23rd Conference of the International Federation of Operational Research Societies
Abstract Submission

1277. Detection of playing styles in professional football through Bayesian Inference

Invited abstract in session FB-5: Bayesian Models for Football Analytics, cluster OR in Sports.

Friday, 11:00-12:30
Room: CE-204

Authors (first author is the speaker)

1. Constanza Encina
Universidad de Chile

Abstract

Football analysis covers a wide range of techniques and applications, including player tracking, performance analysis, optimization of game strategies and scouting. In regards to scouting, professional clubs seek to bring in new players who can contribute to their game strategy for the new season. Currently, they bring in players who were successful at other clubs, without necessarily taking into account the difference between playing styles of the respective teams. This raises the difficult task of integrating them into the team dynamics and adapting them to the game plan.

These new signings must adapt as quickly as possible, so they perform at the same level as in their previous clubs or, even better, they attain superior performances. The objective of this research is to detect team playing styles, to identify similar ones and to improve adaptation of the new player.

To achieve this, an unsupervised classification method is proposed - in particular, a bayesian inference model, to provide a probabilistic framework for relationships between metrics of performance of each team and the playing styles. It incorporates prior knowledge about the data and adjusts the model accordingly, leading to more accurate and robust results.

The project uses InStat data from the 2021 championship of the Chilean first division. The variables considered describe in aggregate form what happened on the field at the match and team level, such as defensive disputes, offensive disputes, passes and shots on goal.

The proposed Bayesian model estimates the probability distribution of the parameters “playing style” and “types of team”, given data. This allows us, through an analysis of the results, to declare certain “playing styles” and the teams belonging to each one, with a certain probability. Subsequently, a sensitivity analysis will be carried out to better adjust the model and define parameters, such as how many clusters are to be obtained, how many performance variables should be considered in the data, and, more importantly, the specific variables.

This model could potentially be extrapolated to many leagues, from many countries and in different team sports, to obtain more accurate results, to facilitate adaptation efficiency of a player into a new team or to analyze a rival team’s playing style.

Keywords

Status: accepted

Back to the list of papers