Abstract

Given the recent advent of granular, tracking, and online data, the art of modeling football data is of ever-increasing relevance. In this talk, I will be dealing with different modeling approaches suited for different data structures: according to data availability, football research may focus in fact on modeling the match results, players performance, competitive balance, and tournaments' evolution.
A strong emphasis will be put on modeling the result of a football match, by considering the inherent amount of uncertainty as provided by both frequentist and Bayesian estimating procedures. According to this aim, many researchers discussed the well-known assumptions underlying the football protocol: goal dependence, overdispersion, and the choice of a suitable distribution for the scoring process of the two opposing teams. I'll be discussing how to criticize/meet these assumptions and, then, how to choose a better model for the football results.
In terms of players' peformance, a large interest relies on tracking data and sophisticated available statistics, such as the number of passes, assists, shots, yellow and red cards of a given player, and so on. However, translating these numbers into some 'statistical wisdom' is an open and unavoidable challenge in the modern field of football modeling. I'll be discussing how some supervised and unsupervised learning techniques could help in determining the drivers for selecting a valuable player.
In general, I'll be providing some practical examples from our most recent experience and contained in our book 'Predictive football analytics', jointly written with Ioannis Ntzoufras and Dimitris Karlis; moreover, an overview of the main computational practices will be provided with the illustration of the main functionalities of the R package 'footBayes'.