SEARCH

SEARCH BY CITATION

Keywords:

  • Bradley;
  • Terry model;
  • Cumulative logit model;
  • Exponentially weighted moving average process;
  • Paired comparisons;
  • Sports tournaments

Abstract

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Description of the data and analyses with non-dynamic abilities
  5. 3. Dynamic Bradley–Terry model
  6. 4. Applications
  7. 5. Conclusions
  8. Acknowledgements
  9. References

Summary.  In the course of national sports tournaments, usually lasting several months, it is expected that the abilities of teams taking part in the tournament will change over time. A dynamic extension of the Bradley–Terry model for paired comparison data is introduced to model the outcomes of sporting contests, allowing for time varying abilities. It is assumed that teams’ home and away abilities depend on past results through exponentially weighted moving average processes. The model proposed is applied to sports data with and without tied contests, namely the 2009–2010 regular season of the National Basketball Association tournament and the 2008–2009 Italian Serie A football season.


1. Introduction

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Description of the data and analyses with non-dynamic abilities
  5. 3. Dynamic Bradley–Terry model
  6. 4. Applications
  7. 5. Conclusions
  8. Acknowledgements
  9. References

The analysis of sports data has always aroused great interest among statisticians. Albert et al. (2005) collected a number of articles that summarize various statistical aspects of interest in sports data including rating of players or teams, evaluation of sport strategies, enhancement of sport rules, illustration of statistical methods and forecasting of results.

Sports data have been investigated from different perspectives, often with the aim of forecasting the results. A first approach consists in modelling the scores of the two opposing teams. Maher (1982) employed independent Poisson distributions for the score of each team with means that depend on the attack and defence strength of teams. Dixon and Coles (1997) proposed an ad hoc adjustment of the Poisson distribution introducing a dependence parameter that modifies the probabilities of the results 0–0, 0–1, 1–0 and 1–1. Dixon and Coles (1997) introduced also a dynamic element in the model updating the parameter estimates including the results up to the last observation and downweighting observations that are distant in time. Karlis and Ntzoufras (2003) suggested applying a bivariate Poisson distribution with a dependence parameter between the number of goals scored by the two teams and then extended the model to inflate the probabilities of draws.

McHale and Scarf (2007) modelled the number of shots of the two teams. They proposed two different types of Archimedean copula with either Poisson or negative binomial distributions for the marginals to account for the negative dependence between shots for and shots against.

Extensions allowing dynamic developments of abilities of the teams were proposed by Rue and Salvesen (2000) and Crowder et al. (2002). Rue and Salvesen (2000) assumed that the attack and defence strength parameters of each team follow a Brownian motion process. The model is estimated by employing Bayesian inference through Markov chain Monte Carlo methods. Crowder et al. (2002) suggested an auto-regressive model for the attack and defence abilities of teams. The original model is then replaced by a derived version that is easier to handle by maximum likelihood.

A second approach to the analysis of sports data consists of modelling the difference in scores. Clarke and Norman (1995) performed a linear regression of the difference in scores on the difference in strength of the two teams. Harville (2003) employed a similar specification but eliminated the incentives for running up the score beyond a predetermined number of points. A dynamic specification of strength in this context was considered in Harville (1980) who proposed an auto-regressive process for the strength of teams in different seasons. Also Glickman and Stern (1998) assumed that the evolution of week-by-week and seasonal strength follows a first-order auto-regressive process; inference is carried out in a Bayesian framework through Markov chain Monte Carlo algorithms.

Finally, sports data can be analysed by considering only the outcomes of the matches (win–draw–loss). Goddard and Asimakopoulos (2004) used an ordered probit model to determine which covariates, e.g. importance of the match, fouls, yellow and red cards, affect the result of the match. An ordered probit model was adopted also by Koning (2000) who specified the probability of the outcome as a function of the difference of abilities of the two teams. Kuk (1995) introduced two strength parameters for each team: one denoting the strength when playing at home and the other when playing away.

Barry and Hartigan (1993) proposed a dynamic extension for the ability parameters of teams; they employed a choice model assuming a prior distribution for strength of teams that changes slowly in time. Fahrmeir and Tutz (1994) considered three possible specifications for the development of abilities: a first- and second-order random walk and a local linear trend model. These models are estimated by using empirical Bayes methods. Glickman (1999) specified a logit model assuming a prior with normal increments for abilities of teams and proposed an approximate Bayesian algorithm for ranking purposes. Knorr-Held (2000) employed a logit model assuming random-walk priors for abilities of teams. The variance of the random walk is estimated through four different predictive criteria whereas the abilities are estimated by means of the extended Kalman filter and smoother.

