Back to projects

Machine learning

NBA Win-Probability Prediction

End-to-end ML system that predicts NBA game win probabilities, served as an interactive web app with game-by-game backtest report cards.


Highlights

  • Engineered a leakage-safe pipeline validated with walk-forward testing over 22,800+ held-out games across 19 seasons.
  • Prioritized proper scoring rules (log loss, Brier) and probability calibration over raw accuracy.
  • Best model (logistic regression on Elo + rolling form/rest features) beat a strong Elo baseline while staying well-calibrated.
  • Benchmarked against the betting market and honestly reported no edge after the house margin after debugging a data error.

Stack

Pythonscikit-learnXGBoostStreamlit

Project Links