A Time Series Investigation of the Duckworth-Lewis Method in One Day International Cricket

cole Polytechnique Fdrale de Lausanne Sui Kai Wong and Anis Younes
Abstract We evaluate the use of the Duckworth-Lewis (D/L) method in limited overs cricket matches. We rst consider the time series of the runs scored in cricket matches and test the validity of the results predicted by the D/L method by examining the errors of the D/L predicted target score with the actual score. We will then enhance the D/L table with results from our time series analysis and design an alternative model for truncated cricket matches. The enhanced D/L table will be based on predictions of the time series model as well as cubic spline smoothing.

Part I

Introduction
The game of Cricket is a popular sport in many Commonwealth countries. It is a bat-and-ball game in which 2 teams of 11 will take turns to bat and the team with the highest score wins. In traditional test matches, cricket is played over 5 days in which the teams bowl and bat twice. In an attempt to garner greater interest, the games governing body, the International Cricket Council introduced shortened versions of the games. In limited-overs cricket matches, target scores have to be predicted in case the game is cut short due to external events (rain, bad lighting, etc...). Duckworth and Lewis (1998)[1] introduced a table which extrapolates the target score based on the number of resources available left at the teams disposal. The Duckworth-Lewis method considers the number of wickets on hand and the number of overs remaining as resources. As the game progresses, the batting team uses up their resources and in case of interruption, their resource pool is diminished. At the end of the game, their scores are scaled by a factor inversely proportionate to their resource allocated. Since its adoption by the ICC for the limited-overs matches, the Duckworth-Lewis method has been the target of criticism [2] by the press and cricket teams alike. In this project we will examine the validity of the Duckworth-Lewis method by pitting its predicted target runs against actual scores of One Day Internationals. In this section we will next provide a brief introduction to limited overs cricket and the DuckworthLewis method. At the end of Part I we will explain how we extracted the data and what methods we will be using in the time series investigation of the data. In Part II we will examine the time series properties of the data that we have extracted. We will rst examine the runs, the raw score that we have extracted and we will try to t a time series model to the score in one inning. Next, we will use the D/L resource table as a non-parametric proxy for the expected consumption level 1

of resources to investigate the shortfall or surplus, if any, of the resource allocation. We look to examine the time series of errors (errors being the dierence between the predicted DuckworthLewis score and the actual score). Once we have an appropriate model, we will rst predict using the errors model a more accurate target runs score and nally we will apply cubic spline smoothing that will ll in target scores where we have missing empirical data. The end result is an enhanced Duckworth-Lewis table. Part III will present our conclusion and some concluding remarks.

Limited Overs Cricket and One Day Internationals

We dene and explain some of the cricket terminologies used in this project. At any one time, there are 13 players on the eld 11 from the elding (non-batting) team and 2 opposing batsmen. A player from the elding team will deliver, or bowl, the ball to either of the batsmen. The bowling teams objective is to dismiss the batsman by striking the wicket or catching the ungrounded ball midair. On the receiving end, the batsmen protect the wicket by striking the ball and score runs in between each delivery. The wicket is the target on the eld in which bowlers deliver the ball towards. There are 2 wickets planted 20 m apart on the eld, 1 behind each batsman for him to protect. The bowlers will try to dismiss a batsman by knocking out the wicket with each delivery. The batsman strikes the ball away, and run in between the wickets. Should the ball be struck over the eld boundaries, the batsman will receive 4 or 6 runs. The batsmen pairs up in a partnership. When a batsman is dismissed, another from his team replaces him in the partnership. This goes on until the 10th wicket is struck and the last batsman is left without a partner. At that point, the batting team is all out and the inning is ended. Innings, Overs and Balls are counts of deliveries the batsman faces. An over is made up of 6 balls. An inning is the collection of all balls faced by the batting team. Depending on the format of the game, there can be limited or unlimited numbers of overs per inning. At the end of an inning, the team switches over and the batting team takes to the eld and the elding team will bat. To shorten the game, there are two formats of limited overs cricket matches: the one-day internationals (ODI) and twenty20 (T20) formats. In these formats, there are a maximum numbers of overs, which each team will play (50 for ODIs and 20 for T20s). In our project, we will consider the ODIs. In ODIs, The inning will end with either all 10 wickets fall or all 300 balls bowled, whichever comes rst. The new formats call for dierent strategies. In test matches, since there are no limits to the number of overs, the most common strategy is a conservative one as batsmen usually vie for longevity in the inning. By being defensive, they are less at risk to dismissal. On the contrary for the limited overs matches, audaces fortuna iuvat (Fortune favours the brave) is the name of the game. Batsmen would try to aim for the boundaries to accumulate more runs before the overs are up. At the same time, they risk being dismissed so it is a trade o they have to manage.

The Duckworth-Lewis Method

Limited overs games are susceptible to interruptions. In occasions where these games need to be interrupted for example, in bad weather or bad light, the two innings may not be symmetric. In such instances, fair comparison between scores is dicult and many methods have been proposed for this purpose.