In this paper, we analyse the results of sport tournaments from the last perspective, i.e. modelling the outcomes of matches. Since we are interested in studying how the strengths of the teams evolve during the season, we develop a dynamic paired comparison model. In particular, we model the evolution in time of the abilities in home and away matches of each team through an exponentially weighted moving average process.

The paper is organized as follows. Section 2 presents two motivating data sets regarding the American National Basketball Association (NBA) league and the Italian major men's football league. Section 3 describes the proposed dynamic version of the Bradley–Terry model, discusses maximum likelihood estimation and considers model validation by Brier and ranking probability scores. The methodology is applied in Section 4 to the data for the two sports. Concluding remarks and future research are summarized in Section 5.

The data and R (R Development Core Team, 2011) code written for implementing the analyses are available from

2. Description of the data and analyses with non-dynamic abilities

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Description of the data and analyses with non-dynamic abilities
  5. 3. Dynamic Bradley–Terry model
  6. 4. Applications
  7. 5. Conclusions
  8. Acknowledgements
  9. References

2.1. National Basketball Association

As our first motivating example we consider the 2009–2010 regular season of the NBA league. There are 30 teams in the league playing 82 games each: 41 at home and 41 away. The total number of matches is 1230. The schedule of the tournament includes a greater number of contests against teams in the same division and in the same conference, whereas competitions between teams in different conferences are less numerous. The regular season started at the end of October 2009 and ended in mid-April 2010. Matches were played on 164 different days. The number of matches per day ranges from 1 to 14; the mean number is 7.5.

The description of the proposed methodology for the analysis of tournaments is simplified by the assumption of an order for the m=1230 matches among the n=30 teams that are involved in the tournament. A convenient choice is to arrange the matches in chronological order, with those played at the same time in alphabetic order of the home team. Let Yi be the binary random variable which denotes the result of the ith match, i=1,…,m, played by the home team hi against the visiting team vi, with inline image, inline image. We arbitrarily code Yi=1 if the home team wins and Yi=0 if the visitors win.

Traditional paired comparison models describe the outcome probability as inline image where F is a cumulative distribution function and inline image and inline image are the parameters representing the abilities of the home and the visiting teams in match i. This simple choice model is commonly termed the Bradley–Terry model (Bradley and Terry, 1952) or the Thurstone–Mosteller model (Thurstone, 1927; Mosteller, 1951) depending on whether F is the cumulative distribution function of a logistic or of a standard normal random variable respectively. In the rest of the paper, we shall consider the Bradley–Terry specification.

The advantage deriving from playing at home is commonly taken into account by including a common home effect parameter η for all teams (Fahrmeir and Tutz, 1994; Knorr-Held, 2000; Harville, 2003), thus leading to the model

  • image(1)

Parameter identifiability requires one constraint in the set of abilities, such as the sum constraint inline image or the reference team constraint ak=0 for some k ∈ {1,…,n}.

Table 1 shows the estimates of the abilities inline image from model (1) with the sum constraint on team abilities. Teams are ranked on the basis of the number of matches won during the season out of the total 82 matches played (the second column). The third column indicates which percentage of the matches won was played at home. On average circa 60% of the matches are won by the home team; therefore the advantage in playing at home seems not negligible. The fourth and sixth columns in Table 1 report the estimated abilities and the ranking according to the estimated abilities. There is a very close agreement between the ranking that is obtained by the estimated abilities and the number of matches won; indeed the Kendall rank correlation τ is 0.97. The estimated home effect (with standard error in parentheses) is inline image. The fifth column of Table 1 reports the quasi-standard errors (Firth and de Menezes, 2004) of the abilities. The quasi-standard errors allow us to reconstruct approximately the uncertainty of pairwise differences inline image used for comparing teams k and k without the need to report also the covariance between inline image and inline image. For example, if it is of interest to test whether the ability of Cleveland is significantly higher than the ability of Orlando, the standard error of the difference between the estimators can be approximated by using the quasi-variances simply as inline image. In this case, the abilities of the best two teams do not appear statistically different.

