Adaptive multi-parameter fusion for sports outcome prediction: a context-aware intelligent platform

Abstract Sports outcome prediction is crucial for performance assessment and strategic decision-making. Traditional models (Elo-based, regression-based) rely on limited historical data and fail to capture dynamic interactions between players, teams, and contexts. We introduce an adaptive multi-parameter fusion (AMPF) framework using hierarchical attention-based neural networks to integrate team ratings, performance trends, and contextual factors. Unlike fixed-weight models, AMPF employs learnable attention mechanisms for adaptive, context-specific feature weighting. Implemented in PyTorch, AMPF is validated on FiveThirtyEight datasets covering 45,000+ football matches (2010–2024) and 20,000+ basketball games (2012–2024). AMPF achieves statistically significant improvements over all baselines: 1.6–2.2% higher accuracy and 3.8–5.6% lower Brier scores compared to deep MLPs, with larger gains (4.7–4.8% accuracy, 10.9–11.8% Brier reduction) over classical Elo ratings. Ablation studies confirm the necessity of contextual embeddings and adaptive fusion. We deploy a platform providing real-time predictions, interpretability visualizations, and provenance tracking, bridging traditional rating systems and modern deep learning for sports analytics.