Synthetic Data Generation for NCAA Basketball Performance Analytics

Abstract

Access to athlete monitoring data in NCAA Division 1 basketball is limited by privacy and proprietary restrictions, despite the growing use of wearable analytics for training, recovery, and performance decision-making. This study introduces a privacy-preserving synthetic data generation framework explicitly designed for collegiate basketball performance analytics. The real dataset consists of wearable-derived workload measurements collected from collegiate women's basketball athletes across practices, training sessions, and games. Key metrics include Total Athlete g-Load, limb-specific load, impact, push-off, intensity, and asymmetry measurements. The proposed pipeline preprocesses and anonymizes raw data, imputes missing values using Multiple Imputation by Chained Equations (MICE), and generates synthetic data via Gaussian resampling based on athlete-specific statistical profiles. Real and synthetic datasets are then used for model development, including a downstream multivariate regression model with regularization and feature-importance analysis. We evaluated the following train–test protocols for assessing synthetic data fidelity: Train Real–Test Real (TRTR, gold-standard baseline), train synthetic–test real (TSTR), train real–test synthetic (TRTS), and train synthetic–test synthetic (TSTS). Predictive results for TSTR (train R2 0.99 / test R2 0.76) closely follow those for TRTR (train R2 0.91 / test R2 0.91), showcasing good generalization ability. The TSTS train and test R2 values are 0.91, indicating good internal consistency of the dataset. Results demonstrate strong statistical fidelity and preservation of meaningful workload relationships, with correlations among load, impact, and push-off metrics consistent with biomechanical expectations. Overall, the framework provides a valuable approach for enabling reproducible research and predictive model development in NCAA basketball while addressing data scarcity and athlete privacy concerns. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.