A Bayesian approach to predict performance in football: A case study

Gabriel G Ribeiro ¹ Lilia C C Da Costa ¹ Paulo Ferreira ^1* Diego Nascimento ²

• ¹ Federal University of Bahia (UFBA), Salvador, Bahia, Brazil
• ² Universidad de Atacama, Copiapó, Chile

Football is the most practiced sport in the world and can be said to be unpredictable, i.e., sometimes presents surprising results such as a weaker team overcoming a stronger one. As an illustration, the Brazilian Championship Series A (Brasileir ão) has historically been shown to be one of the most outstanding examples of this unpredictability, presenting a large number of unexpected outcomes (maybe given its high competitiveness). This study unraveled attack and defense patterns that may help predict match results for the 2022 Brazilian Championship Series A, using data-driven models considering ten variations of the Poisson countable regression model (including hierarchy, overdispersion, time-varying parameters, or informative priors). As informative priors, the 2021 Brazilian Championship Series A's information from the previous season was adopted for each team's attack and defense advantages estimation. The proposed methodology is beyond helpful for match prediction, it is beneficial for quantifying each team's attack and defense dynamic performances too. As a quality of the forecasts, the DeFinetti measure was used, in addition to comparing the goodness-of-fit using the leave-one-out crossvalidation (LOOCV) metric, in which the models presented satisfactory results. According to most of the metrics used to compare the methods, the dynamic Poisson model with zero inflation provided the best results and, to the best of our knowledge, this is the first time this model is used in a subjective football match context. An online framework was developed, unraveling interactively the access of the obtained results of this study implemented into a Shiny app.