Bayesian modelling of football outcomes. 1 Introduction. Synopsis. (using Skellam s Distribution)

Transcription

1 Karlis & Ntzoufras: Bayesian modelling of football outcomes 3 Bayesian modelling of football outcomes (using Skellam s Distribution) Dimitris Karlis & Ioannis Ntzoufras s: {karlis, ntzoufras}@aueb.gr Department of Statistics Athens University of Economics and Business Athens, Greece; 1 Introduction Poisson distribution has been widely used as a simple modelling approach for predicting the number of goals in football (see, for example, Lee, 1997, Karlis and Ntzoufras, 2000). Empirical evidence has shown a (relatively low) correlation between the goals in a football game. This correlation must be incorporated in the model. It is reasonable to assume that the two outcome variables in football are correlated since the two teams interact during the game. 2007, 24-26th June, Manchester Karlis & Ntzoufras: Bayesian modelling of football outcomes 2 Karlis & Ntzoufras: Bayesian modelling of football outcomes 4 Synopsis 1. Introduction. 2. Skellam s Distribution. 3. Bayesian infernece for the Skellam s distribution 4. The General Model. 5. Illustrative example and results. 6. Discussion and further work. We can assume bivariate Poisson distribution (see Karlis and Ntzoufras, 2003). The marginal distributions are simple Poisson distributions, while the random variables are now dependent. Important issue model extra variation of draws (mainly 0-0, 1-1, 2-2). This problem has been effectively handled by Dixon and Coles (1997) and Karlis and Ntzoufras (2003)

2 Karlis & Ntzoufras: Bayesian modelling of football outcomes 5 Karlis & Ntzoufras: Bayesian modelling of football outcomes 7 The Bivariate Poisson Model (Karlis and Ntzoufras, 2003) (X i,y i ) BP(λ 1i,λ 2i,λ 3i ) where BP(λ 1,λ 2,λ 3 ) is the bivariate Poisson distribution with parameters λ i, i =1, 2, 3 and probability function P (X = x, Y = y) =e (λ1+λ2+λ3) λx 1 λ y min(x,y) 2 x y ( ) k λ3 k!. (1) x! y! k k λ 1 λ 2 k=0 Marginals: X P oisson(λ 1 + λ 3 )andy P oisson(λ 2 + λ 3 ) Means and Variance: E(X) =V (X) =λ 1 + λ 3, E(Y )=V (X) =λ 2 + λ 3 Advantages Accounts for the covariance which can be modelled Has he latent variables decomposition which can be used for data augmentation Poisson Marginals Easy interpretation Easy extension to similar models. Covariance: Cov(X, Y )=λ 3 > 0 Can be derived using latent variables W i P oisson(λ i ), for i =1, 2, 3with X = W 1 + W 3 and with Y = W 2 + W 3. Karlis & Ntzoufras: Bayesian modelling of football outcomes 6 Karlis & Ntzoufras: Bayesian modelling of football outcomes 8 Covariates can be linked directly on the means of the latent variables as in Karlis and Ntzoufras (2003) or directly on the marginal means. A Simple Model (Karlis and Ntzoufras, 2003) Response variables (X, Y ) are the home and away goals in each game. Consider the structure of Lee (1997) for λ 1 and λ 2 log(λ 1i ) = µ + H + A HTi + D ATi (2) log(λ 2i ) = µ + A ATi + D HTi (3) µ: constant; H: home effect; A k, D k attacking and defensive parameters of team k; HT i, AT i home and away team in i game. Constant covariance λ 3 Disadvantages How to model λ 3? Is constant λ 3 sufficient? Does λ 3 varies across teams or time? Underestimates the number of draws in certain cases (while it provides better results on this from the double Poisson model) Poisson Marginals while empirical evidence has shown slight over-dispersion: need to adjust The model allows only for positive correlation, if the data show negative correlation (even slow) the model cannot handle it