Table 1.   2009–2010 American NBA league†
Team Won % home Results for static model Results for dynamic model
Ability qse Rank Ability Rank
  1. †The table displays the number of matches won, the percentage of matches won played at home, % home, estimated abilities, quasi-standard errors qse and ranks based on the static Bradley–Terry model, and estimated mean abilities and ranks based on the dynamic Bradley–Terry model.

Cleveland Cavaliers610.571.1890.26710.7691
Orlando Magic590.581.0480.26020.5683
Los Angeles Lakers570.600.9970.25430.6502
Dallas Mavericks550.510.8350.24940.4515
Phoenix Suns540.590.7670.24850.3828
Atlanta Hawks530.640.6920.24780.4247
Denver Nuggets530.640.7440.24670.4824
Utah Jazz530.600.7490.24860.3489
Boston Celtics500.480.4930.243120.4476
Oklahoma City Thunder500.540.5650.242100.26811
Portland Trail Blazers500.520.5380.242110.25312
San Antonio Spurs500.580.5800.24290.27010
Miami Heat470.510.3070.239130.07014
Milwaukee Bucks460.610.2410.239140.05415
Charlotte Bobcats440.700.1550.239150.01916
Houston Rockets420.550.1400.237160.08013
Chicago Bulls410.590.0000.23717−0.06520
Memphis Grizzlies400.58−0.0110.237180.01917
Toronto Raptors400.63−0.0740.240190.01718
New Orleans Hornets370.65−0.1770.24020−0.03719
Indiana Pacers320.72−0.5440.24621−0.38222
Los Angeles Clippers290.72−0.6510.24922−0.30021
New York Knicks290.62−0.7490.25023−0.41523
Detroit Pistons270.63−0.8210.25225−0.44625
Philadelphia 76ers270.44−0.8300.25326−0.45726
Golden State Warriors260.69−0.8040.25424−0.58628
Washington Wizards260.58−0.9040.25728−0.49427
Sacramento Kings250.72−0.8880.25827−0.43624
Minnesota Timberwolves150.67−1.6160.30229−0.83429
New Jersey Nets120.67−1.9720.32730−1.11930

2.2. Association football

The second application concerns the 2008–2009 Italian Serie A football league. This tournament comprises n=20 teams with matches played between August 2008 and May 2009. The tournament has a double-round-robin structure, so each team competes twice against all the other teams in the league: once at home and once away. The total number of matches is thus m=n(n−1)=380. These matches were played on 74 different days. As in the NBA example, there are days with just one match and days with up to 10 matches played. The average number of matches per day is 5.14.

The teams, ranked according to the final points order, are listed in Table 2. In the football tournament, the winning team gains 3 points whereas the losing team receives nothing. If the match is drawn, both teams gain 1 point. On average, 65% of the total points are gained in home matches, with percentages ranging from 45% to 79%, and it is then evident that home advantage of teams should be included in the model.

Table 2.   2008–2009 Italian Serie A football league†
Team pts % home Results for static model Results for dynamic model
Ability qse Rank Ability Rank
  1. †The table displays final points pts, the percentage of points won at home, % home, estimated abilities ability, quasi-standard errors qse and ranks based on the static Bradley–Terry model, and estimated mean abilities and ranks based on the dynamic Bradley–Terry model.

Internazionale840.561.3800.34810.4621
Juventus740.530.9280.32420.3032
Milan740.610.9130.32630.2943
Fiorentina680.650.6430.32750.2054
Genoa680.600.6930.31440.1775
Roma630.680.4220.31060.1086
Udinese580.660.2310.3067−0.00711
Palermo570.750.1240.31980.0717
Cagliari530.70−0.0100.31590.00410
Lazio500.56−0.2330.329120.0288
Atalanta470.70−0.2240.31711−0.00912
Napoli460.76−0.2590.307130.0099
Sampdoria460.70−0.1770.29510−0.09814
Siena440.73−0.4250.31215−0.10915
Catania430.79−0.4090.31914−0.07013
Chievo380.45−0.5270.30616−0.24316
Bologna370.57−0.6270.31517−0.25217
Torino340.74−0.7600.31618−0.25918
Reggina310.58−0.8500.31120−0.33720
Lecce300.63−0.8330.30319−0.27619

