Out-of-Sample Performance of the Premier League Muscular Injury Risk Model

An ensemble model detects early signals of muscular injury risk among Premier League players, built from over 700 features derived exclusively from publicly available data. No club systems, no wearables, no heuristic rules and no synthetic data augmentation.

How it was evaluated

Performance was measured under a locked protocol, fixed before the results were seen. Training and validation used player-grouped splits across the 2020/21 to 2024/25 seasons. The 2025/26 season was held out entirely and used once, as a blind test.

Every feature is computed from data available at the moment of prediction. Nothing is derived from post-injury medical reporting, or from anything recorded after the event date. Validation and test cohorts are separated by both player and time window, so no player’s later observations can inform their earlier ones.

The result

The model scored a Gini coefficient of 0.58 on the blind season. Gini measures how well a model ranks: whether the players it puts at the top are the ones who go on to record injuries.

That figure carries more weight than it might appear to. Muscular injuries develop through complex processes that are only partly predictable, and in this sample an injury occurs on roughly 0.4% of player-days, which makes it a rare event. A stable ranking signal at that base rate, from public data alone and on a season the model had never seen, is what makes the system useful before any club data is introduced.

What it does not claim

The model is an early-warning system, not an oracle. It is not designed to predict every injury as a discrete yes-or-no event, and it is not evaluated as though it were. Its purpose is to rank players by relative risk so that performance and medical staff can concentrate attention where exposure is highest.

Results apply to the Premier League dataset, label definition and evaluation design described in the report. Performance may differ across leagues, injury definitions and deployment contexts.

The feature list, ensemble construction, calibration procedures and source code remain proprietary and are not disclosed.