The most primitive way is a comparison between the run rates. The ratios of the runs against the number of overs played are compared and team with the higher run rate is declared the winner. This is a straight-line extrapolation of the truncated score to the 50 overs. However, this is hardly the case in reality. The run rate evolves not in a straight-line fashion with the overs as batsmen manage their risk of dismissal. Typically, the teams with wickets at their disposal will score more runs towards the end of the inning. The D/L method provides an alternative to the run rate method. The D/L method considers the number of wickets on hand and the number of overs remaining as resources. As the game progresses, the batting team uses up their resources and in case of interruption, their resource pool is diminished. At the end of the game, their scores are scaled by a factor inversely proportionate to their resource allocated. For commercial reasons, only parts of the D/L formulation are disclosed. Partial information suggests that the full formula is an extension to a two factor exponential equation [1]: bu F (w)

Z(u, w) = Z0 F (w) 1 exp

where Z(u, w) denotes the average expected number of runs with u overs remaining and w wickets taken. Z0 F (w) denotes represents the expected number of runs with w wickets taken as u tends to innity. The parameters b, Z0 and F (w) are to estimated from match data. By including the wicket count, DL method is already a more comprehensive one as it considers the style of play given that one has fewer wickets on hand ceteris paribus. One huge advantage for the D/L method is that it is a generic model in which players can use without the technical knowhow. However, this method is by no means foolproof. The comparison of ODI data Z(u, w) vis--vis test cricket data Z0 F (w) meant that the D/L method does not incorporate the time urgency of ODIs. Furthermore, it is unable to the streaks of good or bad form the batsmen undergo in matches as a time series analysis would. Despite the consideration, we can assume that the D/L method is robust in modeling the resource allocation. We consider the example when the team batting rst nished their inning uninterrupted, scoring 300 runs. The second inning started with the team batting second having 100% of their resource1 . After 20 overs and 2 dismissals, rain interrupted the game. It is later resumed but the umpire declared that only 10 more overs could be played due to time constraint. The team has 67.3% of their resource left when the game was rst stopped and 30.8% resource when play resumed. In other words they have lost 36.5% of their resources due to the rain. They will be compensated at the end of their inning by scaling their score by a factor 100/(100-36.5). Similarly, one can scale down the rst teams score of 300 by 63.5% to set the target score of 190.5 for the team batting second to outrun. The D/L Standard Edition Table is given below [3].
1 This will depend on whether the game started out as per normal or if the umpire has already decided that the match will be truncated eventually. Should the umpire already predicted rain and see t that only 30 overs can be played, the team batting second is considered start with 73.5% of the resource.

50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1

10 100.0 99.1 98.1 97.1 96.1 95.0 93.9 92.8 91.7 90.5 89.3 88.0 86.7 85.4 84.1 82.7 81.3 79.8 78.3 76.7 75.1 73.5 71.8 70.1 68.3 66.5 64.6 62.7 60.7 58.7 56.6 54.4 52.2 49.9 47.6 45.2 42.7 40.2 37.6 34.9 32.1 29.3 26.4 23.4 20.3 17.2 13.9 10.6 7.2 3.6

9 93.4 92.6 91.7 90.9 90.0 89.1 88.2 87.3 86.3 85.3 84.2 83.1 82.0 80.9 79.7 78.5 77.2 75.9 74.6 73.2 71.8 70.3 68.8 67.2 65.6 63.9 62.2 60.4 58.6 56.7 54.8 52.8 50.7 48.5 46.3 44.1 41.7 39.3 36.8 34.2 31.6 28.9 26.0 23.1 20.1 17.0 13.8 10.5 7.1 3.6

8 85.1 84.5 83.8 83.2 82.5 81.8 81.0 80.3 79.5 78.7 77.8 76.9 76.0 75.0 74.1 73.0 72.0 70.9 69.7 68.6 67.3 66.1 64.8 63.4 62.0 60.5 59.0 57.4 55.8 54.1 52.4 50.5 48.6 46.7 44.7 42.6 40.4 38.1 35.8 33.4 30.8 28.2 25.5 22.7 19.8 16.8 13.7 10.4 7.1 3.6

7 74.9 74.4 74.0 73.5 73.0 72.5 72.0 71.4 70.9 70.3 69.6 69.0 68.3 67.6 66.8 66.0 65.2 64.4 63.5 62.5 61.6 60.5 59.5 58.4 57.2 56.0 54.7 53.4 52.0 50.6 49.1 47.5 45.9 44.1 42.3 40.5 38.5 36.5 34.3 32.1 29.8 27.4 24.8 22.2 19.4 16.5 13.5 10.3 7.0 3.6

Wickets on hand 6 5 62.7 49.0 62.5 48.9 62.2 48.8 61.9 48.6 61.6 48.5 61.3 48.4 61.0 48.3 60.7 48.1 60.3 47.9 59.9 47.8 59.5 47.6 59.1 47.4 58.7 47.1 58.2 46.9 57.7 46.6 57.2 46.4 56.6 46.1 56.0 45.8 55.4 45.4 54.8 45.1 54.1 44.7 53.4 44.2 52.6 43.8 51.8 43.3 50.9 42.8 50.0 42.2 49.0 41.6 48.0 40.9 47.0 40.2 45.8 39.4 44.6 38.6 43.4 37.7 42.0 36.8 40.6 35.8 39.1 34.7 37.6 33.5 35.9 32.2 34.2 30.8 32.3 29.4 30.4 27.8 28.3 26.1 26.1 24.2 23.8 22.3 21.4 20.1 18.8 17.8 16.1 15.4 13.2 12.7 10.2 9.9 7.0 6.8 3.6 3.5

4 34.9 34.9 34.9 34.9 34.8 34.8 34.8 34.7 34.7 34.6 34.6 34.5 34.5 34.4 34.3 34.2 34.1 34.0 33.9 33.7 33.6 33.4 33.2 33.0 32.8 32.6 32.3 32.0 31.6 31.2 30.8 30.3 29.8 29.2 28.5 27.8 27.0 26.1 25.1 24.0 22.8 21.4 19.9 18.2 16.4 14.3 12.0 9.5 6.6 3.5

3 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 21.9 21.9 21.9 21.9 21.9 21.9 21.9 21.9 21.8 21.8 21.8 21.7 21.7 21.6 21.6 21.5 21.4 21.3 21.2 21.1 20.9 20.7 20.5 20.2 19.9 19.5 19.0 18.5 17.9 17.1 16.2 15.2 13.9 12.5 10.7 8.7 6.2 3.4

2 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.9 11.8 11.8 11.8 11.7 11.6 11.5 11.4 11.2 10.9 10.5 10.1 9.4 8.4 7.2 5.5 3.2

1 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.7 4.6 4.6 4.5 4.2 3.7 2.5

Overs remaining

Figure 1: Duckworth-Lewis Resource Table [3]

In our project, we interpret the proportion of runs scored in each over given the number of wickets at hand as the proportion of the resource consumed on that situation. Thus, cumulatively over the entire match, all 100% of resources will be consumed.