In contrast with basketball, football matches can also end in a draw; hence the random variable Yi has three categories arbitrarily coded as 2 if the home team wins, 1 in the case of a draw and 0 in the case of victory of the visiting team. Accordingly, model (1) is extended to account for draws with a cumulative link specification (Agresti, 2002):

  • image(2)

where inline image are cut point parameters. Parameter identifiability is achieved by imposing the ‘symmetrical’ constraints δ0=−δ and δ1=δ, with δgeqslant R: gt-or-equal, slanted0. These constraints are needed to ensure that two teams with the same ability playing on a neutral field (no home advantage) have the same probability of winning the match. If Yi assumes only two possible values, then the cumulative logit model (2) reduces to the standard Bradley–Terry model (1), and δ becomes 0.

The fourth column of Table 2 shows the estimates of the abilities inline image again with the sum constraint on team abilities. The estimates of the abilities range from −0.850 for Reggina to 1.380 for Internazionale. The ranking that is derived from the estimated abilities is very similar to the final points ranking, as the Kendall τ rank correlation is 0.94. The estimated home effect parameter is inline image and the estimated cut point parameter is inline image. The fifth column in Table 2 reports quasi-standard errors for teams’ abilities. In this case, if, for example, it is of interest to test whether the ability of Internazionale is significantly higher than the ability of Juventus, the standard error of the difference can be computed as inline image, so there is no evidence in this model that they are statistically different even though Internazionale ended the tournament 10 points ahead of Juventus.

In the above static models, parameters ak measure the overall abilities of the teams over a complete season. However, team abilities are expected to change during the season because of injuries to players, tiredness due to participation also in international competitions, team psychology and other factors. In the next section we develop a dynamic version of the Bradley–Terry model in which abilities are allowed to change and to depend on the past performance of the team.

3. Dynamic Bradley–Terry model

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Description of the data and analyses with non-dynamic abilities
  5. 3. Dynamic Bradley–Terry model
  6. 4. Applications
  7. 5. Conclusions
  8. Acknowledgements
  9. References

3.1. The model

We model the match results with the following dynamic Bradley–Terry model:

  • image(3)

where inline image describes the ability of the home team hi in match i played against the visiting team vi at time ti. We specify an evolution of team ability in home matches which depends only on past matches played at home, whereas the ability when playing away depends only on past matches played as visitors. First, consider the ability in home matches and let inline image be the time of the match previous to match i in which hi was the home team. The ability of the home team inline image is assumed to evolve in time following the exponentially weighted moving average (EWMA) process

  • image(4)

for some home-specific smoothing parameter λ1 ∈ [0,1]. The term inline image denotes the mean home ability of team hi based only on the result of the nearest previous match played at home by hi

  • image(5)

with β1 being a home-specific parameter and inline image a variable measuring the result of team hi in the match played at time inline image In the NBA application, we specify inline image as the binary variable equal to 1 if team hi won its previous home match and 0 if it was defeated. Thus, if j denotes the nearest match previous to match i, which was played by hi at home, i.e. inline image, then inline image. Instead, in the Serie A application, we specify inline image as the number of points earned by team hi at time inline image of its previous home match: 3 points for a victory, 1 for a draw and 0 for a loss.

The ability model (4) must be complemented by an initial condition. We assume that all teams start with the same home ability equal to inline image, where inline image is an average of variables inline image over the previous season. In the analysis of the NBA 2009–2010 season, inline image is 0.608, the frequency of victories at home during the NBA 2008–2009 regular season. In the analysis of the Serie A 2008–2009 data, inline image is 1.676 points, which is the average number of points gained by home teams during the Serie A 2007–2008 season.

Suppose that the home team hi has played K matches at home before the match played at time ti. Then, by iterated back-substitution, the model based on the pair of equations (4)–(5) can be reformulated as

  • image(6)

with inline image denoting the time of the rth previous match played at home by team hi. Thus, the ability inline image is a function of the entire past history of home matches, inline image, inline image. The derived covariate inline image is a weighted mean of these past results with weights λ1, inline image, …, inline image geometrically decreasing to 0. The smoothing parameter λ1 specifies the persistence of the dependence on previous home matches. In the limiting case λ1=1, the home team's ability depends only on the previous home match, inline image In contrast, if λ1=0 the home ability is constant in time and equal for all teams, inline image. Values of λ1 ∈ (0,1) specify different levels of smoothing. In particular, home abilities smoothed in time are obtained when λ1 approaches 0.

