Leakage-Safe and Explainable Machine Learning for Predicting NBA Team Outcomes and Performance Deviations

Abstract

Two classification tasks on NBA team standings data are examined: predicting whether a team qualifies for the playoffs from its mid-season record, and predicting whether a team's season will underperform, match, or overperform the level suggested by its previous-season record. Seven classical machine learning models (Logistic Regression, Random Forest, Extra Trees, Gradient Boosting, XGBoost, LightGBM, and CatBoost) are trained and compared under a leakage-safe framework in which features are restricted to information available at the prediction point and the data are split chronologically, with the earliest 80% used for training and the most recent 20% for testing. For playoff qualification, Random Forest is the best-performing model, reaching a test accuracy of 91.7% against a majority-class baseline of 53.3%, although accuracy is considerably lower for teams close to the eighth-seed boundary. For deviation from expected performance, Logistic Regression, tied with Gradient Boosting, is the best-performing model at 70.0% accuracy against the same 53.3% baseline. SHAP analysis identifies conference rank at mid-season as the most influential feature for playoff qualification, followed by win percentage, and the difference between early-season and previous-season win percentage as the most influential feature for deviation from expected performance.