pursue my Masters degree under his supervision.
I would like to extend my heartiest thanks to my co-supervisor, Dr. Paramjit Gill who
made some valuable suggestions on my project. He helped me a lot, especially in the
preliminary stages of the project with the data collection as that has been a real challenge. I
would also like to thank my examining committee member, Dr. Gary Parker, for commenting
on the project. My special thanks to Ian Bercovitz for his guidance in helping me to learn
how to write a paper.
I would also like to thank Dr. Brad McNeney for giving me an oppurtunity to pursue
my studies at SFU. The prompt actions of both Dr. Brad McNeney and Dr. Tim Swartz on
my visa issues were really appreciated.
I am grateful for the financial support provided by the Department of Statistics and Ac-
turial Science. My sincere gratitude to Sadika, Kelly and Charlene for their kind assistance.
Thanks to all the graduate students in the Department for making my studies at SFU
enjoyable. Special thanks to Saman Muthukumarana for helping me to settle down on my
arrival and also for his valuable suggestions throughout my studies, especially while working
vi
on this project. My thanks to Caroll Co, Aruni Tennakoon, Harsha Perera, Joslin Goh,
Jervyn Ang, Ryan Lekivetz, Mathew Pratola, Jia Xu and Monica Lu for your friendship.
Thank you Suddeshna for helping me learn Latex and providing such a friendly atmo-
sphere at home. Thank you to my family for encouraging and standing by my side at all
times. Last, but not the least, I would like to thank you Niladri for your understanding and
endless support in so many ways.
vii
Contents
Approval ii
Abstract iii
Dedication iv
Quotation v
Acknowledgments vi
Contents viii
List of Tables x
xiii
1 Introduction 1
1.1 Different Forms of Cricket and a Brief Background . . . . . . . . . . . . . . . 1
1.2 The Duckworth-Lewis Method and One-day Cricket . . . . . . . . . . . . . . 2
1.3 Motivation of the Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Organization of the Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
viii
3 New Resource Tables for Twenty20 13
3.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Development of the Resource Table . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.1 Construction of R matrix and its Flaws . . . . . . . . . . . . . . . . . 14
3.2.2 Isotonic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 Another Approach in Constructing the R Matrix . . . . . . . . . . . . . . . . 17
3.4 Gibbs Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.5 Comparison of the Suggested Resource Tables with the Duckworth-Lewis
Resource Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4 Conclusion 30
Bibliography 31
ix
List of Tables
x
3.5 A nonparametric resource table for Twenty20 based on Gibbs sampling using
the R matrix in Table 3.1. The table entries indicate the percentage of
resources remaining in a match with the specified number of wickets lost and
overs available. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6 A nonparametric resource table for Twenty20 based on Gibbs sampling using
the R matrix in Table 3.3. The table entries indicate the percentage of
resources remaining in a match with the specified number of wickets lost and
overs available. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
xi
List of Figures
3.1 Heat map of the absolute differences between the Duckworth-Lewis resource
table (Table 2.1) and the nonparametric resource table based on Gibbs sam-
pling (Table 3.5). Darker shades indicate larger differences. . . . . . . . . . . 25
3.2 Heat map of the absolute differences between the Duckworth-Lewis resource
table (Table 2.1) and the nonparametric resource table based on Gibbs sam-
pling (Table 3.6). Darker shades indicate larger differences. . . . . . . . . . . 27
3.3 Heat map of the absolute differences between the nonparametric resource
table based on Gibbs sampling (Table 3.5) and the nonparametric resource
table based on Gibbs sampling (Table 3.6). Darker shades indicate larger
differences. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
xii
xiii
Chapter 1
Introduction
1
CHAPTER 1. INTRODUCTION 2
Melbourne in 1971. As the name suggests, a one-day cricket match is completed in one day.
In this form, each team bats for a maximum of 50 overs. For more information on cricket
see, http://en.wikipedia.org/wiki/Cricket.
The popularity of one-day cricket was at its peak when Twenty20 was introduced in 2001
in the Twenty20 Cup between the English counties. The Twenty20 game, also commonly
known as T20 cricket, involves two teams, each having a single innings, batting for a max-
imum of 20 overs. This version of the game is completed in about three and a half hours,
with each innings lasting around 75 minutes, thus bringing the game closer to the times-
pan of other popular team sports. With the introduction of the biennial World Twenty20
tournament in 2007 and the Indian Premier League in 2008, Twenty20 cricket has gained
widespread popularity.
general, people do not understand the method and how targets are set, but they do agree
that the targets are sensible and are preferable to the target scores set by the approach
based on run rates.
all the international Twenty20 matches involving ICC teams that have taken place from
February 17/2005 through November 9/2009 is used in the construction of the table. The
methodology used to obtain table entries is based on isotonic regression and Markov chain
simulations. The resultant table is compared with the Duckworth-Lewis table. We conclude
with a discussion in chapter 4.
This project is an expansion of the companion paper by Bhattacharya, Gill and Swartz
(2010).