Similarly, the ability of the visiting team is modelled by a second EWMA process

  • image

where λ2 ∈ [0,1] is the visitor-specific smoothing parameter and inline image for a visitor-specific coefficient β2. The starting values for r are computed similarly to those for the home abilities. In the NBA 2009–2010 data inline image is set equal to 0.392, which is the frequency of visitors’ victories during season 2008–2009, whereas in the Serie A 2008–2009 data inline image points, which is the average number of points gained by visitors in the Serie A 2007–2008 tournament.

Thus, in the proposed dynamic Bradley–Terry model, the EWMA specification is used to account for the serial association between match results of the same team, with suitable differences between home and away matches.

3.2. Likelihood inference

EWMA processes are routinely used in time series forecasting (Holt, 2004) and in statistical quality control charts (Montgomery, 2005). In these contexts, the smoothing parameter is often chosen by trials or by using ad hoc methods based on previous experience. However, many have argued that automatic, data-driven choices of the smoothing parameter would be preferable. For example, in classical time series, the smoothing parameter ‘is often estimated by minimizing the sum of squared one-step-ahead forecast errors’ (Chatfield (2000), page 98).

Here, we follow the recommendation to identify the smoothing parameters by using available observations and consider maximum profile likelihood estimation of the two smoothing parameters λ1 and λ2.

Let inline image be the vector of parameters of interest and let inline image be the vector of nuisance smoothing parameters. Under the chosen order for the match results, the likelihood function for inline image is written as

  • image

Given the home smoothing parameter λ1, the home ability can be written as inline image where inline image is the weighted average of past home results inline image with weights inline image; see formula (6). In parallel, given the visitors’ smoothing parameter λ2, the visitors’ ability is inline image where inline image has a specification that is analogous to the home case. Accordingly, the conditional probability for the result of match i is expressed by the cumulative logit model

  • image

Thus, if the smoothing parameters are known, the likelihood function for the interest parameter γ corresponds to that of a standard logistic regression model if draws are not allowed (e.g. basketball) or that of a cumulative logistic regression model with constrained cut points in the case of draws (e.g. football). The simplicity of computation of inline image, the estimates of γ given λ, suggests a two-step maximization of the likelihood. First, the smoothing parameter vector λ is estimated by maximizing the profile likelihood inline image, then γ is estimated as inline image. This approach is employed in the applications that are illustrated in Section 4.

3.3. Model validation

Model validation can be based on comparison of the fitted probabilites from the proposed model with fitted probabilites from the unstructured model (2). The model proposed aims to capture the evolution in time of all teams’ abilities with only four parameters (five in the case of draws), whereas the unstructured model has n free parameters (n+1 when draws are allowed), with n being the number of teams. Clearly, the unstructured model is expected to fit the data better, and thus it may be viewed as a benchmark. The closer the fitted probabilities of the model proposed are to those of the unstructured model, the better is the fit. To summarize the fitted probabilities we consider the Brier score (Brier, 1950) which is defined for match i as

  • image

where 1(yi=q) is the indicator function of the event {yi=q}, Q=2 for sports without draws and Q=3 when draws are allowed, and inline image is the maximum likelihood estimate of θ based on the results of all the matches played, i.e. inline image. When the fit is perfect, giving probability 1 to the observed outcome, the Brier score is equal to 0, whereas a completely erroneous fit produces a Brier score equal to 2.

Some researchers have suggested that in the case of more than two categories it is better to employ an index which accounts for the whole distribution of probabilities, such as the rank probability score (Czado et al., 2009)

  • image

In the analysis of sport tournaments the real interest usually lies in forecasting future results. Hence, it may be more relevant to evaluate the fitted model from a predictive point of view. In this case, we quantify the BSi and RPSi using the maximum likelihood estimate inline image computed only with matches played before the forecasted match i, i.e. only with results inline image.