Data Extraction and Description of Methods Used

Data on cricket matches was not readily available in a format we could use. We therefore had to parse through data from ESPNs cricinfo (www.espncricinfo.com) website to extract data on the matches played. In order to do that we wrote a Java code that crawls through the website and extracts the desired cricket data: innings, overs played, runs scored, and wickets left. Once we had the data we had to initially process the data, for this we used MATLAB to create the necessary tables and lter through falsely extracted data and some data which were unusable. MATLAB was also used to create some of the 3D plots we provide in our project. For the time series analysis of the data and the tted models we have used R. All in all, we have obtained the scores of 75 matches, and we have used the runs scored in the rst inning resulting in a total of 3659 overs played concatenated back to back. In the rst section of Part II we normalize the runs to represent the percentage of resources available to the team. This will allow us to compare across all matches the pattern and rates in which runs are made. We suspect that one-sided matches may have a dierent dynamics. Therefore, we only considered matches of full ICC members, namely Australia, Bangladesh, England, India, New Zealand, Pakistan, South Africa, Sri Lanka, West Indies and Zimbabwe. These 8 teams are considered to be in the highest tier amongst cricket-playing nations and matches between them should be a fairer representation of most matches. All the analysis methods used in our project are based on the material in [4], we have also used models explained in this book to t the data. Since our data has seasonality which is obvious from the data generating process (concatenating matches), we look rst to dierence the data in order to get rid of the seasonality factor, and therefore model the runs in one inning. The choice of Auto-Regressive Moving-Average (ARMA) models for the dierenced data is based on examining the second-order properties of the time series, mainly the Auto-Correlogram and Partial Auto-Correlogram plots which are the sampling equivalent of the autocorrelation and partial autocorrelation functions. We look to have relatively simple models so upon testing dierent orders, we chose not to have more than 6 parameters per model. The particular choice of model was based on minimizing the Akaike Information criteria (AIC) which oers a quantiable measure of the tradeo between the accuracy and complexity of the model. Once we have established a parsimonious model of the D/L errors, we will look at ways that will enhance the D/L method. Here we use a combination of two methods: forecasting using the D/L errors ARMA model to ll in the missing data in our empirical D/L table and nally using cubic spline smoothing per wicket resource to obtain an enhanced D/L table.

Part II

Time Series Investigation

1 Empirical Resource Data
In this section we will investigate the time series properties of the emprical data. As we mentioned in the Data Extraction Section, we have extracted data of the rst inning from 75 matches, resulting in a total of 3659 overs played. Now the average of overs played per inning in our data is close to, but not exactly 50. This is due to a few matches ending prematurely in the rst inning so they had slightly less than 50 overs played. In order to account for that we normalize the data, that is we divide the number of runs scored per over by the total number of overs played in this inning, and then multiply that number by 100. We denote this by the empirical resource data. rEmpirical,t =
ri,t ri,t
t

Oi 50

100% ,

where ri,t is the number of runs in over t of inning i, and Oi is the number of overs played in inning i.

Empirical Data Time Series

10 emp 0 0 2 4 6 8

1000

2000 Overs played

3000

Zoomed in Empirical Data Time Series

8 emp[1:250] 0 0 2 4 6

50

100

150

200

250

Overs played

Figure 2: From Top to Bottom: Time Series Plot of Empirical Data, Auto-Correlation, PartialAutocorrelation, and Spectral Plots.

A rsthand examination of our data demonstrates clear seasonality, this was expected as we have concatenated the innings back-to-back, we expect to have a seasonality factor of 50 overs. So we dierence the data taking into account the seasonality factor and look at the time series plots.
diff(emp, lag = 50)[1:1000]

Season Differenced Empirical Data Time Series

6 6 0

200

400 Overs played

600

800

1000

Season Differenced Data ACF

0.5 0.5

ACF

20

40 Lag

60

80

100

Season Differenced Data PACF

Partial ACF 0.0 0.4 0

20

40 Lag

60

80

100

Season Differenced Data Spectrum

spectrum

1e04

0.0

0.1

0.2

0.3

0.4

0.5

frequency bandwidth = 7.92e05

Figure 3: From Top to Bottom: Time Series Plot of Empirical Data, Auto-Correlation, PartialAutocorrelation, and Spectral Plots.

The time series plot shows that the the dierencing did produce a stationary process. Looking at the ACF and PACF plots clearly cut o at multiples of s=50. We examine SARIMA (p, 0, q) (0, 1, 0)50 models to model this empirical data (ARMA (p, q) model for the dierenced data). Notice the spectrum plot reveals the seasonal frequencies. We choose the Akaike Information Criteria to choose the appropriate model. It tends to choose models that are more complicated, however it is the most stringent criterion and will therefore be the Information criterion of choice upon which we decide which model to use. We will not examine cases where we have more than 6 parameters to estimate as we believe they are too complex and are computationally exhaustive for our purposes. 7

SARIMA (p, 0, q) (0, 1, 0)50 q 0 0 1 p 2 3 4 5 14415.815 14401.09 14347.96 14348.47 14349.98 14351.57 1 14404.41 14371.95 14348.76 14350.20 14279.01 14352.38 2 14350.89 14347.78 14349.74 14351.86 14280.04 3 14348.20 14349.75 14351.77 14352.67 4 14349.64 14351.64 14353.62 5 14351.63 14353.64 -

