TeMPO: A framework for longitudinal mapping of NBA player profiles

Abstract Basketball analytics increasingly relies on statistical modeling and machine learning to evaluate player performance and strategic dynamics. However, most existing approaches analyze player statistics in a static manner, overlooking the temporal dependencies that govern role persistence and transitions across seasons. In this work, we propose TeMPO ( Te mporal M apping of P layers’ O utcomes), a reproducible framework designed to model the structural composition and longitudinal evolution of player performance profiles. The proposed framework integrates correlation-based feature selection, seasonal statistical moment extraction, feature denoising, non-linear manifold learning via Uniform Manifold Approximation and Projection (UMAP), K-Means clustering, and Markovian transition modeling. TeMPO facilitates the identification of interpretable role archetypes and the probabilistic pathways through which players migrate between profiles over time. By combining unsupervised representation learning with stochastic temporal modeling, it offers a scalable methodology for analyzing longitudinal behavioral patterns in structured sports data. We apply the framework to a large-scale dataset containing 31,898 NBA regular-season games between 1997 and 2024, comprising 659,652 player-game observations and 1,709 unique players. The results provide evidence of persistent player archetypes, transitional roles that act as bridges between profiles, and dominant migration pathways across seasons. The analysis also provides a quantitative characterization of structural shifts, including evidence consistent with a progressive consolidation of perimeter-oriented offensive roles. These findings suggest that longitudinal modeling can help uncover the evolving organization of player roles in professional basketball.