About the model
How the win probabilities and season projections are made, and how well they hold up.
Method
Each game gets a home-win probability from a gradient-boosted classifier over form features — rolling goals, expected goals, shots, and high-danger chances over 5/10/20-game windows (home-minus-away differentials), rest days, back-to-backs, and an Elo rating difference — with isotonic calibration so the probabilities mean what they say. A separate model estimates the chance a game passes regulation, which decides the loser point. Season projections Monte-Carlo the remaining schedule 10,000 times with those probabilities.
Currently shipping: GBM (the boosted model beat Elo on held-out seasons).
Walk-forward backtest
Trained on all seasons before the test season, scored on the test season. Lower log loss is better; the model must beat both the Elo baseline and the constant home-win rate.
| Test season | Games | Log loss (model) | Log loss (Elo) | Log loss (home rate) | Brier | Accuracy |
|---|---|---|---|---|---|---|
| 2023–24 | 1259 | 0.6702 | 0.6623 | 0.6903 | 0.2388 | 58.9% |
| 2024–25 | 1258 | 0.6644 | 0.6667 | 0.6872 | 0.2360 | 60.0% |
| 2025–26 | 1259 | 0.6894 | 0.6907 | 0.6922 | 0.2480 | 54.2% |
For scale: home teams win about 54% of NHL games, and strong public models land near 58–62% accuracy. Calibration matters more than accuracy here — the season simulator consumes the probabilities directly.
Prediction log
Every published win probability is logged before the game and scored against the result once it's played. A rolling calibration report appears here once the coming season provides live predictions.