Wearable sensors in team sports enable analysis of massive amounts of movement data collected during training and competition, with researchers primarily focused on determining physiological demands and running profiles of athletes. Recent advancements in machine learning have expanded these analyses, enabling prediction of physiological metrics and supporting data-driven decision-making by coaches and practitioners. However, research applying machine learning to wearable sensor data remains in early stages. This thesis develops a feature engineering framework that transforms raw sensor data from elite Gaelic football into structured datasets, substantially reducing data size while preserving analytical information and enabling machine learning models to forecast high-speed distance covered by players in the second half of matches. Spatio-temporal graphs of player actions reveal spatial movement patterns across consecutive time windows. To anticipate bursts of player effort, a metabolic power-based metric called the Metabolic Work Ratio is introduced for prediction. Machine learning models predict this ratio in upcoming time windows using features from current and previous periods. Finally, to reduce dependence on video footage and manual annotation, shooting events are classified using only single-team sensor data (without ball tracking or opposing team data) through pairwise movement relationships between players, represented as adjacency matrices of sensor variable differences, which serve as machine learning input features.