NBA Win Probability Engine 🏀

Real-time in-game win probability with live SHAP explainability — deployed FastAPI + Streamlit

Results

Metric	Value
Brier Score	0.089
Log Loss	0.213
AUC-ROC	0.941
Calibration Error	0.012

What It Does

• Predicts win probability at every play-by-play event in real time
• Explains each prediction with SHAP values (which factors matter most)
• Exposes a /predict REST endpoint and live Streamlit dashboard
• Trained on 500K+ plays across 5 NBA seasons

Architecture

Raw play-by-play (nba_api)
    → Feature engineering (score diff, pace, possession, time)
    → XGBoost + isotonic calibration
    → SHAP TreeExplainer
    → FastAPI /predict endpoint
    → Streamlit live dashboard

Tech Stack

Python XGBoost SHAP FastAPI Streamlit scikit-learn nba_api Docker Plotly

Quick Start

pip install -r requirements.txt
python data/pull_pbp.py          # Pull NBA play-by-play data
python model/train.py            # Train + calibrate model
uvicorn api.main:app --reload    # Start API
streamlit run app/dashboard.py  # Launch dashboard

API Usage

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"score_diff": 5, "time_remaining": 420, "possession": 1, "pace": 98.4}'
# → {"win_prob": 0.714, "shap_values": {...}}