A context-aware framework to evaluate passing decisions in soccer

Evaluating passing decision by comparing the value of a pass to the value of realistic alternative passes.

Table 1

Resource	Date	Link
Talk presented at the 2026 Cascadia Symposium on Statistics in Sports (CASSIS)	2026-09-12	Presentation (CASSIS 26)
Talk presented at the 2026 Sports Analytics Workshop (SAW)	2026-05-05	Presentation (SAW 26)

Decomposing Pass Value into Decision and Execution

Most action-value frameworks in football (e.g. VAEP, EPV-added) score a pass by comparing game state before and after the action. This conflates two distinct things: the quality of the player’s decision (which target to choose) and the quality of the execution (where the ball actually ended up) — and it credits the passer with downstream value largely determined by teammates and chance.

We propose a framework that instead decomposes an on-ball action into stages: action choice, target choice, execution. Each stage is then valued by conditioning one step further. For passes, this yields an additive decomposition of total pass value (DES) into a Decision Score (value of the chosen receiver relative to the average option) and an Execution Score (value of the realized outcome relative to the chosen receiver’s expectation). Each component is mean-zero by construction, so a positive value means “better than average” and the components can be analysed independently. Estimation combines a pass-completion model, a graph-neural-network mixture-density model for landing location, a pitch-value surface, and a receiver-selection model, with both the successful and unsuccessful branches of a pass modeled explicitly.

Overview (old)

Important

The description below belongs to an old version of the project. The project has profited from new¹ features that our awesome data provider Gradient sports collects. Although the project changed substantially, the old version below is still a valuable approach for addressing the problem of evaluating passing skills more context-aware.

To evaluate passes, modern passing metrics, rely on machine learning models to estimate a value for each pass. While this is a reasonable approach, it does not account for the distribution of alternatives available to the passer in a given situation. As a result, a safe backwards pass under pressure might be rated negatively even if it was the best reasonable options. Conversely, a short progressive pass might obtain a positive value even though much better options were available. In this work, we address this issue and evaluate the decision behind a pass, taking into account the situational context. To do so, we compare the value of a pass, as estimated by an expected pass value (xPV) model, to the average value of alternative passes weighted by their feasibility and likelihood in the given situation. We obtain this context-dependent baseline for each situation by combining a conditional density surface, which capturing where the ball is likely to end up in a given situation, with a feasibility surface, which capturing where passes are structurally feasible given the game situation. The feasibility surface is obtained by estimating a neural inhomogeneous Poisson process (NIPP), while we use mixture density networks (MDNs) to obtain a full conditional density estimate for the likelihood of a pass. We apply our framework to tracking data from the 2022 FIFA World Cup, enabling the evaluation of player decision- making ability that goes beyond simply averaging pass values, rewarding players who consistently identify and exploit the best available options.

Note

This project is realted to the project: Expected Pass Value (xPV): A holistic framework for evaluating passing situations in soccer.

Footnotes

1. Meaning new to us, we only found out about that after already working a while on the problem