Chapter 2
5
CHAPTER 2. TWENTY20 AND DUCKWORTH-LEWIS 6
Now, let us describe how the batsmen bat. The two batsmen stand at opposite ends of the
pitch. For each ball that is bowled, only one of these two batsmen is the “striker”. There
are eleven players on each side. Among the players of the fielding side, one player is the
bowler and another is the wicketkeeper. The bowler bowls from the opposite end of the
pitch to the striker. When a wicket is hit by a ball, and one or both the bails fall off, a
batsman gets “stumped out”. This is one of the most common methods to get a batsman
out. If a ball is hit by the striker, the two batsmen may attempt to change places while the
fielders chase the ball and try to throw it back to one of the wickets. A run is scored every
time the batsmen change places. However, the batsmen should be careful while changing
places as they might get “run out”, if the fielder succeeds in hitting the wicket with the
ball while they are still on the pitch. This is another way of a batsman getting out. Runs
are also scored when the batsman hits the ball and no fielder succeeds in stopping it before
the ball touches the boundary line. Four runs are scored if the ball touches the ground at
least once before touching the boundary and six runs are scored if it goes over the boundary
without touching the ground in between. The wicketkeeper’s job is to stop the ball when
the striker fails to hit the ball.
CHAPTER 2. TWENTY20 AND DUCKWORTH-LEWIS 7
There are other ways in which a batsman may be put out by the fielding team. If the
ball is hit and a fielder catches it before it touches the ground, then a batsman is out. It is
called a “catch out”. Another form of getting out occurs when the ball hits the striker’s leg
and the umpire rules that the ball would have hit the wicket if the leg had not been there.
This is called “lbw” (leg before wicket). When a batsman gets out, we say that a wicket
has been taken or lost. This plays an important role towards ending an innings.
Basically, there are two ways for an innings to end: when all 10 wickets are lost or when
20 overs have been played. An over consists of six balls bowled. There also exists a third
way to end an innings. This happens when the team batting second surpasses the total runs
of the team batting first.
In Twenty20, an innings does not go beyond the 20 overs or, in other words, 120 balls.
However, sometimes more than 120 balls may be bowled in an innings because of “no balls”
and “wides”. The former occurs if the bowler performs any illegal action while bowling and
the latter is called by the umpire if he feels that the ball is unreachable to the striker. In
both cases, the batting side is awarded one run and the ball is also not counted as a part of
the over, thus explaining why more than 120 balls are sometimes bowled.
All of the above rules are common to limited overs cricket. The “laws of cricket” which
differentiate Twenty20 from one-day cricket are given below.
• If the bowler delivers a no ball by overstepping in the pitch, a run is awarded to the
batting team and the next delivery is designated as a “free hit”. A batsman can be
made out from a free hit only by run out and some other uncommon methods.
• No more than five fielders can be on the leg-side at any time during a match. From
the point of view of a right-handed batsman facing the bowler, it is the left hand side
of the cricket field. With a left-handed batsman the leg-side is the right-handed side
of the batsman.
• During the first six overs only two fielders can be outside the 30 yards circle which
surrounds the pitch. This period of the match is known as the “powerplay”. After
the first six overs a maximum of five fielders can be outside the circle.
CHAPTER 2. TWENTY20 AND DUCKWORTH-LEWIS 8
• Umpires may award five penalty-runs at their discretion if they feel that either team
is wasting time.
• If a fielding team does not start to bowl the 20th over within 75 minutes, then the
batting team is awarded an extra six runs for every over bowled after the 75th minute.
Thus we can say that Twenty20 is a faster and even more exciting format than one-day
cricket. The batting team tends to bat more aggressively as only 20 overs are available to
score runs.
Table 2.1: Abbreviated version of the Duckworth-Lewis resource table (Standard Edition).
The table entries indicate the percentage of resources remaining in a match with the specified
number of wickets lost and overs available.
Wickets Lost
Overs Available 0 1 2 3 4 5 6 7 8 9
50 100.0 93.4 85.1 74.9 62.7 49.0 34.9 22.0 11.9 4.7
40 89.3 84.2 77.8 69.6 59.5 47.6 34.6 22.0 11.9 4.7
30 75.1 71.8 67.3 61.6 54.1 44.7 33.6 21.8 11.9 4.7
25 66.5 63.9 60.5 56.0 50.0 42.2 32.6 21.6 11.9 4.7
20 56.6 54.8 52.4 49.1 44.6 38.6 30.8 21.2 11.9 4.7
10 32.1 31.6 30.8 29.8 28.3 26.1 22.8 17.9 11.4 4.7
5 17.2 17.0 16.8 16.5 16.1 15.4 14.3 12.5 9.4 4.6
1 3.6 3.6 3.6 3.6 3.6 3.5 3.5 3.4 3.2 2.5
0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
In constructing the table, Duckworth and Lewis (1998) provide only partial information
citing reasons of commercial confidentiality. In the paper it is stated that 20 parameters
Z0 (w) and b(w), w = 0,1,...,9 are estimated corresponding to the function
where Z(u, w) is the average total score obtained in u overs in an unlimited overs match
where w wickets have already been taken. The parameter Z0 (w) is the asymptotic average
total score from the last (10 - w) wickets in unlimited overs and b(w) is the exponential
decay constant, both of which depend on the number of wickets lost. However, the estimation
procedure is not disclosed. Although we admit the utility of the Duckworth-Lewis resource
table in one-day cricket, some questions naturally arise based on equation (2.1) and the
estimated values in Table 2.1:
• There are many parametric curves available that could be fit to obtain a resource
table. Is (2.1) the best curve to fit? Is there any advantage of using a nonparametric
approach?
