Event-specific versus stable predictors of medal outcomes: A comparative machine learning analysis of the Asian Games and Olympic Games

Elite sporting success reflects interactions among athlete characteristics, performance capacity, and support environments; however, predictor stability across competitions remains unclear. This study applied machine learning models to examine whether predictors of medal achievement generalize across international multisport events. Data from 1011 elite athletes from the Hangzhou Asian Games and Paris Olympic Games were analyzed using gradient boosting classifiers (XGBoost, CatBoost, and LightGBM). Predictors included demographics, sports discipline, physiotherapy variables, and movement performance scores. Model performance was evaluated using nested cross-validation and cross-event transfer testing based on a harmonized common-feature set, with naïve prevalence-based and L2-regularized logistic regression benchmark models. Within-event models demonstrated strong discrimination across medal outcomes. Cross-event transfer revealed two distinct patterns. Gradient boosting models showed substantial transfer declines. In contrast, L2-regularized logistic regression showed partial transportability, outperforming or matching gradient boosting models for bronze and overall medal achievement. Sports discipline remained the dominant predictor, whereas movement performance and physiotherapy variables showed greater event-specific variability. Several predictors retained consistent directions across events. Collectively, these findings indicate that the underlying predictive signal was partially transportable, whereas flexible nonlinear models were more vulnerable to distributional shift, highlighting the importance of benchmarking, model parsimony, and transportability assessment in athlete-level predictive modeling.