4. Applications

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Description of the data and analyses with non-dynamic abilities
  5. 3. Dynamic Bradley–Terry model
  6. 4. Applications
  7. 5. Conclusions
  8. Acknowledgements
  9. References

4.1. Application to the National Basketball Association tournament

We fit the proposed model to the NBA 2009–2010 regular season. Fig. 1 shows the profile log-likelihood for the smoothing parameters λ1 and λ2. This is maximized at inline image and inline image. These two values are close to 0, thus supporting the effect also of remote match results on the estimation of the present ability. The limiting model with one common home advantage parameter and no evolution in time of abilities corresponds to inline image. This pair of values for the smoothing parameters is strongly not supported by the data. In fact, the maximized profile log-likelihood is −752.22 whereas the profile log-likelihood for inline image is much smaller, at −830.56. The maximum likelihood estimates for the home and away coefficients computed at the estimated smoothing parameters are inline image and inline image, with estimated standard errors 0.465 and 0.699 respectively. Both estimates appear strongly significant.

image

Figure 1.  Contour plots of the profile log-likelihood for the smoothing parameters λ1 and λ2 for the NBA 2009–2010 season: (a) plot for all the range of values of λ1 and λ2; (b) zoom of the area of highest likelihood (•, maximum likelihood estimate)

Download figure to PowerPoint

To validate the fitted model, we compute the fitted Brier scores for each of the 1230 matches both with the model proposed, which includes only four parameters, and with the unstructured model involving 31 parameters. There is a high positive association between the Brier scores of the two models; in fact their correlation is 0.784. The mean of the Brier scores for the fitted model is 0.424, whereas for the unstructured model it is 0.379; thus the latter is 10.6% smaller in mean. This result was expected given the larger number of parameters of the unstructured model.

However, it is more interesting to consider the appropriateness of the model proposed by evaluation of its predictive performance. For this, we fit the model to the data coming from half of the competition days, i.e. 82 days, then predict the results of the matches taking place in the following day of competition, the 83rd, and compare them with the observed results by the Brier score. Then, the model is refitted including also the matches in day 83 and used to forecast results in day 84, and so on until the last day of matches (day 164). Following this scheme we compute predictions for a total of 638 matches. Predictions from the unstructured model are similarly computed. These predictions are also compared with those obtained by simply using the empirical proportions of wins and losses of the home team computed on the first half of the competition days. These fixed empirical proportions are employed to forecast all the results of the second half of the tournament. Fig. 2(a) shows the boxplots of the Brier scores computed for each forecasted match by using the unstructured model, the EWMA model and the empirical proportions of wins and losses. Visual inspection of the boxplots shows Brier scores of our model are very similar to those of the unstructured model whereas the Brier scores of the empirical proportions are noticeably worse. Specifically, the mean predictive Brier score for the proposed and the unstructured model are very close: 0.421 and 0.409 respectively. The correlation between the two sets of Brier scores is 0.879. The overall conclusion is that the model proposed is very competitive with the unstructured model from a predictive point of view. The mean Brier score of empirical proportions is 0.497, so it is 15% larger than the score of the model proposed. Furthermore, both the unstructured and the proposed models present Brier scores which are lower than the Brier score of empirical proportions in 66% of the forecasts.

image

Figure 2.  (a) Boxplots of the predictive Brier scores for the NBA data and (b) boxplots of rank probability scores for the football data computed for the unstructured model, the EWMA model and the forecasts based on empirical proportions

Download figure to PowerPoint

Fig. 3 shows the smoothed abilities for nine teams during the complete regular season. The smoothed abilities are computed with parameters estimated from the complete tournament. For each team there is a home and a visiting ability that are plotted in the same graph. The timescale is the sequence of matches; at the end of the regular season each team has played 41 matches at home and 41 away. Cleveland Cavaliers ended the season with the highest number of victories, 61, and, as expected, their ability in both home and away matches increases noticeably from the starting mean ability that is common to all teams. It seems that Cleveland benefits from an important home advantage effect; indeed it won 85% of the matches that it played at home and 63% of the matches played away. Also New Orleans Hornets show an important home effect especially in the first half of the season. This team won 16 matches at home out of the first 20 matches whereas it succeeded only in five matches away out of the first 20. However, in the second part of the season it showed no particular advantage in playing at home since it won eight matches out of the last 21 both at home and away. New Jersey Nets was the team that performed worst during the season. They won a total of 12 matches during the whole tournament and performed poorly both at home and away. The increase in the visiting ability after match 30 is due to their winning three away matches in a row.