Table 1: Moving Average Model AIC values for the Empirical Data

It seems that a SARIMA (4, 0, 1) (0, 1, 0)50 model would best represent the D/L Errors with the lowest AIC value than all models, and we do not see any improvement in larger orders. Here the choice of model is obvious and there is no tradeo, yet the downside is that the model is rather complex and cumbersome to calculate its parameters. The SARIMA (4, 0, 1) (0, 1, 0)50 models parameters are given in the table below. 1 2 3 4 1 1.0363 (0.0166) 0.0686 (0.0239) -0.0978 (0.0239) -0.0411 (0.0167) -1.000 (0.001)

Table 2: MA(3) Model Parameters for the Empirical Data with standard errors

We now examine the suitability of the t with the following graphs testing the properties of the residuals.

Standardized Residuals
4 4 2 0 2

1000 Time

2000

3000

ACF of Residuals

ACF

0.0

0.4

0.8

10

15 Lag

20

25

30

35

p values for LjungBox statistic

q q q q q q q q

0.8

p value

0.4

0.0

q q q q q q q q q q q q q q q q q q q q q q q q

q q q q q

q q q q q q q q q q q

10

20 lag

30

40

50

Figure 4: From Top to Bottom: Time Series Plot of Standardized Residuals, Auto-Correlation of Residuals, p-values of Ljung-Box Test testing autocorrelation of Residuals.

Empirical Data Model Residual QQ

q q qq q q qq qq q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq qq qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q qq qq qq qq qq q q qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq q q qq qq qq qq qq qq qq qq qq qq qq qq qq qqq qq qq qq qqq qqq qqq qqq qq qq qq qqq qqq qqq qq qqq qqq qqq qq q qq q

Sample Quantiles

2
q

0 Theoretical Quantiles

Empirical Data Model Residual CPGRAM

1.0 0.0 0.2 0.4 0.6 0.8

0.0

0.1

0.2 frequency

0.3

0.4

0.5

Figure 5: Normal Q-Q Plot and Cumulative Periodogram of the Residuals.

Examining the plot of the time series of residuals, it seems like white noise. The ACF plot shows no autocorrelation among residuals, consistent with white noise property. The plot of signicance level of the Ljung-Box Test however reveals that the model residuals are not white noise, we do have dependene amongst residuals between lags 10 and 20. The QQ plot to test for the normality of residuals shows a very poor t to the data, the residuals dont seem to be normally distributed, they rather exhibit heavy skewness to the left and leptokurtic symptoms. Finally, the cumulative periodogram of the residuals stay within the bounds but we cannot claim the white noise property of the residuals and therefore conclude that the model is a bad t.

10

Therefore with our SARIMA (4, 0, 1) (0, 1, 0)50 model, we cannot predict the target percentage of resources and hence the target runs per over, but as we have explained, we have to account for the number of wickets in our prediction mechanism. Since the number of runs per over is also dependent on the number of wickets left, we have to account for that dimension as well. Duckworth and Lewis (1998) have suggested a resource table for predicting the target number of runs based on both overs left to play and wickets left. We now move on to the centerpiece of our project: examining the validity of the Duckworth-Lewis method in predicting target run scores.

Errors from Duckworth-Lewis Predicted Runs

In this section we focus on examining the Duckworth-Lewis Method for predicting target runs based on wickets and overs left. In particular we will look at individual runs per over rather than the cumulative runs stated in the Duckworth-Lewis table. This is because errors in the number of runs would eventually sum up, resulting in an autoregressive trend. The errors xw,t that we examine here are the dierence between the resources per over in the empirical data and the resources per over as projected by the Duckworth-Lewis Resource table: xt (w) = Rt (w) Rt (w) where Rt (w) is the total percentage of available resources in over t , given w wickets left of the empirical data, and Rt (w) is the target percentage of resources in over t , given w wickets left from the Duckworth-Lewis table. By referring to the D/L resource table, we assumed the D/L table is robust in representing the mean resource used in the over given the wicket count. Thus, our approach here will both investigate the time series property of the errors as well as the overall goodness of t of the D/L table. Another implication of taking the dierence is that the errors are independent of the number of wickets on hand. A rsthand examination of our data demonstrates clear seasonality, this was expected as we have concatenated the innings back-to-back, we expect to have a seasonality factor of 50 overs.

11

D/L Error Time Series

6 D/L Error 4 0 0 2 4

1000

2000 Overs Played

3000

Zoomed D/L Error Time Series

6 D/L Error 2 0 0 2 4

50

100

150 Overs Played

200

250

300

Figure 6: Time Series Plot of xw,t and a zoomed-in plot demonstrating clear seasonality of 50 overs.

In order to account for this seasonality, we perform a rst order dierencing of the seasonality factor. So in terms of the SARIMA factors, D=1 and s=50. Next, we examine the plots of the seasonality-neutral (dierenced data), its Auto-Correlation and Partial Auto-Correlation plots.

12

Season Differenced D/L Error Time Series

D/L Error 5 5

200

400 Overs Played

600

800

1000

Season Differenced D/L Error ACF

0.5 0.5

ACF

20

40 Lag

60

80

100

Season Differenced D/L Error PACF

Partial ACF 0.0 0.4 0

20

40 Lag

60

80

100

Season Differenced D/L Error Spectrum

spectrum 1e+02 1e04

0.0

0.1

0.2

0.3

0.4

0.5

frequency bandwidth = 7.92e05

Figure 7: From Top to Bottom: Time Series Plot of season-dierenced xw,t , Auto-Correlation, Partial-Autocorrelation, and Spectral Plots.

