Model Analytics
XGBoost classifier trained on ATP matches from 2016–2024, tested on 2025. Isotonic calibration ensures predicted probabilities reflect real-world win rates. Feature importances are averaged across five calibration folds.
Feature Importance
Top 12 features by XGBoost gain — what the model weighs most heavily
Loading…
Calibration Curve
Predicted win probability vs actual win rate — dots near the dashed line mean honest predictions
Loading…
Rolling Accuracy (30-match window)
Model accuracy over time — each point is the last 30 settled matches at that date
Loading…
Accuracy by Surface
How the model performs across different court types
Loading…