Research
The questions this project has put to its NHL data and to the betting market, what each study found, and when. Each links to the page with the method and the evidence.
This season's win chances, graded
After 233 games, the model's win chances have been better than a home-ice-only guess.
Every finished 2026-27 regular-season game with an adjusted win chance recorded before puck drop, graded with that chance, through November 11, 2026. Updated each morning as games are played.
The win chance includes an adjustment for a club playing the second night of a back-to-back. The plain team rating it adjusts is still recorded every night and graded separately: after 298 games, its log loss is 0.6622 against home ice’s 0.6874 (better).
Average log loss per game, lower is better:
model 0.6567 · home ice 0.6879 · coin flip 0.6932
Model minus home ice, where below zero favours the model:
−0.0312, 95% interval −0.0472 to −0.0159
Against a coin flip, the model has been better:
−0.0365, 95% interval −0.0558 to −0.0166
The Brier score, the average squared miss, is shown too, so the reading does not rest on one measure. Against home ice, the model has been better:
model 0.2321 · home ice 0.2474 · coin flip 0.2500 · difference −0.0152, 95% interval −0.0231 to −0.0077
How it is graded — only win chances recorded before puck drop count
Each finished regular-season game is graded with the last win chance this site recorded before its scheduled puck drop. A chance written after a game started is never used. 65 finished games had no adjusted chance recorded before puck drop and are not graded here, including any played before the adjustment began; the plain rating’s chances for those are graded in the record above.
The home-ice guess gives the home team the same chance in every game: 53.7%, the share of 6,560 regular-season games the home team won from 2021-22 through 2025-26. It uses no game from 2026-27. The coin flip gives every game 50%.
Log loss charges a forecast for how surprised it was by each result, so a confident miss costs far more than a cautious one. The comparison is made game by game on the same games, and the 95% interval comes from resampling those games. An interval that includes zero means the model and the guess cannot yet be told apart, and the page says so rather than calling either one better.
| Week ending | Games | Model | Home ice | Difference | 95% interval | Against home ice |
|---|---|---|---|---|---|---|
| Oct 11 | 21 | 0.6595 | 0.6995 | −0.0400 | −0.0943 to +0.0221 | not distinguishable |
| Oct 18 | 71 | 0.6561 | 0.6865 | −0.0304 | −0.0576 to −0.0026 | better |
| Oct 25 | 119 | 0.6609 | 0.6815 | −0.0206 | −0.0416 to +0.0018 | not distinguishable |
| Nov 1 | 166 | 0.6606 | 0.6852 | −0.0245 | −0.0421 to −0.0069 | better |
| Nov 8 | 217 | 0.6540 | 0.6860 | −0.0320 | −0.0477 to −0.0155 | better |
| Nov 15 | 233 | 0.6567 | 0.6879 | −0.0312 | −0.0472 to −0.0159 | better |
The model, against the results
- Who beat expectation, and how much of it should persist?
League-wide finishing and goaltending, each season measured against its own league average.
- Does finishing persist?
Yes, partly — it beat both simpler rules: so far a backtest on 2025-26, 483 projections graded. Nothing logged has come due yet.
- What standings do the model's game odds add up to?
In 2025-26 the same method missed each club’s final points by 10.3 on average, against 10.1 for a flat guess that put every club at the previous season’s league average. None of the three projections, this one or either outside one, was measurably better than the flat guess.
Every regular-season game's win probability, logged before the season and added up to standings points and playoff odds. It is graded when the season ends.
2026-27 projection, logged September 23, 2026
What the model sees · 2026-27 · 32 clubs · rank gap in spots
The Los Angeles Kings are 30th in points. The model ranks them 12th.
The widest gap of the 32 clubs. The chart shows every club rated 7 or more spots from its place in points; this one is LAK.
Drawn: the 13 clubs 7+ spots apart, of 32.
How the gap is measured — rank in points minus rank by xG-Elo; positive means the model ranks a club higher than its points do
Each bar is one club: its rank by xG-Elo, the model’s rating built from the chances a club makes and allows, against its rank in the standings. #1 is the top of each ranking, so the Los Angeles Kings, #30 in points and #12 by xG-Elo, have a gap of 30 − 12 = +18.
Green bars are clubs the model ranks higher than their points; red bars, lower. The words carry the direction too, so the colour is never the only sign.
How the playoff odds are worked out
Every morning the rest of the regular season is played out 40,000 times, starting from the standings as they are: a game already played counts as it ended, and only the games still to play are simulated. A club’s playoff chance is the share of those seasons in which it makes the playoffs, and its projected points are its average final total. The latest run was made on Thu, Nov 12.
Each game’s win chance comes from the xG-Elo rating of both clubs, with home ice, the same model as the preseason projection, adjusted, as the game pages are, when a club plays the second night of a back-to-back. A game goes past regulation 22% of the time, the average of the last 5 seasons, and the loser of such a game takes a point.
What the preseason projection leaves out is that a rating is itself uncertain. So in every simulated season each club’s rating moves by its own random amount, drawn once for the whole season, with a spread of 41.3 rating points. That spread is measured, not chosen: it is the size at which the same method, run over 2025-26 from the ratings it held before that season, misses the final standings by as much as it really did. Without it the odds come out too sure of themselves. It was measured on the rating without the back-to-back adjustment.
The format is the NHL’s: the top three clubs in each division, then the two best of the rest in each conference. Clubs level on points are separated by regulation wins, then by wins, then at random; the league’s later tiebreakers (regulation and overtime wins, head-to-head points, goal difference) are not simulated, because the simulation plays out results, not scores.
The daily change is each club’s playoff chance against the previous morning’s, in percentage points. Each morning uses the same random draws for the games still to play, so a chance moves because of the results in between, not because the simulation was run again. The preseason projection is kept as it was logged before the season, to be graded when the season ends. These are projections of the standings, not betting advice.
The model, against the posted lines
- What does the model project for each skater, against the line a book posted?
The model has not beaten the Pinnacle closing line. How this was measured
The sheet: every game's projected distributions beside the posted lines, one date at a time.
Studies
- How well priced is the closing line?
On shots on goal the Pinnacle close was measurably mis-priced in two consecutive seasons, by about three points, and showed no detectable average mispricing in the third. Betting that mispricing made no money in any season measured.
The Pinnacle closing line on NHL player props, set against what the players went on to do. It needs no model.
Figures updated September 29, 2026
- Does projection ordering hold up?
The order carries information well beyond the top-ranked players, though how much varies by statistic, season and region.
Every pair of skaters ranked close together by projected rate, compared within seven rank regions, by statistic and season.
Figures updated August 19, 2026
Before reading any of it
How to read OddPuck: what the figures are, where they come from, and how the model was tested against the betting market.