BioPerformance

Fixed formulas and rule-based engines fail to capture the non-linear, highly individualized nature of athletic fatigue and recovery.

Run an R&D effort that builds a machine learning pipeline predicting next-day athlete wellness with auditable, explainable predictions.

Trained four separate XGBoost regressors, one per wellness item, each with a group model (cold-start fallback) and an individual athlete model that activates after 14+ days of history. Engineered 52 temporal features per athlete: t0/t1/t2 lags, 7d and 28d rolling statistics, Z-scores, sleep coefficient of variation, Acute:Chronic Workload Ratio, and days since last rest, validated with TimeSeriesSplit to prevent leakage. Added SHAP TreeExplainer attributions with beeswarm and waterfall plots so every prediction is auditable. Validated in two stages: synthetic data with known ground-truth patterns, then real physiology from the PMData dataset (16 athletes, 1,747 observation days).