Location-based movement pattern classification and completed-bout outcome discrimination in taekwondo under the best-of-three system

Event-based taekwondo analyses describe technical and tactical actions but provide limited information about where these actions occur and whether spatial organization is maintained across independently scored rounds. This study classified athlete-level spatial–behavioral profiles and examined their relationships with completed-bout outcome discrimination, scoring efficiency, and between-round win maintenance under the best-of-three system. Publicly available videos of 112 completed bouts involving 121 male university athletes from the 2022 Korea Taekwondo Association President’s Cup were analysed. Event-linked positional coordinates and technical–tactical actions were recorded from 16,665 events. Athlete-level profiles were derived using K-means clustering based on five spatial–behavioural variables. Candidate solutions from K = 2 to K = 10 were evaluated using silhouette, Davies–Bouldin, Calinski–Harabasz, stability, cluster-size balance, and parsimony. Completed-bout outcomes were classified using logistic regression, random forest, support vector machine, and multilayer perceptron models. Performance was assessed using leave-one-weight-class-out cross-validation and ROC AUC, PR-AUC, accuracy, balanced accuracy, F1, and Brier scores. Test-fold permutation importance and external profile analyses of scoring efficiency, scoring location, and Round 1-to-Round 2 win maintenance were also conducted. The two-cluster solution showed the best overall balance of separation and stability. C1, representing a central stable–offensive style, comprised 71 athletes, whereas C2, representing a peripheral variable–defensive style, comprised 50 athletes. Random forest achieved the highest OOF AUC (0.749, 95% CI 0.658–0.821) and OOF accuracy (68.8%). Overall scoring efficiency did not differ between profiles ( p = 0.652). In descriptive zone-specific comparisons, raw scoring efficiency was 6.52% versus 5.45% in the central zone, 7.97% versus 5.96% in the middle zone, and 8.66% versus 7.32% in the peripheral zone for C1 and C2, respectively. In mixed-profile bouts, Round 1 wins were maintained into Round 2 more often when the Round 1 winner was C1 than C2 (90.2% vs 68.2%; OR = 4.32, 95% CI 1.10–16.94, p = 0.039). Event-linked positional data identified two reproducible spatial–behavioural profiles. Overall scoring efficiency was comparable, while the raw zone-specific percentages descriptively characterized how scoring success was distributed across court locations in each profile. The central stable–offensive C1 profile showed stronger Round 1-to-Round 2 win maintenance than the peripheral variable–defensive C2 profile. Random forest provided the highest observed cross-validated discrimination.