Football Pass Modeling via Expected Possession Value Components with Fully Convolutional Networks

Abstract

Evaluating individual actions in soccer is hindered by the low frequency of goals relative to the volume of observed events. The Expected Possession Value (EPV) framework addresses this by decomposing possession value into subcomponents associated with passes, ball drives, and shots, but adapting it to new datasets is far from straightforward. We adapt the EPV pass components to 52 Premier League matches (2022--2023), building a representation pipeline that encodes each game state into 17 spatial channels and feeds them into the SoccerMap architecture. Even with roughly 12× fewer matches than the original study, the success model yields well-calibrated surfaces that outperform a distance-only baseline, while the selection model ranks the true pass destination among the top 10 pixels in 67.6% of cases out of 7.072 candidates (median rank 3). We further exploit these surfaces through a what-if analysis that perturbs the target player's position on failed passes, revealing actionable positioning adjustments and demonstrating the model's practical value for post-match tactical review.