From rookie to contract extension: Modeling NBA player career trajectories with machine learning

Abstract

This paper investigates machine learning-based prediction of players’ long-term career trajectories at different career stages. Specifically, we predict whether a player is likely to achieve Elite Player, High-Impact Starter or Rotation Player status based on data from their first one, three, and five NBA seasons. The target classes are expert-defined, subjective categories intended to capture meaningful differences in players’ career success and on-court impact. We collected, analyzed, and preprocessed a dataset of 942 players, described by 60 attributes representing their performance metrics and physical characteristics. We compare the classification performance of Multilayer Perceptron, Random Forest, and XGBoost, along with Logistic Regression as a benchmark. Furthermore, we conducted two experiments: one that takes into account the effect of NBA era player was active and another that neglects it. Finally, we identify the most important attributes contributing to classification outcome. Random Forest achieves the best performance after the rookie season (ROCAUC = 0.671 ± 0.039), indicating that career outcomes can be predicted only to a limited extent from rookie-season data. Performance improves substantially after three seasons (0.793 ± 0.025) and five seasons (0.859 ± 0.023), suggesting that additional performance history provides much stronger signals of long-term career success. PER is identified as the most important performance indicator.