The top plot reveals that the dierencing produces a more stationary looking process. Looking at the above plots, we can clearly see that both ACF, and PACF plots clearly tail o at multiples of s=50. The ACF demonstrates signicant autocorrelations up till a lag of 3, while the PACF tails o at a lag of 2. We therefore look to examine SARIMA (p, 0, q) (0, 1, 0)50 models with p and q values similar to those mentioned. In other words this is just an ARMA (p, q) model for the dierenced data. One should take note of the spectrum plot which demonstrates the seasonality at the seasonal frequencies.

Seasonal Auto-Regressive Integrated Moving-Average, SARIMA (1, 0, 3) (0, 1, 0)50 Model

We suggest tting an SARIMA (p, 0, q) (0, 1, 0)50 model (ARMA (p, q) of the 50th dierence data i.e. ARIMA(p, 50, q) ) to the Duckworth-Lewis Errors. In order to do that we examine the 13

AIC of possible SARIMA (p, 0, q) (0, 1, 0)50 models. The choice of the AIC is for the same reasons mentioned in Section 1.
SARIMA (p, 0, q) (0, 1, 0)50 q 0 0 1 p 2 3 4 5 14518.74 14491.96 14395.93 14382.95 14380.08 14377.65 1 14499.01 14381.39 14371.75 14360.93 14362.41 14364.25 2 14414.56 14374,33 14368.56 14368.99 14364.56 3 14400.49 14360.57 14362.28 14316 4 14392.56 14362.29 14364.84 5 14386.57 14364,25

Table 3: SARIMA (p, 0, q) (0, 1, 0)50 Model AIC values for the Duckworth-Lewis Errors.

Combining the results of the above table with what we deduced from the ACF and PACF plots, it seems that a SARIMA (3, 0, 3) (0, 1, 0)50 has the lowest AIC value. But we choose the SARIMA (1, 0, 3) (0, 1, 0)50 model to represent the D/L Errors with the second lowest AIC value than all models, and a much simpler model. Here the choice of model is a tradeo between a more complex model: the SARIMA (3, 0, 3) (0, 1, 0)50 and one which is simpler, the SARIMA (1, 0, 3) (0, 1, 0)50 but also represents well the data. When simulated in R, the estimated variance of residuals of the SARIMA (3, 0, 3) (0, 1, 0)50 is degenerate (NaN) which also makes us choose the former model. The SARIMA (1, 0, 3) (0, 1, 0)50 models parameters are given in the table below. 1 1 2 3 0.9142 (0.0229) -0.8596 (0.0287) 0.0990 (0.0221) -0.0760 (0.0179)

Table 4: SARIMA (1, 0, 3) (0, 1, 0)50 Model Parameters (with standard errors) for the Duckworth-Lewis Errors

In other words our model is the following: (B)(I B 50 )xt (w) = (B)t , where t N (0, 2 ), and B is the backshift operator. We now examine the suitability of the t with the following graphs testing the properties of the residuals up till lag 50 which is the period of our seasonality (50 overs per inning).

14

Standardized Residuals
4 4 2 0 2

1000 Time

2000

3000

ACF of Residuals

ACF

0.0

0.4

0.8

10

15 Lag

20

25

30

35

p values for LjungBox statistic

q q q q q q q q

p value

0.8

0.4

q q q q q q q q q q q q q q q q q

q q q

q q q q q q q

q q q q q

0.0

q q q q q q q q q

10

20 lag

30

40

50

Figure 8: Model Fit: (From Top to Bottom) Time Series Plot of Standardized Residuals, AutoCorrelation of Residuals, p-values of Ljung-Box Test testing autocorrelation of Residuals.

15

D/L Error Model Residual QQ

q

qq q q q q q q q qq qq q qq qq qq q q q qq qq q q q q q q q q q q q q q q q q qq qq q q q qq q q qq qq qq qq q q q qq qq qq qq q q q q q q qq qq qq qq q q q qq q q qq qq qq qq q qq qq q q qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq qq q q qq qq qq qq qq qq qq qq qq qq qq qq qq qq q qq qq qq qq qq qq qq qq qq q qq qq qq q q qq qq qq qq qq qq qq q qq qq qq qq qq qq qq qq qq q q q q q qq qq qq q q q q qq q q q q qq q q q q q q q q q q q qq q qq qq qq q q q q q q q q q q q

Sample Quantiles

0 Theoretical Quantiles

D/L Error Model Residual CPGRAM

1.0 0.0 0.2 0.4 0.6 0.8

0.0

0.1

0.2 frequency

0.3

0.4

0.5

Figure 9: Normal Q-Q Plot and Cumulative Periodogram of the Residuals.

Examining the plot of the time series of residuals, we can see that it looks like stationary white noise. The ACF plot shows no autocorrelation among residuals, consistent with white noise property. The plot of signicance level of the Ljung-Box Test reveals that the model residuals are indeed white noise (a slight exception is present in the case of lags 10 till 14). We also examine the QQ plot to test for the normality of residuals: the linearity of the model shows an excellent t to the data. The ordered values of the residuals and the theoretical quantiles seem almost perfectly aligned with a normal distribution with slightly fatter tails, but this slight leptokurtocy is insignicant. Finally, the cumulative periodogram of the residuals stay within the bounds suggesting they are indeed white noise. 16

If we were to go further to check wheter the residuals have a t-distribution, we examine the QQ plot of residuals versus the t-distribution theoretical quantiles with 3 degrees of freedom, which seem like a slightly better t to the residuals than a normal distribution but outliers are signicant.

t(3) QQ Plot
q

q q q q q q q q q q q qq qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q qq q q q q q

q q

Sample Quantiles

20

10

10

20

30

rt(3659, df = 3)

Figure 10: Student t-distribution with 3 Degrees of Freedom Q-Q Plot of Residuals.