image

Figure 3.  Smoothed home (inline image) and visiting (inline image) abilities for nine teams for the NBA 2009–2010 season: (a) Cleveland Cavaliers, (b) Los Angeles Lakers; (c) Phoenix Suns; (d) Oklahoma City Thunder; (e) New Orleans Hornets; (f) New York Knicks; (g) Washington Wizards; (h) Minnesota Timberwolves; (i) New Jersey Nets

Download figure to PowerPoint

Finally, the seventh column in Table 1 reports the ranking that is derived from the proposed model based on the average of the team abilities in each of the 82 days in which the team played. The reported abilities are computed so as to sum to 0, by analogy with the unstructured model. The Kendall rank correlation between the rankings of the unstructured and the proposed model is 0.89. In this case we cannot report quasi-standard errors since the abilities are not individual parameters as in the static Bradley–Terry model. Suppose that we are interested in testing whether the ability of Cleveland Cavaliers is significantly higher than the ability of Orlando Magic. The ability of the former team is 0.769 whereas the ability of the latter is 0.568, so their difference amounts to 0.201. This difference appears statistically significant since its standard error is 0.018. This result is different from the result that is given by the unstructured model because the model proposed has fewer parameters and thus the ability estimates are more precise.

4.2. Application to the Serie A tournament

The second data set concerns the 2008–2009 Serie A football tournament which allows also for ties. The profile log-likelihood is maximized at inline image and inline image: again a pair of values that are quite close to 0. The support for this model, in contrast with the static model resulting when inline image, is given by the maximized profile log-likelihood which is equal to −383.27, whereas it takes the value −393.68 in the model with smoothing parameters equal to 0. The maximum likelihood estimates for the coefficients computed at the estimated values of the smoothing parameters are inline image and inline image, with standard error 0.173 and 0.294 respectively, whereas the estimated cut point parameter is inline image with standard error 0.054.

Since in football data there are three possible results, it is more appropriate to employ the rank probability score to validate the model. The mean rank probability score for the whole tournament, which comprises m = 380 matches, is 0.416 for the proposed model and 0.369 for the unstructured model; hence the proposed model has a mean rank probability score which is 11.2% higher than that of the unstructured model. The correlation between rank probability scores for the two models is 0.715.

As for the NBA example, it is, however, more interesting to consider the predictive performance of the model. To forecast the results we employ the same scheme as used for the basketball data. The model is estimated by using the matches in the first half of the competition days (37 days) and the results of matches occurring in the 38th day of competitions are predicted. Then we updated the model with the matches performed in day 38 and predict the following matches. The procedure is repeated until the end of the tournament with a total of 198 matches predicted. The mean rank probability scores for the unstructured and the EWMA model are essentially equal: 0.450 and 0.451 respectively. The correlation of the RPSi in the two models is 0.762. Fig. 2(b) shows the boxplots of the rank probability scores of the two models and the scores computed by using the empirical proportions of home wins, draws and losses computed in the first 37 days. It is evident that the medians of the RPSi for the unstructured and the EWMA models are equal; however, the EWMA model presents a shorter box; thus with respect to the unstructured model it assigns higher probability to results actually observed in fewer cases, but also it less often assigns higher probabilities to results that do not occur. The mean RPSi of empirical proportions is 0.468, which is 4% larger than the value for the EWMA model. Furthermore, the rank probability scores of the empirical proportions are higher than the scores computed for the proposed model in 58% of the forecast matches.

Fig. 4 shows the estimated abilities for nine of the teams which competed in the 2008–2009 Serie A league. Internazionale won the tournament, ending the season with 84 points, whereas Juventus and Milan achieved second position, gaining 74 points each. The plots show that the home performance of Internazionale is more stable during the season; in particular they never lost a home match, whereas Juventus and Milan lost two home matches each. In general, teams tend to perform better at home than away. This is particularly evident for Siena which won nine matches and drew five matches out of the 19 home matches, whereas in the other 19 away matches they won only three times and drew three matches.

image

Figure 4.  Smoothed home (inline image) and visiting (inline image) abilities for nine teams for the 2008–2009 football season data: (a) Internazionale; (b) Juventus; (c) Milan; (d) Roma; (e) Atalanta; (f) Siena; (g) Chievo; (h) Reggina; (i) Lecce

Download figure to PowerPoint

The seventh column in Table 2 reports the mean abilities of football teams resulting from the EWMA model. The Kendall rank correlation between the rankings derived from the unstructured and the proposed models is 0.86. In this case, the standard error needed to test whether the teams Internazionale and Juventus have the same ability is 0.037. Since the estimated ability of Internazionale is 0.462 and the estimated ability of Juventus is 0.303, yielding a difference of 0.159, the two teams appear to have statistically different abilities whereas the unstructured model led to the opposite conclusion.