3 Karlis & Ntzoufras: Bayesian modelling of football outcomes 9 Karlis & Ntzoufras: Bayesian modelling of football outcomes 11 marginal means λ 3 win draw loss ratio Table 1: The effect on the predicted probabilities when ignoring the covariance; the BP increases the probability of a draw. Skellam s distribution (1) For any pair of variables (X, Y )thatcanbewrittenas X = W 1 + W 3, Y = W 2 + W 3 with W 1 P oisson(λ 1 ), W 2 P oisson(λ 2 ) and W 3 any distribution with parameters θ 3 then Z = X Y PD(λ 1,λ 2 ) (Poisson difference or Skellam s distribution with parameters λ 1 and λ 2 ). Karlis & Ntzoufras: Bayesian modelling of football outcomes 10 Karlis & Ntzoufras: Bayesian modelling of football outcomes 12 Skellam s distribution (2) Diagonal Inflated Bivariate Poisson Model Under this approach a diagonal inflated model is specified by (1 p)bp(x, y λ 1,λ 2,λ 3 ), x y P D (x, y) = (1 p)bp(x, y λ 1,λ 2,λ 3 )+pd(x, θ), x = y, where D(x, θ) is discrete distribution with parameter vector θ. Such models can be fitted using the EM algorithm. Important: diagonal inflation improves in several aspects: better draw prediction, overdipsersed marginals, introduce correlation (4) Z = X Y PD(λ 1,λ 2 ) Mean E(Z) = λ 1 λ 2 Poisson difference or Skellam s distribution with Variance Var(Z) = λ 1 + λ 2 and density function ( ) z/2 f PD (z λ, λ 2 )=P(Z = z λ 1,λ 2 )=e (λ1+λ2) λ1 I z (2 ) λ 1 λ 2. (5) λ 2 for all z Z,whereI r (x) is the modified Bessel function of order r ( ) k ( x ) r x 2 4 I r (x) = 2 k!γ(r + k +1). k=0 (see Abramowitz and Stegun, 1974, pp. 375).

4 Karlis & Ntzoufras: Bayesian modelling of football outcomes 13 Karlis & Ntzoufras: Bayesian modelling of football outcomes 15 Prob lambda1 = 1,lambda2= x Prob lambda1 = 1,lambda2= x (see, Karlis and Ntzoufras,2006, SIM) Bayesian approach Used Discrete differences of dental data to test for differences of means between two repeated measurements test for zero inflated components Prob lambda1 = 1,lambda2= Prob lambda1 = 2,lambda2= using the Bayes Factor approach (and RJMCMC to estimate it). The same approach can be used to quantify (using the Bayes factor) Used Discrete differences of dental data to the importance of home effect (by testing for the equality of expected home and away goals) and the excess of draws (by testing the importance of the zero inflated component). x x Karlis & Ntzoufras: Bayesian modelling of football outcomes 14 Advantages Removes additive covariance (do not need to model it) Has an Poisson latent variable interpretation Does not assumes Poisson marginals (i.e. one may assume an overdispersed marginal distribution) Easy interpretation Simpler Model set-up than the corresponding Biv. Poisson model Discards part of the information Disadvantages Cannot model covariance and final full score (only differences). Karlis & Ntzoufras: Bayesian modelling of football outcomes 16 Skellam s Distribution for Football Scores Response variable: Z = X Y the goal difference in each game. Same structure for parameters λ 1 and λ 2 as in Bivariate Poisson: log(λ 1i ) = µ + H + A HTi + D ATi (6) log(λ 2i ) = µ + A ATi + D HTi (7) µ: constant; H: home effect; A k, D k attacking and defensive parameters of team k; HT i, AT i home and away team in i game. Use the zero inflated variation of Skellam s distribution to model the excess of draws. Hence we define the zero inflated Poisson Difference (ZPD) distribution as f ZPD (0 p, λ 1,λ 2 )= p + (1 p)f PD (0 λ 1,λ 2 ) and f ZPD (z p, λ 1,λ 2 )= (1 p)f PD (z λ 1,λ 2 ), (8) for z Z\{0}; wherep (0, 1) and f PD (z λ 1,λ 2 ) is given by (5).

5 Karlis & Ntzoufras: Bayesian modelling of football outcomes 17 Karlis & Ntzoufras: Bayesian modelling of football outcomes 19 Posterior Distributions The key element for building an MCMC algorithm for the above models is to use augmented data Sample latent data w 1i and w 2i from f(w 1i,w 2i z i = w 1i w 2i,λ 1i,λ 2i ) λw1i w 1i! where I(A) =1ifA is true and zero otherwise. Sample [δ i z i,λ 1i,λ 2i ] Bernoulli( p i )from p p i = p +(1 p)f PD (z i λ 1i,λ 2i ) [δ i indicates the mixing (zero inflated or PD) component] λ w2i 2i w 2i! I(z i = w 1i w 2i ) Then we MCMC algorithm for model parameters is similar to the one used in Poisson regression models. We can additionally model p as in logistic regression models Fulham (38) Watford (29) Chelsea (64) Arsenal (63) Tottenham (57) Everton (52) Bolton (47) Reading (52) Portsmouth (45) Blackburn (52) Aston Villa (43) Middlesbrough (44) Newcastle (38) Man City (29) West Ham (35) Wigan (37) Sheff Utd (32) Charlton (34) Man Utd (83) Liverpool (57) % Posterior intervals for attacking coefficients (A i ) Karlis & Ntzoufras: Bayesian modelling of football outcomes 18 Karlis & Ntzoufras: Bayesian modelling of football outcomes 20 Example: Premiership season data Data from the premiership for the season are used to fit the PD and ZPD models. Results using PD Posterior summaries for µ and home effect Min. 1st Qu. Median Mean 3rd Qu. Max. s.d. mu home Man Utd (27) Chelsea (24) Liverpool (27) Arsenal (35) Tottenham (54) Everton (36) Bolton (52) Reading (47) Portsmouth (42) Blackburn (54) Aston Villa (41) Middlesbrough (49) Newcastle (47) Man City (44) West Ham (59) Fulham (60) Wigan (59) Sheff Utd (55) Charlton (60) Watford (59) % Posterior intervals for defensive coefficients [( 1) D i ]