Therefore our SARIMA (1, 0, 3) (0, 1, 0)50 model ts extremely well the errors of resources per over as predicted by the Duckworth-Lewis method and we will use this model to predict the discrepancy of the D/L method and hence predict a more realistic score based on the empirical data.

Enhanced D/L table

Theoretical D/L table

The empirical D/L table is constructed using the Empirical Resource Data which was modelled in Section 1. Resources were rst sorted by overs remaining and wickets in hand, we then cumulated the values to obtain a cumulative resource table for the inning. The 3D plots below shows the Empirical Cumulative D/L table of resources, the Theoretical Cumulative D/L table of resources (see Figure 1), and the heat map displays the dierences between the two plots (darker shades translate to larger dierences). There are much missing data from the samples we collected, especially in the quadrant with many wickets lost within few overs. We can consider the D/L resource table to be composed of 10 individual time series one for each wicket. At the beginning of the game, the time series starts from the 50th over with 10 wickets on hand. As the game progresses and wickets lost, the time series hop to chains down the wicket count. It is very unlikely in reality that most of the batsmen get struck out already early in the game. As such, information may be lacking in that quadrant.

17

H a Ma o Df rn e b te nT e rta a d e t p f i e c s ew e h oei l n f e c E i a C mu teDLR s uc s mp i l u l i / e o re r c av

Figure 11: Top Left to Right: Theoretical D/L Cumulative Resources, Empirical D/L Cumulative Resources. Bottom: Heat map of the dierence.

Notice the dark shades especially in the o-diagonal resources where the D/L method fails to replicate the empirical data.

Enhanced D/L table

Using the SARIMA (1, 0, 3) (0, 1, 0)50 model, we use forward prediction to ll in the values for the Empirical D/L Error that was modelled in Section 2. The prediction is done per wicket basis, that is D/L errors were rst sorted by overs remaining and wickets in hand to obtain an Empirical D/L Error table. The time series of errors per wicket (column) was used for forward prediction and lling in the missing values in our table. Next, we cumulated the D/L errors obtained from the rst step and added them to the Theoretical Cumulative D/L table of resources. In order to account for the values where we were not able to use our forward prediction (due to lack of initial data points), we use the cubic spline smoothing function in R applied to the 10 time series of columns. The result is an Enhanced Cumulative D/L 18

table of resources. We also plot to compare: the Empirical Cumulative D/L table of resources, the Enhanced Cumulative D/L table of resources (see Figure 1), and the heat map displays the dierences between the two plots (darker shades translate to larger dierences) .

H a Ma o Df rn eb te nE h n e a d e t p f i e c ew e n a c d n f e E i a C mu teDLR s uc s mp i l u l i / e o re r c av

Figure 12: Top Left to Right: Enhanced D/L Cumulative Resources, Empirical D/L Cumulative Resources. Bottom: Heat map of the dierence.

Notice the dominance of the very light shades in the heat map compared with the heat map of the dierence with the Theoretical D/L table where darker shades dominate. This shows that our Enhanced D/L table replicates the empirical data better than the theoretical D/L, the improvement is most noticeable in the o-diagonal resources. Our Enhanced D/L table values are given below.

19

50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1

10 100.0 98.5 97.1 95.5 94.0 92.3 90.6 88.8 87.0 85.3 83.5 81.8 80.1 78.5 76.9 75.4 73.9 72.5 71.2 69.9 68.6 67.2 65.7 64.1 62.5 61.0 59.5 58.1 56.7 55.4 53.8 52.1 50.2 48.2 46.3 44.6 43.2 41.8 40.5 38.9 36.9 34.4 31.5 28.3 24.8 20.9 16.9 13.1 9.2 4.7

9 93.1 92.0 90.9 89.7 88.5 87.2 85.8 84.4 82.9 81.4 79.8 78.2 76.6 74.9 73.3 71.6 70.0 68.4 66.9 65.4 64.0 62.6 61.1 59.6 58.1 56.5 54.9 53.2 51.5 49.7 48.0 46.1 44.3 42.4 40.6 38.7 36.9 35.1 33.1 30.8 28.2 25.3 22.2 19.0 16.0 13.2 10.5 7.7 5.1 2.8

8 84.2 83.2 82.2 81.3 80.3 78.9 77.6 76.6 75.2 73.9 72.7 71.7 70.6 69.2 67.8 66.6 65.4 64.1 62.8 61.3 60.0 58.7 57.3 56.0 54.8 53.4 52.0 50.6 49.0 47.3 45.6 43.9 42.4 40.9 39.4 37.7 35.7 33.8 32.3 30.9 29.4 27.4 25.0 22.8 20.9 18.8 16.6 13.6 9.6 5.0

7 77.5 76.7 76.1 75.3 74.5 73.7 72.9 71.9 71.1 70.1 68.6 67.9 66.9 65.5 64.7 63.6 62.5 61.3 60.3 58.8 57.3 55.8 54.4 53.2 51.9 50.3 48.6 47.0 45.4 43.7 42.2 40.1 38.7 37.0 35.5 33.7 32.3 30.6 28.9 27.1 25.3 23.5 21.4 18.8 16.7 14.6 11.6 9.0 6.5 3.2

Wickets on hand 6 5 63.3 60.2 63.0 60.1 62.5 59.9 62.0 59.6 61.5 59.4 61.1 59.3 60.6 59.1 60.1 58.8 59.5 58.5 58.8 58.3 58.2 58.0 57.5 57.6 56.8 57.2 56.2 56.8 55.7 56.4 55.3 56.0 54.5 55.6 53.2 55.2 52.2 54.8 51.0 53.8 49.9 52.3 48.6 50.8 47.1 49.5 45.8 48.2 44.6 46.9 43.2 45.4 41.8 43.9 40.3 42.6 38.8 41.2 37.3 39.7 35.7 38.3 34.0 36.9 32.3 35.4 30.7 33.7 29.2 32.0 27.6 30.3 26.1 28.8 24.7 27.5 23.0 26.2 21.2 24.7 19.4 23.3 17.7 21.8 15.7 20.3 13.6 18.7 11.5 16.8 9.1 14.4 6.3 11.8 3.4 9.0 1.5 5.8 0.8 2.5

