Testing the hockey model
An initial season-form baseline was evaluated against past game results. This is a starting point for model development, not validation of the existing historical playoff simulator.
How the test works
Start every team equally. Predict before each game using only earlier days; update after all games on that date. Use pre-January results for warmup and evaluate January onward. Exclude exhibitions, ties in evaluation, and opponents outside the imported D1 directory. No parameter selection on the evaluation set.
Every team starts at an Elo rating of 1500, using a fixed update factor of 20 and a 400-point probability scale. No final-season ranking or postseason poll enters the prediction for an earlier game.
The table below shows probabilities for the first-listed team. Missing home-ice modeling can introduce systematic bias, which the observed-win columns help reveal.
D1 men
554 decided games evaluated, January 1 through 2026-04-11, 2026.
Winner accuracy: 63.5%. Brier score: 0.225, compared with 0.250 for a 50/50 prediction. Lower Brier scores are better.
| Predicted range | Games | Average prediction | Observed wins |
|---|
| 20–30% | 34 | 26.6% | 20.6% |
| 30–40% | 95 | 36.0% | 27.4% |
| 40–50% | 169 | 44.8% | 34.9% |
| 50–60% | 158 | 54.8% | 50.6% |
| 60–70% | 84 | 63.9% | 65.5% |
| 70–80% | 14 | 73.4% | 78.6% |
D1 women
370 decided games evaluated, January 1 through 2026-03-22, 2026.
Winner accuracy: 75.1%. Brier score: 0.191, compared with 0.250 for a 50/50 prediction. Lower Brier scores are better.
| Predicted range | Games | Average prediction | Observed wins |
|---|
| 10–20% | 9 | 17.0% | 0.0% |
| 20–30% | 30 | 26.3% | 6.7% |
| 30–40% | 77 | 35.6% | 19.5% |
| 40–50% | 100 | 44.8% | 31.0% |
| 50–60% | 76 | 54.3% | 59.2% |
| 60–70% | 47 | 64.7% | 76.6% |
| 70–80% | 26 | 73.2% | 92.3% |
| 80–90% | 5 | 82.4% | 100.0% |
What these results cannot tell us
- One season only; no independent season holdout.
- This baseline is not the existing playoff simulator, which uses retrospective NPI and postseason polls. These results do not validate that simulator.
- No home ice, injuries, player changes, automatic bids or selection probabilities.
- Frozen Four percentages in the simulator are conditional on the selected field. They are not the probability of qualifying for the tournament.
The current playoff simulator combines retrospective NPI and poll snapshots. Its percentages assume your selected field and do not model the chance of receiving a tournament bid. They must not be presented as calibrated current-season forecasts.
Full simulator methodology · Download test results · Men’s results source · Women’s results source
