From Prediction to Understanding: Explainable Machine Learning for Elite Winter Sports Performance

Machine learning (ML) has become increasingly important in sports performance analysis because of its ability to model complex relationships in high-dimensional biomechanical and contextual data. However, many high-performing models operate as black boxes, limiting their practical value in elite sport contexts where coaches, athletes, and sport scientists require interpretable and trustworthy insights. This thesis investigates the use of explainable machine learning for winter sports performance analysis, with a specific focus on ski flying and ski jumping. The research is organised around two complementary case studies. The first case study examines ski flying performance using wearable-sensor-derived flight data and official competition results. Machine learning models were developed to predict official landing distance from phase-specific features representing take-off, early flight, stable flight, and landing preparation. Explainable AI (XAI) methods were then applied to identify the most influential variables and compare the predictive value of different flight phases. The results showed that landing preparation and stable flight contained the strongest predictive information, while speed-related variables, particularly horizontal speed and official take-off speed, were central drivers of model predictions. The second case study examines ski jumping performance using video-derived biomechanical features combined with official competition variables. A data-processing pipeline was developed to clean pose-coordinate data, extract biomechanical features, integrate official result information, and train machine learning models across sex-specific, discipline-specific, hill-type-specific, and subgroup-specific datasets. The explainability analysis showed that both contextual variables, such as gate and wind, and biomechanical variables related to body orientation, lower-limb movement, and take-off mechanics contributed to predicted jump distance. Across both case studies, the findings demonstrate that explainable machine learning can support winter sports performance analysis by linking prediction with biomechanical and contextual interpretation. Explainable AI was useful not only for interpreting model behaviour, but also for auditing feature validity, detecting potentially problematic predictors, and strengthening confidence in model outputs. In addition to quantitative evaluation of explanation fidelity and robustness, a questionnaire-based expert evaluation was conducted with seven participants, comprising four ski-jumping experts and three machine-learning experts, to assess the understandability, trustworthiness, plausibility, and usefulness of the generated explanations. Overall, the thesis contributes an applied framework for explainable machine learning in elite winter sports and shows how predictive models can move beyond numerical accuracy toward interpretable, domain-relevant understanding.