The Past Informs the Future: Temporal Sliding Window Graph Attention Networks for Real-Time Football Match Outcome Prediction

Real-time football match outcome prediction is challenged by the sport’s inherent dynamic nature. While recent graph-based approaches have shown promise, they primarily rely on cumulative passing networks that aggregate data from kickoff, potentially obscuring critical short-term tactical shifts. We propose a Temporal Sliding Window Graph Attention Network (TSW-GAT) that addresses this limitation through a novel dual-graph architecture. Unlike prior models that rely solely on cumulative representations, our framework simultaneously processes global team structures and localized momentum signals through a compact integration of cumulative and interval-specific graphs. Furthermore, we introduce a substitution-aware construction method to ensure network consistency during tactical changes. Trained and validated on 2,806 English matches, TSW-GAT achieves 81.85% accuracy at full-time and 77.94% at halftime. Notably, our model outperforms cumulative-only baselines by up to 20.58% points at the 45-minute mark, demonstrating superior performance when in-match evidence is limited. Cross-league evaluation on 2,800 matches (700 per league) from Bundesliga, La Liga, Ligue 1, and Serie A further confirms strong generalization, with an average accuracy of 84.24% at 90 min. These results highlight the efficacy of temporal decomposition in capturing evolving match contexts. The source code for this work is publicly available at: https://github.com/merthann/The-Past-Informs-the-Future.