Machine learning (ML) is increasingly used in team sports, yet existing reviews vary widely in scope, focus, and methodological rigour. This umbrella review synthesizes evidence from 12 reviews encompassing 263 primary studies to provide a structured overview of ML applications in this field. Reviews were classified by task (Match Outcome Prediction, Performance Prediction, Performance Evaluation, Playing Style Identification, Player Evaluation), sport, algorithm family, and data modality. Soccer was the most frequently studied sport, followed by basketball, while other football codes, handball, and ice hockey were underrepresented. Outcome prediction dominated the literature, although performance prediction tasks typically reported higher predictive accuracy. Low-scoring sports such as soccer showed lower predictability than high-scoring sports like basketball. Neural networks and probabilistic approaches consistently performed well, whereas ensembles displayed heterogeneous results and linear or single-tree models generally underperformed. Tracking data yielded better results than purely notational data, though combining modalities did not consistently improve accuracy. Persistent gaps include the absence of benchmark datasets, inconsistent evaluation metrics, and limited consideration of ethical issues such as bias, interpretability, and fairness. Addressing these gaps is critical for translating ML advances into reliable and actionable insights for sports practice.
Klemp et al. (2026) studied this question.