• The function (2.1) uses an asymptote (i.e. unlimited overs cricket). Is there any
advantage in taking the limited overs nature of the one-day game into account?
CHAPTER 2. TWENTY20 AND DUCKWORTH-LEWIS 10
• In Table 2.1, the last two columns have some identical entries going down the columns.
Though very few matches occur under these conditions, the question remains whether
it is sensible for resources to remain constant as the number of available overs are
decreasing. This may be a consequence of the asymptote imposed by (2.1).
For ease of discussion, we convert the Duckworth-Lewis resource table to the context of
Twenty20 cricket. That is, we scale the entries in the table so that the resources available to
a team with 20 overs and 10 wickets in hand corresponds to 100% resources. In Table 2.2,
the full Duckworth-Lewis resource table (Standard Edition) for Twenty20 is given. In this
table the entries are obtained by dividing the corresponding entries in Table 2.1 by 0.566
(which is the resources remaining in a one-day match where 20 overs are still available and
10 wickets are available).
CHAPTER 2. TWENTY20 AND DUCKWORTH-LEWIS 11
Table 2.2: The Duckworth-Lewis resource table (Standard Edition) scaled for Twenty20.
The table entries indicate the percentage of resources remaining in a match with the specified
number of wickets lost and overs available.
Wickets Lost
Overs Available 0 1 2 3 4 5 6 7 8 9
20 100.0 96.8 92.6 86.7 78.8 68.2 54.4 37.5 21.3 8.3
19 96.1 93.3 89.2 83.9 76.7 66.6 53.5 37.3 21.0 8.3
18 92.2 89.6 85.9 81.1 74.2 65.0 52.7 36.9 21.0 8.3
17 88.2 85.7 82.5 77.9 71.7 63.3 51.6 36.6 21.0 8.3
16 84.1 81.8 79.0 74.7 69.1 61.3 50.4 36.2 20.8 8.3
15 79.9 77.9 75.3 71.6 66.4 59.2 49.1 35.7 20.8 8.3
14 75.4 73.7 71.4 68.0 63.4 56.9 47.7 35.2 20.8 8.3
13 71.0 69.4 67.3 64.5 60.4 54.4 46.1 34.5 20.7 8.3
12 66.4 65.0 63.3 60.6 57.1 51.9 44.3 33.6 20.5 8.3
11 61.7 60.4 59.0 56.7 53.7 49.1 42.4 32.7 20.3 8.3
10 56.7 55.8 54.4 52.7 50.0 46.1 40.3 31.6 20.1 8.3
9 51.8 51.1 49.8 48.4 46.1 42.8 37.8 30.2 19.8 8.3
8 46.6 45.9 45.1 43.8 42.0 39.4 35.2 28.6 19.3 8.3
7 41.3 40.8 40.1 39.2 37.8 35.5 32.2 26.9 18.6 8.3
6 35.9 35.5 35.0 34.3 33.2 31.4 29.0 24.6 17.8 8.1
5 30.4 30.0 29.7 29.2 28.4 27.2 25.3 22.1 16.6 8.1
4 24.6 24.4 24.2 23.9 23.3 22.4 21.2 18.9 14.8 8.0
3 18.7 18.6 18.4 18.2 18.0 17.5 16.8 15.4 12.7 7.4
2 12.7 12.5 12.5 12.4 12.4 12.0 11.7 11.0 9.7 6.5
1 6.4 6.4 6.4 6.4 6.4 6.2 6.2 6.0 5.7 4.4
CHAPTER 2. TWENTY20 AND DUCKWORTH-LEWIS 12
Figure 2.2: Average number of runs Z(u, w) from u overs remaining with w wickets lost:
Duckworth and Lewis (1998).
Chapter 3
13
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 14
After running the Java script we get the match results on a ball-by-ball basis. For each
match, define x(u, w(u)) as the runs scored from the stage in the first innings where u
overs are available and w(u) wickets have been taken until the end of the first innings. We
calculate x(u, w(u)) for all values of u and w(u) that occur in the first innings of a Twenty20
match. The variable u takes the values 0,1,...,20 and w(u) takes the values 0,1,...,10 with
w(20)=0.
We now define ruw as the estimated percentage of resources remaining when u overs
are available and w wickets have been taken. Thus the matrix R = (ruw ) becomes our
first attempt as a new resource table for Twenty20. We calculate (100%)ruw by averaging
x(u, w(u)) over all the matches where w(u) = w wickets have been taken and dividing
by the average of x(20, 0) over all the matches considered. Note that the denominator is
simply the average number of first innings runs over all matches. In the case of u=0, we set
ruw =r0w =0.0%. In this matrix, r20,0 =100% is also desired. The calculated matrix is given
in Table 3.1.
The new estimated resource matrix R is calculated in a non-parametric fashion and it
does not make any assumptions concerning the scoring patterns in Twenty20. In spite of
these facts, the estimated matrix is less than ideal. In Table 3.1 there are many missing
entries as the data for those situations are missing. There are other flaws associated with
this table. For example, the table entries do not exhibit the monotonicity that we expect.
