Early Prediction of Physical Performance in Elite Soccer Matches-A Machine Learning Approach to Support...

Area of Science:

• Sports Science
• Data Science
• Machine Learning

Background:

• Player substitutions are crucial for tactical adjustments and managing player performance in elite soccer.
• Predicting individual player physical performance early in a match can significantly enhance coaching decisions.

Purpose of the Study:

• To develop machine learning models for predicting individual player physical performance in the early stages of a soccer match.
• To provide coaches with data-driven insights for optimizing player substitution strategies.

Main Methods:

• Utilized tracking data from 302 elite soccer matches (Dutch Eredivisie, 2018-2019 season).
• Focused on physical performance variables: distance covered, distance by speed category, and energy expenditure in the power category.
• Applied Random Forest and Decision Tree algorithms to predict if players would achieve 100%, 95%, or 90% of their average performance.

Main Results:

• The Random Forest model, using energy expenditure in the power category, demonstrated high precision in predicting performance.
• Precisions for predicting performance at 100%, 95%, and 90% thresholds after 15 minutes were 0.91, 0.88, and 0.92, respectively.
• Machine learning models successfully predicted early-phase physical performance, outperforming a Naïve Bayes baseline.

Conclusions:

• It is feasible to predict individual player physical performance during the early phase of a soccer match.
• These predictive capabilities offer valuable support for coaches in making timely and informed substitution decisions.
• The study highlights the potential of machine learning in enhancing strategic decision-making in elite soccer.