5. Conclusions

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Description of the data and analyses with non-dynamic abilities
  5. 3. Dynamic Bradley–Terry model
  6. 4. Applications
  7. 5. Conclusions
  8. Acknowledgements
  9. References

We have described a dynamic paired comparison model for the results of matches in sport tournaments. The model specification describes the temporal evolution of teams’ abilities by separate EWMA processes for the home and away results. The two applications to basketball and football tournaments show that the model proposed seems to capture the relevant aspects of the evolution in time of team abilities, thus providing sensible forecasts if compared with the unstructured model that fits one ability parameter for each team.

The model proposed uses only information about the final result of previous matches. It seems reasonable that using more detailed information about previous matches may result in more accurate data fitting and improved forecasts. The inclusion of additional information about previous matches is easily handled under the model framework that is described in this paper; for example one could substitute inline image and inline image with vectors inline image and inline image and thus consider vector home and visitor parameters β1 and β2. With a possible large number of covariates from previous matches, it could be sensible to consider some form of shrinkage to avoid overfitting, e.g. by a lasso penalty (Tibshirani, 1996) on parameters β1 and β2.

The model proposed requires starting values for the covariates. We considered equal values for all teams, inline image and inline image, based on the results in the previous season. It is possible to consider team-specific starting values, as for example the proportions of wins and losses at home for each team in the previous season. The use of team-specific starting values for the NBA data leads to somewhat different estimates of the model parameters, but the fitting and predictive qualities of the resulting model, as evaluated by the Brier score, are almost identical to those of the model with all equal starting values. This exercise seems not possible for the football data since the teams in the Serie A league change season by season: at the end of the regular season the last three teams in the league are demoted to the lower league and the three best teams of the lower league are promoted to Serie A. Hence, the teams in the league are not the same in different seasons and it seems inappropriate to use team-specific values computed for teams in different leagues.

The issue of whether team-specific home advantages should be included in paired comparison models has been considered by many researchers with contrasting conclusions. Knorr-Held (1997) did not find much evidence of home advantage heterogeneity among teams in the Bundesliga. Nor do the results in Harville and Smith (1994) show much difference in home field advantages among college basketball teams. Analyses for some other contexts do, however, support heterogeneity in home advantages; see Clarke and Norman (1995), Kuk (1995) and Glickman and Stern (1998) for different analyses of the English Premier Football League. The model that is proposed in this paper can be seen as a convenient way to induce home abilities which vary between teams and in time depending on past performances of teams.

We considered sports with matches that may end with two or three different results. There are other competitions where more than three levels of results are possible. An example is volleyball where points are assigned as follows: if a match ends 3–0 or 3–1 the winning team gains 3 points and the losing team remains empty handed, whereas, if the match ends 3–2, the winning team gains 2 points and the losing team is awarded 1 point. As suggested by a referee, the analysis of volleyball matches would require a categorical variable Yi with four ordered levels. There is no special difficulty in extending our modelling framework to this case. More generally, assume that Yi is a categorical variable that may assume Q different categories where Q−1 denotes the best result achievable for the home team and 0 denotes the best result for the visiting team. The cumulative logit model (2) is easily extended to handle Q levels

  • image(7)

where inline image are cut point parameters. To preserve model identifiability, the symmetrical constraint now becomes inline image, q=0,…,Q−1, and δQ/2−1=0 when Q is even.

Acknowledgements

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Description of the data and analyses with non-dynamic abilities
  5. 3. Dynamic Bradley–Terry model
  6. 4. Applications
  7. 5. Conclusions
  8. Acknowledgements
  9. References

The authors thank the Associate Editor and the referee for detailed comments about the first submission. The first two authors have been partially supported by the Progetti di Rilevante Interesse Nazionale 2008 grant of the Italian Ministry of Instruction, University and Research. The third author's work was supported by the UK Engineering and Physical Sciences Research Council.

References

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Description of the data and analyses with non-dynamic abilities
  5. 3. Dynamic Bradley–Terry model
  6. 4. Applications
  7. 5. Conclusions
  8. Acknowledgements
  9. References