This AI Model Blames Football Defenders for Their Positioning Errors — And It's Changing How We Analyze Defense

A team of researchers (Bischofberger, Ma, Bauer, Arnsmeyer, Baca) from the multi-agent systems domain have published a paper tackling a long-standing gap in football analytics: measuring off-ball defensive performance. Traditional metrics only count visible actions like tackles and interceptions, ignoring the continuous spatial impact defenders have. The team formulated this as an attribution problem over multi-agent spatiotemporal trajectories, using defensive pressure areas (DPAs) to compute player involvement scores. By establishing role-conditioned baselines within automatically detected team structures, they determine each defender's expected responsibility for threat created by any pass.

The framework was validated on an extensive cross-gender and cross-competition dataset: 64 men's World Cup matches, 116 women's German Bundesliga matches, and 336 men's German 3. Liga matches. In the absence of ground truth, they combined multiple weak proxies into robust summary scores, achieving a validity improvement of roughly one standard deviation over top action-based metrics. Notably, the 'blame' score—measuring how much responsibility a defender bears for conceding high-value actions—showed strong correlations with external ratings and market values, making it the first published metric that reliably captures positioning errors. All code is open-source to support reproducibility.