4 52.2 52.2 52.2 52.2 52.1 52.1 52.1 51.9 51.9 51.7 51.7 51.6 51.6 51.4 51.3 51.1 50.9 50.8 50.6 50.3 50.1 49.0 48.7 46.9 45.1 44.4 42.6 40.8 40.0 39.1 37.7 36.8 34.1 32.8 32.0 30.1 28.1 27.0 25.6 24.0 22.6 21.3 19.9 18.2 16.8 14.6 12.7 10.2 6.9 2.9

3 36.5 36.5 36.5 36.5 36.5 36.5 36.5 36.5 36.5 36.5 36.5 36.5 36.4 36.4 36.4 36.4 36.4 36.4 36.4 36.3 36.2 36.2 36.2 36.1 36.0 35.9 35.9 35.7 35.6 35.3 34.5 33.5 32.7 31.4 29.7 28.1 26.7 25.3 23.7 21.8 20.3 19.2 18.1 16.6 14.8 12.8 10.8 8.8 6.4 3.7

2 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.7 30.8 30.8 30.8 30.9 30.8 30.4 29.7 28.6 27.1 25.5 23.9 22.4 20.8 19.2 17.3 15.4 13.5 11.5 9.6 7.6 5.6 3.5

1 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.7 8.8 8.8 8.8 8.6 8.3 7.6 6.7 5.5 4.1 2.6 1.0

Figure 13: Enhanced Duckworth-Lewis Resource Table

Overs remaining

20

Part III

Conclusion
In this project we looked to examine the Duckworth-Lewis method which predicts a target run score in limited overs cricket. We rst examined the time series properties of normalized runs and found that a SARIMA (4, 0, 1) (0, 1, 0)50 was not sucient to model the data and therefore we might need a more complex model, this was expected since the model accounted for runs using overs (time) as the only dimension whereas we need to account for the number of wickets. The Duckworth-Lewis method provides a semi-parametric table that allows for predicting target score runs in the event of the truncation of a match. So we next examined the time series of errors which are the dierences between the runs per over in the empirical data and the runs per over as projected by the Duckworth-Lewis Resource table. We nd that a SARIMA (1, 0, 3) (0, 1, 0)50 model ts the errors extremely well. We then used this model to ll in missing values in our empirical D/L table. Finally, we smoothed the table per column of wicket resources using cubic splines to obtain an enhanced D/L table which seems to better mimic empirical data than the theoretical D/L table. We observed that there is indeed a time series aecting the gameplay. As such, we can use the tted model to predict the nal scores. An alternative to using the resource table is to use the time series model directly to predict the nal score. The time series model would incorporate the dynamics of that specic match. At this point, the run rate will also play a part in determining the result. We can extend the time series approach to predict test cricket scores. As mentioned above, test cricket are played over 5 days. Each team can play 2 innings of unlimited numbers of overs each (instead of 1 50-over inning) over the 5 days. The winner will be determined by the nal accumulated run counts. After 5 days of play, there are (many) instances where the game will still end up as a draw due to the lack of time. For example, consider that team A has batted twice and team B is batting their second inning nearing the end of the fth day. Team B is trailing by 400 runs with only 1 wicket on hand. Considering that the last batsmen are usually the weakest, it is almost impossible for team B to win the match. However, if the last batsmen can hold up until the end of the day, they can still salvage a draw despite team As commanding lead. Suppose we can model the fall of wicket like a stopping time or waiting time problem, then we can extend the ARMA model to predict the expected number of runs at the expected stopping time. We can expand the set of resources to incorporate powerplays. Powerplays are segment of the games when a restriction is imposed to the elding team. Fielding restrictions limit the number of players outside a certain radius on the eld. This is to promote aggressive batting and scoring. The rst 10 overs of an inning is a mandatory period of powerplay. Both teams will need to call for another 5 overs of powerplays whenever they deem advantageous. As there is more aggressive style of play during these powerplays, we can consider these 20 overs can be considered another set of resource and further enhance the D/L table. An interesting investigation would be to backtest the dierent resource tables. We can evaluate if the match results will dier given truncation of various length at various stages of the game. Due to the multiple permutations of truncations we may have, there are bound to be some cases where the table fails to declare the ultimate winner. We understand that both tables are not comprehensive, but its main advantage that it can be easily implemented still remains. Using the time series model to predict the nal score for truncated matches may be less transparent and more

21

complicated than necessary.

References
[1] F. C. Duckworth and A. J. Lewis, A fair method for resetting the target in interrupted one-day cricket matches, The Journal of the Operational Research Society, vol. 49, no. 3, pp. pp. 220227, 1998. [Online]. Available: http://www.jstor.org/stable/3010471 [2] R. Ramachandran. (2002, November 23 - December 6) For a fair formula. [Online]. Available: http://www.hinduonnet.com/ine/1924/stories/20021206004410400.htm [3] I. C. Council. Duckworth lewis standard edition table. [Online]. Available: cricket.yahoo.net http://icc-

[4] A. Davison, Time Series, C. o. S. Mathematics Institute (MATHAA), Ed. cole Polytechnique Fdrale de Lausanne, 2011.

