Predictive machine learning models for match outcomes in taekwondo based on competitive history

Abstract

Taekwondo is an Olympic combat sport where performance depends on speed, strength, and tactical precision. Although data-driven methods are advancing in sports science, predictive modeling in taekwondo remains limited. Most existing studies focus on physical metrics or more popular disciplines, leaving a gap in outcome prediction based on competitive history. In this study, we analyze the contribution of technical and contextual features to match outcomes, aiming to identify the most relevant predictors of success. We also propose a dual-structured dataset: one version models individual match sequences, and the other captures pairwise confrontations. This design allows evaluation under both temporal and head-to-head prediction frameworks. Using official data from the Peruvian Taekwondo Sports Federation, we trained and compared eight machine learning models. LightGBM achieved the highest F1-score (84.00%) in the sequence-based format, while XGBoost performed best (75.00%) in the pairwise version. Feature importance analysis revealed second-round actions—clean points and penalties—as key predictors. Our findings demonstrate that machine learning can effectively identify technical and contextual variables that influence match outcomes, offering valuable support for performance improvement, training optimization, and strategic planning in high-performance taekwondo.