[Paper Review] A framework for the fine-grained evaluation of the instantaneous expected value of soccer possessions

Javier Fernández, Luke Bornn+1 more|arXiv (Cornell University)|Nov 18, 2020

Sports Analytics and PerformanceEconomics, Econometrics and Finance26 references74 citations

TL;DR

This paper proposes a decomposed deep learning framework for fine-grained, real-time evaluation of the expected possession value (EPV) in soccer by modeling passes, ball drives, and shots as separate components. Using convolutional neural networks and spatial-temporal features, the method produces calibrated, visually interpretable probability surfaces for potential actions, enabling coaches to assess both observed and hypothetical in-game decisions with high precision and interpretability.

ABSTRACT

The expected possession value (EPV) of a soccer possession represents the likelihood of a team scoring or receiving the next goal at any time instance. By decomposing the EPV into a series of subcomponents that are estimated separately, we develop a comprehensive analysis framework providing soccer practitioners with the ability to evaluate the impact of both observed and potential actions. We show we can obtain calibrated models for all the components of EPV, including a set of yet-unexplored problems in soccer. We produce visually-interpretable probability surfaces for potential passes from a series of deep neural network architectures that learn from low-level spatiotemporal data. Additionally, we present a series of novel practical applications providing coaches with an enriched interpretation of specific game situations.

Motivation & Objective

• To develop a flexible, interpretable framework for estimating the instantaneous expected possession value (EPV) in soccer, capturing dynamic in-game states.
• To decompose EPV into distinct on-ball action components—passes, ball drives, and shots—for granular analysis of their individual contributions.
• To enable coaches and practitioners to evaluate not only observed actions but also potential alternative actions through visualized probability surfaces.
• To produce calibrated, spatially aware predictions using low-level spatiotemporal data from player and ball tracking.
• To bridge the gap between complex deep learning models and practical decision-making in professional soccer by enhancing model interpretability.

Proposed method

• The EPV is modeled as the expected value of the next goal outcome (G ∈ {−1, 1}), estimated frame-by-frame using a decomposition into three on-ball action components.
• For full-field prediction of potential passes, a fully convolutional neural network (SoccerMap) is used, processing layered game snapshots at multiple resolution scales (1x, 1/2x, 1/4x) to capture local and global spatial features.
• The network employs 32 and 64 filters with 5×5 kernels, max-pooling for downsampling, and fusion layers to merge predictions across scales before final 1×1 convolutional output.
• For single-location predictions, shallow neural networks are trained on a rich set of novel spatial and contextual features, including dynamic pressure lines (vertical/horizontal) derived from player clustering.
• Dynamic pressure lines are defined as centroids of k=3 clusters (using complete-linkage clustering) to represent team formations (e.g., forwards, midfielders, defenders).
• A baseline expected goals model is trained using XGBoost on OPTA event data (117,948 shots), with features including shot location, distance, angle, and attack type, achieving a test log loss of 0.2540 and ECE of 0.00594.

Experimental results

Research questions

• RQ1 How can the expected possession value (EPV) be decomposed into interpretable components for passes, ball drives, and shots to improve model transparency?
• RQ2 Can deep neural networks produce calibrated, visually interpretable probability surfaces for potential passes using only single-location ground-truth data?
• RQ3 To what extent do novel spatial and contextual features—such as dynamic pressure lines—improve the accuracy and interpretability of EPV estimation?
• RQ4 How can the framework support practical coaching applications by evaluating hypothetical actions not taken during a match?
• RQ5 What is the performance of the proposed models in terms of calibration and predictive accuracy for EPV and related components?

Key findings

• The proposed framework successfully decomposes EPV into three on-ball action components—passes, ball drives, and shots—enabling fine-grained analysis of their individual impacts on possession value.
• The SoccerMap architecture produces calibrated and visually interpretable probability surfaces for potential passes, even when trained with only single-location ground-truth data for the full field.
• The XGBoost baseline model for expected goals achieved a test log loss of 0.2540 and ECE of 0.00594, indicating strong calibration and predictive performance on OPTA event data.
• Dynamic pressure lines, derived from k=3 complete-linkage clustering of player positions, effectively represent team structure and enhance spatial representation in the model.
• The decomposition approach improves interpretability over single-model EPV estimation, allowing coaches to evaluate the impact of alternative actions not executed in real time.
• The framework enables novel practical applications such as real-time visualization of possession value evolution and simulation of alternative tactical decisions during live matches.