Learning state representations and Markov models in football analytics

Abstract

The increasing availability of spatio-temporal data of football matches in recent years has prompted the interest of many clubs in performing automated tactical analysis using machine learning techniques to gain competitive advantage. The low-scoring nature of the sport, the highly dynamic interactions and the presence of contextual circumstances that change continuously present challenges for automated analysis. Using data from football matches of FC Barcelona B, this work aims to automatically learn a meaningful state representation using high-level features that include contextual information about the game and to estimate basic Markov models from the transition probabilities between the states to help coaches to understand player and team behavior. Multiple clustering techniques have been tested to define states and a basic Markov model has been estimated for different teams. This allows modeling how possessions can unfold in any given number of passes, as well as estimating the probabilities of keeping possession or for it resulting in either turnover, shot or goal. It has been shown that even a simple model yields useful results for the club analytics team, that can be used to analyze how a team plays. Also, that this highlevel representation can help significantly to facilitate the communication between coaches and analysts thanks to its interpretability.