<p>Traditional sports outcome prediction models predominantly rely on aggregate team-level statistics and scalar rating systems, treating teams as monolithic entities and overlooking the intricate web of player interactions, positional dependencies, and tactical coordination that fundamentally shape competitive performance. Recent advances in graph neural networks (GNNs) and relational learning offer a promising avenue for modeling sports outcomes: representing teams as dynamic interaction graphs where nodes correspond to individual players and edges encode cooperative or adversarial relationships. This paper presents a temporal graph neural network (T-GNN) framework for football match outcome modeling that explicitly captures evolving player-level interactions, spatial formations, and time-varying team structures from event-stream data. Unlike conventional feature-fusion approaches, our method constructs match-specific dynamic graphs—derived from pass frequencies, defensive duel partnerships, and spatial co-occurrence patterns—and applies graph convolutions combined with recurrent temporal encoders to capture both local interaction patterns and global tactical evolution. We validate T-GNN on two publicly available football datasets—StatsBomb Open Data (covering 11 competitions with over 3,500 matches) and Wyscout match event data (five European leagues, over 1,900 matches)—using strictly held-out test sets. Graph edges are constructed from observed pass frequencies, defensive duel partnerships, and spatial co-occurrence patterns, all directly extracted from timestamped event records. Experimental results show that T-GNN improves over non-graph baselines—including logistic regression, gradient boosting, and standard recurrent neural networks—by 3–9 percentage points in accuracy and 5–17% in Brier score (relative reduction); improvements over the best sequential baseline (LSTM) are 3.2 pp and 5.3% respectively, while larger gains are observed over statistical baselines; all improvements are statistically significant (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(p < 0.01\)</EquationSource></InlineEquation>). Beyond predictive performance, the learned representations support post-hoc relational analysis of critical player roles, interaction pathways, and formation stability dynamics. The framework remains constrained by the granularity and completeness of event-based data and does not model high-frequency tracking trajectories or off-ball dynamics that are absent from the available event logs. A prototype decision-support interface is developed to enable post-match relational exploration, assisting coaches and analysts in examining tactical patterns derived from event-stream representations.</p>

Temporal graph neural networks for sports outcome prediction: modeling dynamic player interactions and evolving team structures

期刊论文
出版模式：

开放获取
发表日期：

2026年7月29日

查看全文

Yanxiao Zhang,
Shuangcong You

摘要

Traditional sports outcome prediction models predominantly rely on aggregate team-level statistics and scalar rating systems, treating teams as monolithic entities and overlooking the intricate web of player interactions, positional dependencies, and tactical coordination that fundamentally shape competitive performance. Recent advances in graph neural networks (GNNs) and relational learning offer a promising avenue for modeling sports outcomes: representing teams as dynamic interaction graphs where nodes correspond to individual players and edges encode cooperative or adversarial relationships. This paper presents a temporal graph neural network (T-GNN) framework for football match outcome modeling that explicitly captures evolving player-level interactions, spatial formations, and time-varying team structures from event-stream data. Unlike conventional feature-fusion approaches, our method constructs match-specific dynamic graphs—derived from pass frequencies, defensive duel partnerships, and spatial co-occurrence patterns—and applies graph convolutions combined with recurrent temporal encoders to capture both local interaction patterns and global tactical evolution. We validate T-GNN on two publicly available football datasets—StatsBomb Open Data (covering 11 competitions with over 3,500 matches) and Wyscout match event data (five European leagues, over 1,900 matches)—using strictly held-out test sets. Graph edges are constructed from observed pass frequencies, defensive duel partnerships, and spatial co-occurrence patterns, all directly extracted from timestamped event records. Experimental results show that T-GNN improves over non-graph baselines—including logistic regression, gradient boosting, and standard recurrent neural networks—by 3–9 percentage points in accuracy and 5–17% in Brier score (relative reduction); improvements over the best sequential baseline (LSTM) are 3.2 pp and 5.3% respectively, while larger gains are observed over statistical baselines; all improvements are statistically significant (\(p < 0.01\)). Beyond predictive performance, the learned representations support post-hoc relational analysis of critical player roles, interaction pathways, and formation stability dynamics. The framework remains constrained by the granularity and completeness of event-based data and does not model high-frequency tracking trajectories or off-ball dynamics that are absent from the available event logs. A prototype decision-support interface is developed to enable post-match relational exploration, assisting coaches and analysts in examining tactical patterns derived from event-stream representations.