Logically, we require a resource table that is decreasing as we go from left to right along
the rows and also as we go down the columns. This property of the resource table can be
explained by a simple example. Consider the situation in a match where 16 overs are left to
bat with 10 wickets in hand and the other situation is that 16 overs are left in a match with
only 2 wickets in hand. In the former case the batsmen bat aggressively since a lost wicket is
not a big problem. In the latter case, the batsmen bat less aggressively as they try to retain
the two wickets in hand. Thus, in the former case the team has more resources available.
This explains why the resources are decreasing from left to right along rows. Similarly, we
consider two different situations in a match to explain the decreasing property of the table
down the columns. Suppose the match situation is such that 16 overs are left to bat with
10 wickets in hand and the other situation is 5 overs left with 10 wickets in hand. Clearly
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 15
in the former case, the extra overs provide extra opportunity to score runs.
There are also some dubious entries in Table 3.1. One such entry is 110.2% resources
corresponding to 19 overs available and 2 wickets taken. This entry is clearly misleading as
all the entries in the table should be less than 100%. This particular entry arises due to the
fact that the sample size corresponding to the given situation is very small, only 2 matches.
We emphasize that the investigation in this project is one of discovery rather than an
attempt to replace the existing Duckworth-Lewis table for Twenty20.
Table 3.1: The matrix R = (ruw ) of estimated resources for Twenty20 (calculated by taking
ratio of the average of x(u, w(u)) over all matches where u overs are left to bat and w(u)=w
wickets have been taken to the average number of first innings runs over all the matches).
Missing entries correspond to match situations where data are unavailable.
Wickets Lost
Overs Available 0 1 2 3 4 5 6 7 8 9
20 100.0
19 93.6 83.0 110.2
18 90.4 85.8 78.3
17 86.7 80.5 82.8 53.7
16 81.7 74.5 81.9 70.7 32.8
15 76.5 71.4 71.5 65.9 59.9
14 68.3 69.1 67.6 66.2 58.4
13 63.8 68.2 62.4 62.9 59.0 24.3
12 62.1 62.3 60.6 57.3 58.8 44.1
11 60.5 56.3 57.0 53.6 61.0 39.7
10 57.6 49.6 52.1 52.8 48.1 38.6 41.0 35.2
9 54.9 52.1 43.6 49.0 44.1 33.8 35.0 29.7
8 51.0 46.4 41.7 42.2 41.2 36.7 27.5 28.7
7 48.6 45.8 38.9 35.9 39.1 34.8 24.1 25.5
6 54.0 37.9 36.6 30.3 36.2 31.3 20.9 21.4 26.7
5 44.0 32.5 25.4 28.7 29.4 23.9 17.1 14.9
4 28.2 23.4 22.5 22.2 20.9 14.3 10.6 6.7
3 20.6 19.9 16.9 17.8 15.8 12.4 7.6 1.2
2 21.2 17.6 11.9 13.4 10.6 11.0 7.2 1.4
1 8.7 5.2 7.3 6.0 5.5 6.0 2.6
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 16
with respect to the matrix Y = (yuw ) where the double summation corresponds to u =
1,...,20 and w = 0,...,9. The quw are the weights and the minimization is subject to the
constraints yu,w ≥ yu,w+1 and yu,w ≥ yu−1,w . We impose, y20,0 = 100 and y0,w = 0 for w =
0,...,9 and yu,10 = 0 for u = 1,...,20.
The fitting of Y is completely nonparametric. But still there are some arbitrary choices
made in the minimization of (3.1). The “squared error” discrepancy in (3.1) is one of many
functions that could be minimized. Also, there are various weights that might be chosen.
We note that minimization of the function F in (3.1) with the squared error discrepancy
is equivalent to the method of constrained maximum likelihood estimation where the data
ruw are independently distributed normal variates with means yuw and variances 1/quw .
Again, in this project, we consider the matrix Y : 20 × 10 based on overs. Instead,
we might have considered a larger matrix Y : 120 × 10 based on balls. But we prefer the
former case as it involves less missing data and leads to less computation. Not only that,
if we have a matrix Y based on overs, it is always possible to find the Y matrix based on
balls by simple interpolation.
Since we are considering a weighted least squares approach, the choice of weights plays
an important role and may affect the result. We choose a very simple set of weights. We
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 17
take 1/quw equal to the sample variance with respect to the calculation of ruw . The reason
behind this choice is that when ruw is less variable, yuw should be close to ruw .
In Table 3.2, a nonparametric resource table based on the minimization of (3.1) is given.
The algorithm for isotonic regression in two independent variables was first introduced by
Dykstra and Robertson (1982). Fortran code was subsequently developed by Bril, Dykstra,
Pillers and Robertson (1984). The algorithm for bivariate isotonic regression has recently
been updated in R in the “Iso” package on the Cran website (www.cran.r-project.org). In
this project, we have used the R code to implement bivariate isotonic regression.
The program requires 27 iterations to achieve convergence. But, this table also has some
drawbacks. First, there exist adjacent table entries that are of the same value which is not
desirable. Clearly, resources cannot be constant with decreasing overs left or with increasing
wickets lost. Secondly, this table also has a large number of missing values corresponding to
those match situations where there is no data available. This is natural as it is impossible
to have matches, for example where 20 overs are left but at the same time 9 or even 10
wickets have been taken.
