Forecasting sports outcomes through machine learning

Abstract

Machine learning has increasingly been applied to forecast sports outcomes, assess player performance, estimate injury risk, and model in-game events, yet the reported evidence remains fragmented across sports and validation practices. To address this fragmentation, this systematic review followed PRISMA 2020 reporting principles and analyzed 118 English-language journal articles published between 2020 and 2025 in Scopus and Web of Science, coding each study by prediction domain, sport discipline, algorithm family, dataset type, evaluation metric, and validation approach. Match outcome prediction and player performance assessment accounted for most studies, whereas injury prediction and in-game event forecasting were less frequent and drew on smaller per-domain samples. Random Forest, Gradient Boosting, XGBoost, Support Vector Machines, Neural Networks, and Logistic Regression were the most common methods: ensemble models predominated in structured tabular datasets, whereas deep learning was concentrated in video, tracking, temporal, and multimodal contexts. Reported performance was difficult to compare directly because studies used heterogeneous metrics, partitions, datasets, and validation schemes. These patterns indicate that data access, task design, and validation quality matter as much as algorithm choice; future work should therefore prioritize open benchmarks, external validation, and causal modeling, while privacy-aware data governance, uncertainty estimation, and interpretability remain underdeveloped priorities for the field.