22

Appendix
Sample R code:
## Time Series plots of D/L Errors ## par(mfrow=c(2,1)) plot(ind,type="l",main="D/L Error Time Series",xlab="Overs Played",ylab="D/L Error") plot(ind[1:300],type="l",main="Zoomed D/L Error Time Series",xlab="Overs Played",ylab="D/L Error") x11() par(mfrow=c(4,1)) plot(di(ind,lag=50)[1:1000],type="l",main="Season Dierenced D/L Error Time Series",xlab="Overs Played",ylab="D/L Error") acf(di(ind,lag=50),lag.max=100,main="Season Dierenced D/L Error ACF") pacf(di(ind,lag=50),lag.max=100,main="Season Dierenced D/L Error PACF") spectrum(di(ind,lag=50),main="Season Dierenced D/L Error Spectrum") ## Test of Fit ## tIND0=arima(ind,order=c(0,0,0),seasonal=list(order=c(0,1,0),period=50)) tIND1=arima(ind,order=c(0,0,1),seasonal=list(order=c(0,1,0),period=50)) tIND2=arima(ind,order=c(0,0,2),seasonal=list(order=c(0,1,0),period=50)) tIND3=arima(ind,order=c(0,0,3),seasonal=list(order=c(0,1,0),period=50)) tIND4=arima(ind,order=c(0,0,4),seasonal=list(order=c(0,1,0),period=50)) tIND5=arima(ind,order=c(0,0,5),seasonal=list(order=c(0,1,0),period=50)) tIND0$aic tIND1$aic tIND2$aic tIND3$aic tIND4$aic tIND5$aic tINDA0=arima(ind,order=c(0,0,0),seasonal=list(order=c(0,1,0),period=50)) tINDA1=arima(ind,order=c(1,0,0),seasonal=list(order=c(0,1,0),period=50)) tINDA2=arima(ind,order=c(2,0,0),seasonal=list(order=c(0,1,0),period=50)) tINDA3=arima(ind,order=c(3,0,0),seasonal=list(order=c(0,1,0),period=50)) tINDA4=arima(ind,order=c(4,0,0),seasonal=list(order=c(0,1,0),period=50)) tINDA5=arima(ind,order=c(5,0,0),seasonal=list(order=c(0,1,0),period=50)) tINDA0$aic tINDA1$aic tINDA2$aic tINDA3$aic tINDA4$aic tINDA5$aic tINDA10=arima(ind,order=c(1,0,0),seasonal=list(order=c(0,1,0),period=50)) tINDA11=arima(ind,order=c(1,0,1),seasonal=list(order=c(0,1,0),period=50)) tINDA12=arima(ind,order=c(1,0,2),seasonal=list(order=c(0,1,0),period=50)) tINDA13=arima(ind,order=c(1,0,3),seasonal=list(order=c(0,1,0),period=50)) tINDA14=arima(ind,order=c(1,0,4),seasonal=list(order=c(0,1,0),period=50)) tINDA15=arima(ind,order=c(1,0,5),seasonal=list(order=c(0,1,0),period=50)) tINDA10$aic tINDA11$aic tINDA12$aic tINDA13$aic tINDA14$aic tINDA15$aic tINDA21=arima(ind,order=c(2,0,1),seasonal=list(order=c(0,1,0),period=50)) tINDA22=arima(ind,order=c(2,0,2),seasonal=list(order=c(0,1,0),period=50)) tINDA23=arima(ind,order=c(2,0,3),seasonal=list(order=c(0,1,0),period=50)) tINDA24=arima(ind,order=c(2,0,4),seasonal=list(order=c(0,1,0),period=50)) tINDA25=arima(ind,order=c(2,0,5),seasonal=list(order=c(0,1,0),period=50))

23

tINDA21$aic tINDA22$aic tINDA23$aic tINDA24$aic tINDA25$aic tINDA31=arima(ind,order=c(3,0,1),seasonal=list(order=c(0,1,0),period=50)) tINDA32=arima(ind,order=c(3,0,2),seasonal=list(order=c(0,1,0),period=50)) tINDA33=arima(ind,order=c(3,0,3),seasonal=list(order=c(0,1,0),period=50)) tINDA31$aic tINDA32$aic tINDA33$aic tINDA41=arima(ind,order=c(4,0,1),seasonal=list(order=c(0,1,0),period=50)) tINDA42=arima(ind,order=c(4,0,2),seasonal=list(order=c(0,1,0),period=50)) tINDA41$aic tINDA42$aic tINDA51=arima(ind,order=c(5,0,1),seasonal=list(order=c(0,1,0),period=50)) tINDA51$aic ## Model Diagnostics ## x11() tsdiag(tINDA13,gof.lag=50) x11() par(mfrow=c(2,1)) qqnorm(tINDA13$residuals,main="D/L Error Model Residual QQ") abline(0,1) cpgram(tINDA13$residuals,main="D/L Error Model Residual CPGRAM") x11() qqplot(rt(length(tINDA13$residuals),df=3), tINDA13$residuals, main="t(3) Q-Q Plot", + ylab="Sample Quantiles") abline(0,1) ##Smoothing of Errors## smoothDL<-matrix(nrow=50,ncol=10) for (i in 1:10){ smoothDL[,i]<-smooth.spline(newDL[,i])$y } ##HeatMap Codes## matEN<-data.matrix(heatEN) heatmapEN <- heatmap(matEN[50:1,], Rowv=NA, Colv=NA, col = cm.colors(256), scale="column", margins=c(5,10),ylab="Overs",xlab="Wickets") heatOR <-read.csv("/Users/skwong/Desktop/Timeseries/dlEN.csv",header=T) matOR<-data.matrix(heatOR) heatmapOR <- heatmap(matOR[50:1,], Rowv=NA, Colv=NA, col = cm.colors(256), scale="column", margins=c(5,10),ylab="Overs",xlab="Wickets")

24