NHL Model Performance
Walk-forward replay of the 2022-23 through 2025-26 regular seasons with the same windows and goal map, kept separate from live forecasts. Log loss and Brier score judge the home win probability (a coin flip scores 0.693 and 0.250); total MAE and bias judge the model total in goals. The comparison with the closing line follows.
| Season | Games | Log loss | Brier | Accuracy | Home wins | Total MAE | Total bias |
|---|---|---|---|---|---|---|---|
| 2025-26 | 1312 | 0.7085 | 0.2563 | 54.0% | 52.2% | 1.90 | -0.23 |
| 2024-25 | 1289 | 0.6812 | 0.2430 | 58.3% | 56.3% | 1.93 | -0.16 |
| 2023-24 | 1312 | 0.6831 | 0.2443 | 58.2% | 54.1% | 1.93 | +0.01 |
| 2022-23 | 1312 | 0.6734 | 0.2402 | 57.7% | 52.4% | 1.86 | 0.00 |
| All | 5225 | 0.6866 | 0.2460 | 57.0% | 53.7% | 1.90 | -0.10 |
Against the closing line
Closing moneylines and totals come from ESPN's public odds data. The moneyline is converted to a home win probability with the vig removed. Model and market are scored on the same games: only those with a close, so these figures differ from the full-sample ones above. Lower is better.
| Season | Games | Model log loss | Market log loss | Gap | Total games | Model total MAE | Market total MAE |
|---|---|---|---|---|---|---|---|
| 2025-26 | 1312 | 0.7085 | 0.6817 | +0.0268 | 1309 | 1.90 | 1.85 |
| 2024-25 | 1289 | 0.6812 | 0.6579 | +0.0233 | 1289 | 1.93 | 1.90 |
| 2023-24 | 1312 | 0.6831 | 0.6552 | +0.0279 | 1311 | 1.93 | 1.87 |
| 2022-23 | 1221 | 0.6707 | 0.6495 | +0.0213 | 0 | – | – |
| All | 5134 | 0.6862 | 0.6613 | +0.0249 | 3909 | 1.92 | 1.87 |
Calibration
Observed home win rate against the predicted probability in ten bins. Points above the diagonal mean the model was too pessimistic about the home side, below it too confident.