Machine learning-based winner prediction in elite kickboxing: a technical-tactical performance analysis

To develop and validate machine learning models for predicting match outcomes in elite kickboxing based on technical-tactical performance indicators and to identify the key variables distinguishing winners from defeated athletes. Video analysis was conducted on 142 elite kickboxing matches from WAKO-sanctioned international competitions, yielding 284 individual observations. Fifty technical-tactical variables encompassing punch techniques, kick techniques, and landing rates were extracted through standardized notational analysis. Five machine learning algorithms were evaluated using group k-fold cross-validation to prevent data leakage between paired competitors, with 95% confidence intervals for performance metrics estimated through match-level bootstrap resampling. Feature importance was assessed through permutation analysis, and winner-defeated differences were quantified using Wilcoxon signed-rank tests, appropriate for the paired structure of same-match observations, with Cliff’s Delta effect sizes and Benjamini-Hochberg correction for multiple comparisons. The Support Vector Machine achieved the highest predictive performance with an area under the receiver operating characteristic curve of 0.845 (95% CI: 0.795–0.883) and accuracy of 78.5% (95% CI: 74.3–82.7%). Landing variables, particularly total kick landings, demonstrated the strongest predictive importance and largest paired effect sizes when comparing winners to defeated athletes. Kick landing rate emerged as a more powerful predictor than punch landing rate or attempt frequencies, and winners also showed substantially higher landing accuracy percentages than defeated athletes. Machine learning models demonstrated promising internally validated performance for classifying winner and defeated-athlete observations within the analyzed elite WAKO K-1 match sample. The findings indicate that landing-related variables, particularly kick landings and landing accuracy, were the strongest technical-tactical indicators associated with competitive success. These results may inform training program design and tactical analysis, but external validation across independent competitions is required before broader generalization.