A machine learning framework for off-ball defensive evaluation applied to corner kicks

Abstract Evaluating off-ball defensive performance in football is challenging, as traditional metrics do not capture the nuanced coordinated movements that limit opponent action selection and success probabilities. While possession value models focus on on-ball actions, their application to defense remains limited. Existing counterfactual methods, such as ghosting models, help extend these analyses but often rely on simulating “average” behaviour that lacks tactical context. To address this in the highly structured context of set pieces, we introduce a covariate-dependent Hidden Markov Model tailored to corner kicks. Incorporating domain expert-informed constraints, such as independent defender modelling and fixed spatial emissions, our label-free model infers time-resolved man-marking and zonal assignments directly from player tracking data. We propose a novel evaluation framework culminating in the Group Coverage Advantage (GCA) metric. This approach: (1) infers latent tactical roles; (2) generates locally plausible, role-conditioned counterfactuals ("ghosts”) rather than generic averages; and (3) evaluates defensive positioning via a downstream risk model. Applying GCA to over 13,000 Premier League corners reveals that while average defensive quality remains stable regardless of delivery trajectory (Cohen’s d < 0.11), outswinging deliveries create significantly higher outcome volatility ( p < 10 −40 ). We show how these contributions provide an interpretable evaluation of defensive performance against context-aware baselines.