Por Rui Martins (CEAUL and Department of Statistics and Operations Research, University of Lisbon).

Statistical modeling of sports results has become trendy. Different types of models have been proposed to model these data depending on the objectives: from predicting the outcome of a game or team rankings in national championships to the identification of player characteristics that can improve their performance.

Our work shows that the Multinomial model is a reasonable approach when it comes to predicting the result of a game in terms of win, tie, or loss. Whereas if one wants to predict the goals scored by each team in a match a modified–Poisson distribution might be the best way to go. The Bayesian context facilitates the predictions for a new game because they are naturally accommodated in terms of the posterior predictive distribution.

We do not account for possible information embedded in covariates and from this point of view, the models are simple and not dependent on contextual features. The approach is underpinned in terms of random effects that inform about the attack and defense abilities of the teams enrolled. There is also one parameter that represents the home advantage. From our understanding of the game, a home-team has generally some unobserved advantage. This particular feature is also referred to throughout the literature.