Table 3.2: A nonparametric resource table based on Table 3.1 for Twenty20 using isotonic
regression. The table entries indicate the percentage of resources remaining in a match
with the specified number of wickets lost and overs available. Missing entries correspond to
match situations where data are unavailable.
Wickets Lost
Overs Available 0 1 2 3 4 5 6 7 8 9
20 100.0
19 93.6 85.5 85.5
18 90.4 85.5 80.8
17 86.7 80.8 80.8 64.7
16 81.7 77.4 77.4 64.7 55.9
15 76.5 71.5 71.5 64.7 55.9
14 68.8 68.8 67.6 64.7 55.9
13 66.6 66.6 62.6 62.6 55.9 38.4
12 62.2 62.2 60.6 57.3 55.9 38.4
11 60.5 56.8 56.8 54.8 54.8 38.4
10 57.6 52.1 52.1 52.1 48.1 38.4 34.1 29.3
9 54.9 52.1 46.5 46.5 44.1 36.3 34.1 29.3
8 51.0 46.4 42.0 42.0 41.2 36.3 28.6 28.6
7 48.6 45.8 38.9 37.3 37.3 34.8 25.3 25.3
6 39.7 39.7 36.6 32.8 32.8 31.3 23.0 21.4 21.4
5 39.7 32.5 28.0 28.0 28.0 23.0 17.1 15.5
4 27.9 23.4 22.5 22.2 20.9 14.3 10.7 10.7
3 20.7 19.9 17.4 17.4 15.8 12.4 7.7 7.7
2 20.7 17.6 12.5 12.5 10.8 10.8 7.2 1.8
1 8.7 6.6 6.6 6.0 5.7 5.7 1.8
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 19
Table 3.3: The matrix R = (ruw ) of estimated resources for Twenty20 (calculated by taking
the average of the ratio of x(u, w(u)) to x(120, 0) over all matches where u overs are left to
bat and w(u)=w wickets have been taken). Missing entries correspond to match situations
where data are unavailable.
Wickets Lost
Overs Available 0 1 2 3 4 5 6 7 8 9
20 100.0
19 95.8 95.9 97.6
18 91.8 91.1 90.8
17 86.0 86.1 86.6 81.0
16 80.2 80.1 83.0 80.7 73
15 73.6 74.3 76.1 78.4 67.8
14 67.1 69.3 71.6 73.4 69.2
13 61.6 66.8 65.6 69.7 67.1 54.1
12 57.7 61.5 61.0 63.3 66.0 57.6
11 54.3 55.5 57.0 58.9 63.5 53.1
10 50.2 50.1 51.2 56.1 53.4 48.6 48.1 57.4
9 44.6 47.9 45.6 49.9 48.7 42.7 41.3 44.8
8 40.7 42.2 41.6 42.8 43.7 43.0 35.2 42.5
7 37.6 41.0 37.0 36.3 41.2 38.1 31.6 37.7
6 40.8 32.3 34.4 30.5 35.9 34.6 26.2 29.2 38.3
5 34.5 29.0 26.2 28.1 30.5 28.2 23.7 21.7
4 23.1 23.9 21.9 22.8 24.0 17.0 15.4 14.9
3 16.7 18.5 17.0 17.2 17.7 14.4 11.4 2.7
2 16.7 14.3 11.6 12.9 11.5 12.4 9.3 2.3
1 7.0 5.3 7.0 6.3 5.9 6.7 3.4
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 20
This table also suffers from some of the same problems as Table 3.1. For example, the miss-
ing values in the table are a drawback. As before, we ran the isotonic regression program on
this estimated resource table. The resultant nonparametric resource table is given in Table
3.4. In this case, the program requires 32 iterations to achieve convergence.
Table 3.4: A nonparametric resource table based on Table 3.3 for Twenty20 using isotonic
regression. The table entries indicate the percentage of resources remaining in a match
with the specified number of wickets lost and overs available. Missing entries correspond to
match situations where data are unavailable.
Wickets Lost
Overs Available 0 1 2 3 4 5 6 7 8 9
20 100.0
19 96.2 96.2 96.2
18 91.8 91.2 91.2
17 86.2 86.2 86.2 86.2
16 81.3 81.3 81.3 81.3 81.3
15 75.9 75.9 75.9 75.9 75.9
14 71.3 71.3 71.3 71.3 71.3
13 67.2 67.2 67.2 67.2 67.2 67.2
12 62.5 62.5 62.5 62.5 62.5 62.5
11 58.3 58.3 58.3 58.3 58.3 58.3
10 53.3 53.3 53.3 53.3 53.3 48.7 48.7 48.7
9 53.3 48.1 48.1 48.1 48.1 45.3 45.3 45.3
8 53.3 43.1 43.1 43.1 43.1 43.1 43.1 43.1
7 53.3 41.0 38.6 38.6 38.6 38.6 38.6 38.6
6 53.3 39.0 34.4 33.8 33.8 33.8 33.8 33.8 33.8
5 39.0 29.0 28.5 28.5 28.5 28.5 28.5 28.5
4 27.1 23.9 22.6 22.6 22.6 19.6 19.6 19.6
3 27.1 18.5 17.2 17.2 17.2 14.9 14.9 14.9
2 27.1 14.3 12.2 12.2 11.9 11.9 9.3 3.8
1 7.0 6.6 6.6 6.3 6.1 6.1 3.8
Table 3.4 also cannot be regarded as ideal. In fact, we hoped that the new approach would
give us a more reliable resource table. But in this table we also see equal adjacent values.
