Using Machine Learning to Rate Attributes of Football Players

Sammanfattning

This thesis proposes a data-driven framework for constructing interpretable football player profiles from event data. While substantial research and development has been conducted in football analytics, there is still a lack of comprehensive frameworks that summarise football player attributes in an interpretable way. This research gap is addressed by combining several modelling and validation techniques into a unified player profiling framework. Using StatsBomb event data from the 2015/16 Premier League, Bundesliga, Serie A, La Liga and Ligue 1 seasons, the thesis includes 1028 matches and 2153 eligible outfield players. The framework combines expected-event models, residual-based attribute estimation, empirical success-rate measures, proxy-based attributes, empirical Bayes shrinkage and clustering based on dimensionality reduction. Player attributes are normalised to a 1–99 rating scale, while play-style profiles are derived separately within positional groups. The expected pass model and expected goal-on-target model showed strong event-level performance on external open-data matches. Residual models for vision and offensive positioning also generalised reasonably well and improved on baseline predictors. Attribute-level validation showed that directly observable attributes, such as dribbling, short passing and heading, aligned well with football expectations and external FIFA 17 ratings. More abstract or proxy-based attributes, including vision, composure and first touch, showed weaker or less consistent validation results. Play-style clusters were only moderately separated according to internal clustering metrics, but their feature profiles remained interpretable from a football perspective. The results suggest that event data can support the construction of interpretable player profiles, particularly for attributes closely linked to observable on-ball actions. However, attributes depending on off-ball behaviour, contextual interpretation or low-frequency events should be interpreted with caution. The proposed framework should therefore be viewed as an analytical support tool rather than an objective replacement for human judgement.