When prediction is not the answer

My NBA player-data study began with a predictive question: can individual statistics explain on-court point differential? I compared linear regression, XGBoost, and CatBoost. The best reported R² was approximately 0.14.

That weak result was a reason to reconsider the question. The work shifted toward SHAP explanations and K-Means clustering, grouping players into six performance-based archetypes. The dataset became more useful as a way to inspect patterns than as a confident prediction engine.

A model explanation is not proof of causality. The useful lesson was methodological: a research project can become stronger when the evidence changes its direction.