This table also suffers from missing entries. Thus though we hoped to see a better resource
table, it can hardly be regarded as an improvement to the previous one at this stage.
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 21
• Metropolis-Hastings
• Gibbs sampling
Gibbs sampling is a MCMC algorithm that is applicable when the associated conditional
distributions are tractable.
We saw that the nonparametric resource tables generated by isotonic regression have
many shortcomings. To address those criticisms, we now use Gibbs sampling as another
approach to estimation. Note that the minimization of (3.1) arises from the maximization
of the normal likelihood
1 XX
exp{− quw (ruw − yuw )2 } (3.2)
2
Thus, we now consider a Bayesian model where the unknown parameters in (3.2) are the y’s.
A flat default prior is assigned to the y’s subject to the required monotonicity constraints
as mentioned in (3.1). Thus the posterior density takes the form (3.2) and Gibbs sampling
can be carried out via sampling from the full conditional distributions
1
[yuw | ·] ∼ Normal(ruw , ) (3.3)
quw
subject to the local constraints on yuw in the given iteration of the algorithm. Sampling
from (3.3) is easily carried out using a normal generator and rejection sampling according to
the constraints. Here it should be mentioned that though we are taking a parametric form
for the random variables in (3.3), we still refer to the approach as nonparametric since we
are not assuming any functional relationship that is imposed on y. In Table 3.5 and Table
3.6 we present the resource matrices that we get after carrying out Gibbs sampling using
the R matrices from Table 3.1 and Table 3.3 respectively. These are the alternate resource
tables for Twenty20 cricket.
In Table 3.5, the table entries are the estimated posterior means of the y’s obtained
through Gibbs sampling. In this project we have used Fortran code where the estimates
stabilize after 50,000 iterations. Note that the R matrix has many missing entries. So to
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 22
address this problem, we actually impute the missing ruw ’s with the corresponding entries
in the scaled Duckworth-Lewis table given in Table 2.1. This imputation is in the spirit of a
Bayesian approach as we are using prior information. A nice feature about this table is that
it is a complete resource table and devoid of any missing values. No adjacent table entries
are identical. Thus one major drawback of the previous resource tables are been overcome.
We should also mention that the approach is able to take into account expert opinion. For
example, if there is an opinion that the table entry yij should be a particular value k, then
to set the table entry to that value we simply take rij = k and assign a sufficiently small
standard deviation.
Table 3.5: A nonparametric resource table for Twenty20 based on Gibbs sampling using the
R matrix in Table 3.1. The table entries indicate the percentage of resources remaining in
a match with the specified number of wickets lost and overs available.
Wickets Lost
Overs Available 0 1 2 3 4 5 6 7 8 9
20 100.0 96.9 93.0 87.9 81.3 72.2 59.9 44.8 29.7 17.6
19 95.6 90.9 87.7 83.0 76.9 68.3 56.5 42.0 27.2 15.3
18 91.7 86.7 82.9 78.7 73.2 65.4 54.2 40.2 25.7 13.9
17 87.7 82.3 78.9 73.8 69.7 62.8 52.2 38.7 24.6 12.8
16 83.5 78.2 75.3 70.5 66.4 60.2 50.3 37.4 23.5 12.0
15 79.2 74.3 70.9 66.9 62.6 57.4 48.4 36.2 22.7 11.2
14 75.1 70.7 67.3 63.7 59.3 54.6 46.4 35.0 21.8 10.5
13 71.5 67.4 63.6 60.3 56.2 51.5 44.3 33.8 21.0 9.8
12 68.3 63.7 60.2 56.8 52.9 47.5 41.9 32.6 20.2 9.1
11 65.0 59.9 56.6 53.3 49.7 43.9 39.3 31.3 19.4 8.5
10 61.3 56.0 52.6 50.1 46.0 40.8 36.1 30.0 18.6 7.9
9 57.9 52.3 47.9 46.1 42.5 37.8 33.1 28.3 17.7 7.2
8 54.0 48.3 44.3 41.7 38.9 34.9 30.2 26.1 16.7 6.6
7 49.3 44.2 40.2 37.4 35.4 32.1 27.2 23.4 15.7 5.9
6 41.7 38.5 35.7 33.0 31.7 29.0 24.2 20.0 14.5 5.2
5 36.2 33.4 31.0 28.6 27.3 25.5 21.5 17.0 12.2 4.4
4 30.8 28.0 26.1 24.1 22.4 20.7 18.3 14.2 10.0 3.5
3 25.4 22.8 21.1 19.4 17.7 16.5 14.4 11.6 7.9 2.5
2 19.7 17.2 15.5 14.1 12.7 11.9 10.6 9.3 6.2 1.6
1 13.7 11.3 9.7 8.5 7.3 6.7 6.0 5.2 4.2 0.9
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 23
Table 3.6 is the resource table that we obtained from Gibbs sampling using Table 3.3 as the
R matrix. Again, we impute the missing values with the corresponding Duckworth-Lewis
table entries given in Table 2.2. The resultant resource table has some advantages over
the earlier resource tables. Note that the table entries are slightly different from those in
Table 3.5. In the next section we compare the two suggested resource tables with the scaled
Duckworth-Lewis resource table given in Table 2.2.