6 Karlis & Ntzoufras: Bayesian modelling of football outcomes 21 Karlis & Ntzoufras: Bayesian modelling of football outcomes 23 Number of games Goal difference 95% Posterior intervals for predicted differences (in red=observed differences; - indicates the posterior median) Karlis & Ntzoufras: Bayesian modelling of football outcomes 22 Karlis & Ntzoufras: Bayesian modelling of football outcomes 24 Posterior Predictive Table Observed Final Table Pred.(Obs.) Post. Expectations Obs. Obs. values Rank Team Pts G.Dif. Rank Team Pts G.Dif. 1 (1) Man Utd ManUtd (2) Chelsea Chelsea (3) Arsenal Liverpool (3) Liverpool Arsenal (6) Everton Tottenham (8) Reading Everton (5) Tottenham Bolton (9) Portsmouth Reading (10) Blackburn Portsmouth (11) Aston Villa Blackburn (12) Middlesbrough Aston Villa (7) Bolton Middlesbrough (13) Newcastle Newcastle (14) Man City Man City (15) West Ham West Ham (17) Wigan Fulham (18) Sheff Utd Wigan (19) Charlton Sheff Utd (16) Fulham Charlton (20) Watford Watford Deviations between observed and predictive measures Comparison Deviation 1. Relative Frequencies (counts/games) 1.06% 2. Frequencies (counts) Relative Frequencies of win/draw/lose 2.20% 4. Frequencies of win/draw/lose Expected points Expected goal difference 0.28 Deviation = 1 K ( E(Q Pred i y) Q obs i K where K is the lengths of vector Q (K = 13 for comparisons 1-4, K =20for comparisons 5-6). i=1 ) 2

7 Karlis & Ntzoufras: Bayesian modelling of football outcomes 25 Results using ZPD Posterior summaries for µ and home effect Min. 1st Qu. Median Mean 3rd Qu. Max. s.d. mu home p The posterior distribution of p is close to zero Small deviations between µ for PD and ZPD Home effect is similar in both models Results from PD Min. 1st Qu. Median Mean 3rd Qu. Max. s.d. mu home Karlis & Ntzoufras: Bayesian modelling of football outcomes 27 Deviations between observed and predictive measures Deviation Comparison PD ZPD 1. Relative Frequencies (counts/games) 1.06% 1.32% 2. Frequencies (counts) Relative Frequencies of win/draw/lose 2.20% 2.80% 4. Frequencies of win/draw/lose % 5. Expected points Expected goal difference Comments Zero inflation component does not improves the performance (as expected) Prediction of goal differences for ZPD are worse than the corresponding ones for PD Nevertheless, differences in the final rankings are minimal Karlis & Ntzoufras: Bayesian modelling of football outcomes 26 Karlis & Ntzoufras: Bayesian modelling of football outcomes 28 Number of games Goal difference 95% Posterior intervals for predicted differences (in red=observed differences; - indicates the posterior median) Comment: No major differences between the two models are observed. ZPD PD Discussion and further research Although in the simple and bivariate Poisson model we underestimate draws here using PD we have overestimated draws. No zero inflation is needed. We might need to add a component which will reduce the predicted draws. Can covariates on p improve the model? Apply Bayesian variable selection and Bayesian model averaging techniques. Apply other distributions defined on the same range.

8 Karlis & Ntzoufras: Bayesian modelling of football outcomes 29 Our Related Publications 1. Karlis D. and Ntzoufras, I. (2006). Bayesian analysis of the differences of count data. Statistics in Medicine, 25, Karlis, D. and Ntzoufras, I. (2005). Bivariate Poisson and Diagonal Inflated Bivariate Poisson Regression Models in R. Journal of Statistical Software, Volume 10, Issue Karlis, D. and Ntzoufras, I. (2003). Analysis of Sports Data Using Bivariate Poisson Models. Journal of the Royal Statistical Society, D,(Statistician),52, Karlis, D. and Ntzoufras, I. (2000). On Modelling Soccer Data. Student, 3, Karlis & Ntzoufras: Bayesian modelling of football outcomes 30 Additional References Lee, A.J. (1997). Modeling scores in the Premier League: Is Manchester United really the best? Chance, 10, Dixon, M.J. and Coles, S.G. (1997). Modelling association football scores and inefficiencies in football betting market. Applied Statistics, 46,