Multi-Output Regression for the Prediction of World-Class Performances in Women’s Handball

The purpose of this study was to investigate the relationship between workload and in-game technical and athletic performance. To achieve this, multi-output regression models were used to predict 7 performance indicators based on previous training and game athletic workloads measured by Inertial Measurement Units (IMU) indicators, previous in-game actions annotated by staff members and game contextual factors. We compared 4 single-output models (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i>NN, regression tree, random forest and neural networks), their multi-output counterparts and a dummy baseline (predicting the average performance of each player over the last month) in terms of average RMSE (aRMSE) during a chronological evaluation where previous trainings and games data are used to train models to predict the next game performances. Overall, the use of multi-output regression models enabled a decrease of the average predictive error (aRMSE = 4.23) in regards to their single-output counterparts (aRMSE = 4.35) while providing a significant decrease of average computation times (4.75 to 0.82 seconds). Among the 4 multi-output models, only the kNN (aRMSE = 3.852) and random forest (aRMSE = 3.888) performed better than the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">dummy</i> regressor (aRMSE = 3.944). These results point towards that physical training may have a limited impact on game performance.