2026-27 season to date · 0 games scored
Every game prediction on Tonight is locked in before puck drop and scored here once the final is in.
The same model replayed on history, walk-forward: each season is predicted by a model fitted only on the seasons before it, so no game is scored by a model that has seen it. Over 16,312 games it scored 0.6648 log-loss and picked 59.5% of winners, against 0.6682 for player Elo alone and 0.6885 for guessing the home team every time.
Running total over one full season, game by game. Below zero means guessing the home team was ahead. The model was behind the naive guess for 650 of 1178 games and last below zero at game 807; it finished +6.2, against +3.5 for Elo alone, and led Elo after 95% of games. A season is noisy: expect weeks where the model trails.
Games sorted into ten equal groups by the model’s home-win probability. Dots on the diagonal mean a 65% call wins about 65% of the time.
Log-loss by season (lower is better). The bar is the model’s edge over player Elo alone: blue to the right when the model was better, red to the left when Elo was.
| Season | Games | Model | Elo alone | Home guess | Picks correct | Edge vs Elo (log-loss) |
|---|---|---|---|---|---|---|
| 2012-13 | 720 | 0.6728 | 0.6748 | 0.6856 | 57.9% | +0.0020 |
| 2013-14 | 1,227 | 0.6715 | 0.6771 | 0.6905 | 57.6% | +0.0056 |
| 2014-15 | 1,224 | 0.6627 | 0.6632 | 0.6898 | 59.1% | +0.0005 |
| 2015-16 | 1,230 | 0.6791 | 0.6812 | 0.6918 | 57.5% | +0.0021 |
| 2016-17 | 1,229 | 0.6672 | 0.6668 | 0.6868 | 59.4% | −0.0004 |
| 2017-18 | 1,269 | 0.6627 | 0.6643 | 0.6859 | 61.5% | +0.0015 |
| 2018-19 | 1,270 | 0.6727 | 0.6771 | 0.6906 | 59.4% | +0.0044 |
| 2019-20 | 1,080 | 0.6759 | 0.6760 | 0.6914 | 56.5% | +0.0001 |
| 2020-21 | 865 | 0.6527 | 0.6585 | 0.6912 | 61.3% | +0.0058 |
| 2021-22 | 1,241 | 0.6428 | 0.6491 | 0.6864 | 63.5% | +0.0063 |
| 2022-23 | 1,266 | 0.6527 | 0.6579 | 0.6904 | 62.0% | +0.0053 |
| 2023-24 | 1,259 | 0.6548 | 0.6600 | 0.6862 | 60.4% | +0.0052 |
| 2024-25 | 1,254 | 0.6568 | 0.6641 | 0.6821 | 61.0% | +0.0073 |
| 2025-26 | 1,178 | 0.6852 | 0.6874 | 0.6904 | 54.4% | +0.0023 |
Log-loss charges each prediction −ln(probability given to what happened): calling a winner at 70% costs 0.36, calling the loser at 70% costs 1.20. It rewards being confident only when it is warranted. A coin flip scores 0.693 on every game, and the best NHL models live around 0.66: hockey results are close to random, so small differences in the third decimal are real skill. Picks correct counts the team given 50% or more; ties in the NHL are settled in overtime or a shootout, and the final score decides.
The four models: the game model behind Tonight (player Elo, recent net goals of the lineup, goalies, rest); player Elo alone, the same Elo ratings without the lineup terms; the standings sim’s per-game probability from projected rosters (as in Projected Standings); and the home-team guess, the home win rate of past seasons (54.7%) for every game. Every live prediction is written to the log the morning of the game and never edited after puck drop.