Table 3.6: A nonparametric resource table for Twenty20 based on Gibbs sampling using the
R matrix in Table 3.3. The table entries indicate the percentage of resources remaining in
a match with the specified number of wickets lost and overs available.
Wickets Lost
Overs Available 0 1 2 3 4 5 6 7 8 9
20 100.0 98.0 95.1 89.8 82.5 72.7 60.1 44.8 29.8 17.8
19 96.0 95.5 92.4 85.9 78.6 69.0 56.8 42.1 27.3 15.5
18 91.9 91.0 89.5 82.9 75.7 66.3 54.5 40.2 25.8 14.1
17 86.7 86.1 85.3 80.5 73.2 63.9 52.6 38.8 24.7 13.1
16 81.6 81.0 80.5 78.2 71.1 61.7 50.7 37.5 23.7 12.2
15 76.2 75.5 75.1 74.3 68.7 59.5 49.0 36.3 22.8 11.5
14 71.9 71.3 70.9 70.0 66.6 57.5 47.2 35.1 22.0 10.8
13 67.8 67.0 66.3 65.8 63.8 55.5 45.4 34.0 21.2 10.1
12 63.8 62.6 61.9 61.3 60.1 53.7 43.4 32.8 20.4 9.5
11 59.7 58.3 57.7 57.1 56.2 51.1 41.4 31.5 19.6 8.9
10 54.8 53.8 53.3 52.9 51.5 47.1 39.1 30.2 18.8 8.3
9 51.3 49.5 48.3 47.8 46.8 42.8 36.9 28.8 18.0 7.7
8 47.5 44.9 43.3 42.5 41.6 39.5 34.1 27.4 17.1 7.1
7 44.5 41.4 39.3 38.3 37.7 36.2 31.1 26.0 16.2 6.5
6 40.9 36.6 34.9 33.6 33.1 32.2 28.0 24.4 15.2 5.8
5 36.1 33.5 30.8 29.0 28.4 27.9 25.4 21.9 14.1 5.2
4 30.5 27.5 25.4 24.0 22.8 22.0 20.9 17.2 12.8 4.4
3 24.9 22.1 20.2 18.8 17.7 16.9 15.9 13.9 10.8 3.4
2 19.7 17.3 15.7 14.1 12.6 12.2 11.5 10.9 8.5 2.4
1 13.4 10.9 9.2 7.8 7.0 6.7 6.3 5.8 5.2 1.6
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 24
table (Table 3.5) suggests that (0.439)240 → 106 runs is expected. It is interesting to note
that in spite of the aggressive batting style of Twenty20 cricket the table is suggesting a
lower target score. As an explanation, Twenty20 has more aggressive batting than one-
day cricket throughout the innings. One-day cricket begins its innings in a less aggressive
fashion. Therefore, at the “halfway” point of an innings, Twenty20 has used up more of
its resources. Thus the Duckworth-Lewis table (Table 2.2) does not seem to capture this
characteristic as it is based on scoring patterns of one-day cricket. So, the target of 106
runs seems to be more sensible in Twenty20 cricket.
Figure 3.1: Heat map of the absolute differences between the Duckworth-Lewis resource
table (Table 2.1) and the nonparametric resource table based on Gibbs sampling (Table
3.5). Darker shades indicate larger differences.
Figure 3.2 shows the heat map of the absolute differences between the resource table (Table
3.6) and the Duckworth-Lewis resource table for Twenty20 (Table 2.2). In this figure, we can
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 26
see that the greatest difference lies in the top-right hand corner and the bottom-left hand
corner of the heat map. We can also see a little difference when 11 to 16 overs are remaining
and 4 wickets are lost. In these situations we see that the new resource table assigns nearly
6% to 8% more resources than the resource table in Table 3.5. Thus it assigns more resources
in these situations than the Duckworth-Lewis table in Table 2.2, when our intuitions tells us
that fewer resources should be available in these areas. A meagre difference also lies when
2 to 4 overs are left and 9 wickets are lost, though it can be regarded as insignificant.
The heatmap given in Figure 3.3 shows the absolute difference of the two new suggested
resource tables (Table 3.5 and Table 3.6). These two tables differ in situations where 8 to
15 overs are left with no wickets lost and also where 9 to 19 overs are available with 1 to
5 wickets taken. As mentioned earlier, Table 3.6 assigns more resources than Table 3.5 in
these areas whereas that is not desirable. These situations arise often in Twenty20 cricket
matches. We thus regard Table 3.5 preferable to Table 3.6 since Table 3.6 does not seem to
capture the aggressive scoring pattern of the Twenty20 game.
3.6 Discussion
In this section, our intention is to provide a discussion about the use of alternate resource
tables for Twenty20. We recall that we have only considered international Twenty20 matches
played between ICC teams up to November 9, 2009. Until that time there were only four
such instances where the Duckworth-Lewis method was applied.
