How good are these forecasts?
Every forecast scored on seasons the model never saw while its settings were chosen.
Held-back seasons
| Forecast | Forecasts | Log loss | Favourite won |
|---|---|---|---|
| The model | 2700 | 0.684 | 56% |
| Base rate | 2700 | 0.694 | 51% |
| Recent form | 2700 | 0.691 | 53% |
Against the betting markets
Each market's price before the start, scored on the same finished games as the model. A negative figure means the market was more accurate.
| Market | Since | Priced | Model log loss | Market log loss | Model vs market |
|---|---|---|---|---|---|
| Bookmakers (Pinnacle, DraftKings, bet365) | 27 Jan 2026 | 763 | 0.692 | 0.690 | −0.3% |
| Polymarket | 25 Nov 2025 | 1132 | 0.689 | 0.687 | −0.3% |
| Kalshi | 16 Jan 2026 | 866 | 0.690 | 0.689 | −0.2% |
The AHL is the NHL's top development league: rosters change all season as players are called up and sent down, which makes it noisier to forecast than the NHL. The model keeps more of each rating's change from game to game for that reason.
The playoff odds follow the league's format: five, six or seven teams from each division (by its size), a best-of-three first round for the lower seeds, best-of-five division semi-finals and finals, then best-of-seven conference finals and Calder Cup Finals.
NHL call-ups and assignments are not in the model; a team that loses its best players to its parent club mid-season is only caught as its results change.
Goaltending is not in the rating.
Questions and answers
How accurate are PodiumOracle's AHL predictions?
On 2,700 games from held-back seasons (2024–25 to 2026–27), which the model never saw while it was tuned, the side it favoured won 56% of the time. Its log loss was 0.684, against 0.694 for the base rate and 0.691 for recent form. Against the betting markets on the same finished games, its log loss was 0.692 against the bookmakers' 0.690 over 763 games, 0.689 against Polymarket's 0.687 over 1,132 games and 0.690 against Kalshi's 0.689 over 866 games (lower is better).