This study used in-game markerless motion capture to develop a Super Learner ensemble model for predicting fastball velocity, with the aim of identifying key biomechanical contributors and evaluating generalisability across conferences. Data were collected from 322 NCAA Division I pitchers (ACC: n = 122; SEC: n = 200) between 2022 and 2024. The ensemble combined five algorithms-elastic net, random forest, support vector machines, gradient boosting, and multivariate adaptive regression splines (MARS)-and was trained using 10-fold cross-validation with internal-external validation by conference. Internal performance yielded RMSE 2.70 mph (95% CI 2.44-2.99), MAE 2.10 mph (95% CI 1.93-2.29), R2 = 0.29 (95% CI 0.21-0.36), and calibration slope = 0.99 (95% CI 0.83-1.15). Internal-external performance was direction-dependent (SEC→ACC RMSE = 3.04 mph, 95% CI 2.51-3.63; ACC→SEC RMSE = 2.55 mph, 95% CI 2.33-2.77). Height was the most influential predictor (81.2%), followed by mass (2.47%). Algorithm weights favoured elastic net (62.6%), followed by support vector machines (19.8%), gradient boosting (13.6%), and random forest (4%). Findings suggest that biomechanical strategies for generating velocity vary between groups and highlight the potential of ensemble machine learning to identify biomechanical contributors to fastball velocity and inform individualised analysis in baseball.
Keywords: Baseball kinematics; anthropometrics; multi-model learning; sports performance; throwing mechanics.