Method
What each number means, how it is checked, and what we are not allowed to claim.
1. Data
- Source: the public NHL statistics services. No paid source.
- Regular season only. Playoffs are a different regime and are not mixed in.
- A game is kept only when three routes agree: official statistics, game summary, play-by-play. Disagreements are set aside with their reason.
- Team goals are counted in regulation time, without empty-net goals. Shootout goals are not goals of the game.
- The win is the final win, overtime and shootout included.
- Home ice is measured without neutral-site games. Utah counts as one team across its name change, with no history inherited from Arizona.
Verified games on hand: 2,870
2. Reliability of a ranking
Over n games, the spread of team averages mixes the true gap between teams (τ²) and chance (σ²/n). Reliability is τ² divided by the observed spread.
A ranking is shown only at a reliability of at least 0.70, measured on all teams.
Goal differential per game, season 2026-27
| Games | Teams | Reliability | Ranking |
|---|---|---|---|
| 1 | 32 | 0.42 | no |
| 5 | 32 | 0.38 | no |
| 10 | 32 | 0.39 | no |
Goal differential per game, season 2025-26
| Games | Teams | Reliability | Ranking |
|---|---|---|---|
| 1 | 32 | -0.02 | no |
| 5 | 32 | -0.11 | no |
| 10 | 32 | -0.34 | no |
| 20 | 32 | 0.00 | no |
| 41 | 32 | 0.38 | no |
| 82 | 32 | 0.72 | allowed |
Goal differential per game, season 2024-25
| Games | Teams | Reliability | Ranking |
|---|---|---|---|
| 1 | 32 | 0.29 | no |
| 5 | 32 | 0.17 | no |
| 10 | 32 | 0.45 | no |
| 20 | 32 | 0.51 | no |
| 41 | 32 | 0.48 | no |
| 82 | 32 | 0.72 | allowed |
With about thirty teams, each value carries an uncertainty of roughly ±0.1. Negative values mean the true gap is small compared with chance.
3. Frozen back-test
Marker: final win of the home team. Parameters are frozen on season 2024-25, then judged once on season 2025-26, never seen by the model.
- Judged only from 1,200 predictions over 20 weeks.
- Publishable only if its Brier score beats the reference (the home-win frequency of the calibration season) AND no reliability bin is at fault.
- A bin (at least 50 predictions) is at fault when its gap exceeds both 5 points and 2.5 standard deviations of chance.
- A losing model is retired, not patched. Correcting data does not reopen frozen parameters.
| Variant | Predictions | Weeks | Brier | Reference | Coin flip | Bins at fault | Verdict |
|---|---|---|---|---|---|---|---|
| A — actual starter (known afterwards) | 1,312 | 26 | 0.2494 | 0.2513 | 0.2500 | 0 | publishable |
| B — estimated starter | 1,312 | 26 | 0.2516 | 0.2513 | 0.2500 | 1 | retired |
Verification runs against this season so far: 1. Only the first one counts.
Reading: a Brier score only means something next to its reference and to a coin flip (0.25). Beating the reference by a margin of the same size as chance does not demonstrate predictive power.
This site shows no win probability. None will be shown before a live log has been judged on at least 2,000 predictions, with the starting goaltender confirmed before the game.