Predicting NBA Rookie Impact from NCAA Statistics: A Machine Learning Approach Using Regularized Adjusted Plus-Minus (RAPM)

Abstract

Projecting which college basketball players will succeed in the NBA remains one of the most challenging tasks in professional sports. This study utilizes publicly available NCAA statistics to investigate whether machine learning models can predict players’ rookie window impact, using Regularized Adjusted Plus-Minus (RAPM) over their first four NBA seasons as the proxy metric. Using data on 546 drafted players and undrafted free agents from 2008 to 2021, the analysis applies extensive feature engineering including era-adjusted efficiency metrics, per-100 possessions standardization, team quality ratings and team strength of schedule. A temporal train (2008-2018) / test (2019-2021) split is employed to evaluate whether the model can generalize on new, unseen data. Two models are compared: a regularized linear model (ElasticNet) and a gradient-boosted tree ensemble (XGBoost), with hyperparameter tuning via cross-validation grouped by draft year. Both models explain approximately 7-8% of the variance in the rookie window RAPM on the held-out test set. However, despite the modest R² values, both models demonstrated meaningful ranking ability, outperforming the aggregated pre-draft media consensus board (pooled Spearman ρ of 0.265 for XGBoost and 0.233 for ElasticNet vs 0.050 for consensus). Feature importance analysis shows that both models converge on defensive stops per 100 possessions, free-throw percentage, assists to turnover ratio and draft age as the strongest predictors of early career NBA impact. The findings indicate that while college statistics provide a limited but real signal for NBA projection, interpretable models can still add value as a supplementary tool to traditional scouting by identifying undervalued or overvalued prospects relative to consensus expectations.