Winning beyond scoring: an explainable machine learning framework for identifying key performance indicators in the Chinese Basketball Association.

backgroundUnderstanding why teams win in professional basketball requires more than examining isolated statistics. Performance adapts under changing competitive constraints, and key performance indicators (KPIs) must therefore be interpreted within context. This study aimed to develop an explainable, multi-model machine learning (ML) framework to identify score-line dependent KPIs associated with winning outcomes in the Chinese Basketball Association (CBA).

methodsA total of 3,249 games from the 2018–2022 CBA regular seasons were analyzed. Based on final score margin, games were classified into close (1–9 points) and balanced (10–18 points) contexts using k-means clustering. For each context, game-related statistics and play-type variables were initially screened using conventional statistical procedures. A dual-stage feature selection process combining lasso and stepwise logistic regression retained indicators consistently selected by both models. Ten commonly used ML classifiers were then trained and evaluated using nested cross-validation. Rather than benchmarking predictive accuracy alone, a convergence-based strategy was adopted, treating KPI relevance consistently identified across classifiers as stronger structural evidence. Shapley Additive exPlanations (SHAP) were used to quantify the magnitude and direction of each KPI's contribution within the top-performing classifiers.

resultsModel performance differed across score-line contexts. In close games, the BP neural network, support vector machine, and logistic regression achieved the best performance (accuracy: 71.2%–76.4%; AUC: 0.80–0.83), while tree-based ensemble classifiers (e.g., XGBoost, CatBoost, and LightGBM) showed superior discrimination in balanced games (accuracy: 83.9%–86.7%, AUC: 0.92–0.93). SHAP analysis revealed distinct context-dependent KPI structures. Close game wins were characterized by greater defensive rebounding and fewer turnovers, defensive fouls, and missed perimeter attempts. Screen-off made actions emerged as the only retained play-type indicator. Balanced games displayed a broader KPI structure involving higher relevance of rebounding, contested field-goal efficiency, steals, and play-type behaviors including screen-off, pick-and-roll roller, and pick-and-pop actions.

conclusionWinning outcomes in the CBA depend less on scoring volume alone and more on how teams regulate coordination and possession stability under different score-line constraints. Close games favor defensive integrity and the control of destabilizing events, while balanced games reward sustained possession management and diversified off-ball coordination. These findings provide actionable guidance for context-sensitive training design and tactical decision-making in professional basketball.