Cross-individual generalizability of machine learning models for ball speed prediction in baseball pitching

A new study by Ryota Takamido and colleagues evaluates how well machine learning models for baseball pitching ball speed prediction perform when applied to pitchers not seen during training. Using a dataset of 50 pitchers ranging from amateur to expert, the team employed leave-one-subject-out cross-validation. Results showed a dramatic drop in predictive power: within-individual R-squared was 0.91, but cross-individual R-squared fell to 0.38 — a 58% reduction. The models also exhibited systematic bias, overestimating the performance of intermediate-level pitchers compared to experts (p < .05).

However, not all motion features were equally problematic. The trunk and pivot leg showed relatively high generalization performance, with the pivot leg demonstrating notable generalizability (R² > 0.25) even during the weight-shift initiation phase. The findings underscore that current ML models struggle to transfer across athletes without individual calibration, but specific biomechanical features like pivot leg motion may offer a more robust foundation for future models. The study appears on arXiv (2605.05487) and highlights the critical need for cross-individual evaluation in sports analytics.