To demonstrate our proposed resource table, we revisit the match between England
and West Indies during the 2009 World Cup on June 15, 2009. This was a very im-
portant game as the winner advanced to the semi-finals. In the first innings, England
scored 161 runs at the expense of 6 wickets using 100% of their resources. The second
innings was shortened to 9 overs with a target of 80 runs for West Indies. West Indies
scored 82 runs in 8.2 overs with 5 wickets still in hand and eliminated England from the
tournament. The English fans were upset and a report in the Guardian claimed that
the Duckworth-Lewis method will be reviewed to take into account Twenty20 matches
(http://www.cricinfo.com/ci-icc/content/story/409482.html). Note that in this match, the
Professional Edition of the Duckworth-Lewis Table (which is not available from the 2008-
2009 ICC Playing Handbook) was used to set the target score of 80. If we use the scaled
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 27
Figure 3.2: Heat map of the absolute differences between the Duckworth-Lewis resource
table (Table 2.1) and the nonparametric resource table based on Gibbs sampling (Table
3.6). Darker shades indicate larger differences.
Duckworth-Lewis Table given in Table 2.2 (based on the Standard Edition), the target score
would have been (0.518)161 → 84 runs. However, the alternative resource table in Table
3.5, sets an even higher target score of (0.579)161 → 94 runs. And if we use the alternate
resource table in Table 3.6, the target score is set at (0.513)161 → 83 runs.
In a Twenty20 match between England and South Africa on November 13, 2009, the
Duckworth-Lewis method was again used. In that match, England scored 202 runs losing 6
wickets in the full first innings. But due to rain, the second innings was reduced to 13 overs
with a target score of 129 using the Duckworth-Lewis method. But South Africa scored 127
runs losing 3 wickets in 13 overs. Thus England won the match. To set the target score
the Professional Edition of the Duckworth-Lewis table was used. Using Table 2.2 (based on
the Standard Edition), the target score is (0.71)202 → 144 runs. If we use our Table 3.5,
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 28
Figure 3.3: Heat map of the absolute differences between the nonparametric resource table
based on Gibbs sampling (Table 3.5) and the nonparametric resource table based on Gibbs
sampling (Table 3.6). Darker shades indicate larger differences.
the target score is (0.715)202 → 145 runs and Table 3.6 gives a target of (0.678)202 → 137
runs. We see that all targets are higher than the target score used in the match. Taking
into account the aggressive scoring pattern of the Twenty20 game, we can say that 145 runs
is a preferable target in 13 overs where the average runs scored in the first innings is 202.
Both of these examples question the validity of the existing Duckworth-Lewis method
when applied to Twenty20. It is important to emphasize that we are not claiming our non-
parametric resource table (Table 3.5) is a replacement for the Duckworth-Lewis resource
table in Twenty20. Our resource table is based on only 85 matches. This is too small a
sample to provide confident table entries. However, the table does suggest that there may
be a significant difference in the scoring pattern of one-day cricket and Twenty20 cricket.
When more Twenty20 matches become available, we endorse a review of the use of the
CHAPTER 3. NEW RESOURCE TABLES FOR TWENTY20 29
Conclusion
The goal of this project was to develop a resource table for the resetting of target scores
in interrupted Twenty20 matches. In 1998, Duckworth and Lewis developed a resource
table for one-day cricket. However, after the introduction of Twenty20 cricket, the existing
Duckworth-Lewis table is used for resetting target scores in Twenty20. But it is a known
fact that the scoring pattern in Twenty20 is much more aggressive than it is in one-day
cricket. Thus the question arises whether it is sensible to use the Duckworth-Lewis table
for Twenty20.
In this project we tried to produce an alternate resource table taking into account the
scoring pattern of the Twenty20 game. We tried to frame our resource table in a nonpara-
metric fashion using isotonic regression. But due to the lack of data or no data in some of
the match situations, the resultant resource table remains incomplete. We therefore con-
sidered a Bayesian model as it has the capability to impute missing values. Posterior mean
estimates obtained from Gibbs sampling provide us with a new nonparametric resource table
for Twenty20 cricket.
In chapter 3, we argued that Table 3.5 exhibits the characteristics of the scoring patterns
of Twenty20 game. This table suggests that there are differences between the scoring rates
in the two types of limited overs cricket. We reiterate that we are not claiming that our
resource table is a viable replacement of the existing Duckworth-Lewis resource table since
our table is based on a very small amount of data. But with the availability of more data,
a review of the Duckworth-Lewis approach should be undertaken.
30
Bibliography
[1] Bhattacharya, R., Gill, P.S. and Swartz, T.B. Duckworth-Lewis and Twenty20 Cricket.
under review, 2010.
[2] Bril, G., Dykstra, R.L., Pillers, C. and Robertson. Isotonic regression in two independent
variables. Journal of the Royal Statistical Society, 33:352–357, 1984.
[3] Duckworth, F.C. and Lewis, A.J. A fair method for resetting targets in one-day cricket
matches. Journal of the Operational Research Society, 49:220–227, 1998.
[4] Duckworth, F.C. and Lewis, A.J. A successful operational research intervention in one-
day cricket. Journal of the Operational Research Society, 55:749–759, 2004.
[5] Dykstra, R.L. and Robertson, T. An algorithm for isotonic regression for two or more
independent variables. Annals of Statistics, 10:708–716, 1982.
[6] Swartz, T.B., Gill, P.S. and Muthukumarana, S. Modelling and simulation for one-day
cricket. The Canadian Journal of Statistics, 37:143–160, 2009.
31