From Patterns to Playbooks : Identifying offensive and defensive team playing styles from possession chains with machine learning and its possible practical applications in management of football teams

Analyzing playing styles is an important task in many parts of football, from performance evaluation to scouting of both opponents and potential player signings. Manual analysis can be time-consuming and subjective and current data-driven approaches face a number of challenges including 1) a lack of established definitions for playing styles 2) under exploration of defensive playing styles, 3) mainly being based on event data, with few studies using tracking data (especially for defence), 4) a lack of granularity, with many studies analysing playing styles on aggregated match or season level statistics and 5) a lack of studies combining methodological and applicational perspectives. The study aims to bridge this gap by using unsupervised machine learning to cluster possession chains, using both tracking and event data, enabling both a more granular analysis and the creation of novel features for both offensive and defensive play. The study was performed on a dataset obtained from Twelve Football, containing both event and tracking data from the 2022 season of Swedish first division Allsvenskan. The dataset is split in possession sequences and 13 features for offensive play and 12 for defensive play are generated. The features are scaled and then clustered using K-Means clustering. In total 6 offensive and 6 defensive clusters were found, and 4 different applications for football teams were proposed. The method, results and applications were discussed and presented to both football researchers and club professionals who showed great interest in the methodology and results for application in football analytics. In conclusion, applying unsupervised machine learning methods like K-Means clustering on possession chains with tracking and event data can unravel interesting insights of the playing styles adopted by teams in both offense and defence. By doing this at the possession level, both more detailed insights and potential new applications in scouting and performance evaluation can be enabled due to the ability to link offensive and defensive playing styles together with specific player or team statistics from the event and tracking data.