Revisiting Expected Possession Value in Football: Introducing a U-Net Architecture, Reward and Risk for Passes, and a Benchmark

This paper presents an Expected Possession Value (EPV) model for football with three main new components: a U-Net-inspired convolutional neural network architecture, ball height as a feature, and a dual-component pass value model that analyzes reward and risk. We furthermore introduce the Overmeer-Janssen-Nuijten Pass Expected Possession Value benchmark (OJN-Pass-EPV benchmark), which enables a quantitative evaluation of EPV models by using pairs of game states with given relative EPVs. The presented EPV model achieves good results in model loss and Expected Calibration Error on a dataset containing Dutch Eredivisie and 2022 FIFA Men's World Cup matches and correctly identifies the higher value state in 78% of the game state pairs in the OJN-Pass-EPV benchmark, demonstrating its ability to accurately assess goal-scoring potential. Our findings enable more precise EPV estimations, support risk-reward analysis for passing decisions, and establish quality control standards for EPV models.