NHL / EVIDENCE BEFORE CONFIDENCE
Show the work.
A forecast should earn your trust on games it has not seen. This model learns on earlier seasons and is evaluated on the held-out 2025–26 regular season.
Against simple baselines
| Method | Accuracy | Brier | Log loss |
|---|---|---|---|
| Sports Desk | 54.1% | 0.248 | 0.690 |
| Elo + home venue | 53.3% | 0.250 | 0.694 |
| Training home-win rate | 52.2% | 0.251 | 0.694 |
Winner includes overtime and shootouts. Accuracy alone rewards picking favourites; Brier and log loss also test whether the probabilities are sensible.
Calibration by home-win probability
| Band | Games | Predicted | Actual |
|---|---|---|---|
| 20–40% | 29 | 36.7% | 24.1% |
| 40–60% | 964 | 52.3% | 51.6% |
| 60–80% | 319 | 64.5% | 56.7% |
Within each band, average predicted home-win chance is compared with the observed home-win rate. Small groups carry more uncertainty.
THE CONTRACT
What goes into a forecast
The regularised logistic model uses home venue, Elo difference, last-30 and last-10 scoring margins, days off, back-to-back status, venue-city travel distance, eastward time-zone changes and games in seven days. All features are computed before the game outcome enters the history.
At least 20 prior regular-season games per team warm the ratings. Eligible games through 2024–25 train the evaluation model. The 2025–26 test season does not choose its coefficients. NBA Cup finals are explicitly excluded even when the publisher labels them regular season. Future forecasts are refitted on completed regular-season games. Pregame forecasts are saved before the scheduled start and frozen once play begins.
Current injury reports, projected goalies and market prices are displayed as context, not model inputs. Historical point-in-time availability and player-impact estimates are needed before they can be validated. Preseason and postseason forecasts are withheld because their rotations and incentives differ.
Travel uses approximate venue-city centroids and scheduled games, not flight routes or actual departure times. Missing travel comparisons are treated as neutral model contributions and shown as unknown in the interface.
This is an initial historical holdout evaluation, not a prospective track record or proof of a betting advantage. More data points only help when they improve out-of-sample performance.