Precision Training Via Causal Machine Learning: Modeling Rating of Perceived Exertion in Professional Soccer Players

Abstract

Purpose: This study aimed to explore the use of predictive and prescriptive machine-learning models for managing training loads in professional soccer, with a focus on the rating of perceived exertion (RPE). Using data from a Belgian Pro League club, we evaluated the effectiveness of these models in predicting and prescribing optimal training regimens. Methods: Data from 14 players across a full competitive season were analyzed. Predictive models including linear regression, random forest, and XGBoost were compared using the root-mean-square error and the mean absolute error. SHapley Additive exPlanations values were used to interpret feature importance. A prescriptive model based on the counterfactual recurrent network was developed to optimize training inputs for desired outcomes. Results: The XGBoost model demonstrated the best predictive performance (root-mean-square error: 1.262), with session distance identified as the most significant driver of RPE. While the prescriptive counterfactual recurrent network model showed slightly lower predictive accuracy (root-mean-square error: 1.379), its unique advantage lies in estimating counterfactual outcomes, allowing for the simulation of future RPE trajectories under different potential training plans and providing actionable insights for personalized training prescription. Conclusions: Predictive modeling effectively estimates RPE, and prescriptive modeling offers the added benefit of optimizing training strategies. The integration of these approaches supports data-driven decisions in professional soccer, enhancing player performance and recovery. Future research should expand sample sizes and validate these methods across diverse sports and contexts.