Related Experiment Videos
When Timing Matters: Evaluating Temporal Leakage in Machine Learning Models of Football Pass Turnovers
Andrew Peters1,2, Nimai Parmar1, Michael Davies1,2
1Middlesex University.
Research Quarterly for Exercise and Sport
|June 3, 2026
Summary
Correcting temporal leakage in Expected Pass Turnovers (xPT) models improves real-time football analytics. Leakage-corrected models offer tactical utility without significantly sacrificing predictive performance.
Area of Science:
- Sports Analytics
- Machine Learning in Football
- Predictive Modeling
Background:
- The Expected Pass Turnovers (xPT) model quantifies turnover probability in football.
- Temporal leakage from post-pass features limits real-time tactical application of xPT models.
Purpose of the Study:
- To compare the original xPT framework with leakage-corrected alternatives.
- To evaluate four modeling approaches: logistic regression, penalized logistic regression, random forest, and gradient boosting.
- To assess the impact of removing post-execution features on model performance.
Main Methods:
- Utilized 256,433 passes from the 2020-21 English Premier League.
- Compared leakage-inclusive and corrected feature sets using grouped cross-validation.
- Evaluated models using ROC-AUC, accuracy, sensitivity, specificity, F-measure, and Brier score.
Main Results:
- Removing post-execution features reduced ROC-AUC by 0.082-0.183.
- Gradient boosting (ROC-AUC=0.742) approached the performance of the default mixed-effects logistic model (ROC-AUC=0.789).
- Leakage-corrected models shifted focus to pre-execution variables like pressing intensity.
Conclusions:
- Leakage-corrected models retain substantial predictive signal and acceptable performance for real-time tactical deployment.
- Default xPT models suit retrospective analysis; leakage-corrected models are better for real-time tactical decision-making.
