Our 2026-27 NHL Projection, and Exactly How Much to Discount It
2026-09-29 · projections · predictions · analytics
Carolina 111.6 points, Vancouver 79.1 - and luck alone gives the average team an 80% range 21.8 points wide, two-thirds of the league's entire spread. Our measured error says the honest range is closer to 32. The projection beats the honest baseline, and here is exactly how much of the table you should refuse to read as a ranking.
Figures computed 2026-09-27, against the models as they stood that day. We change them — see the methodology page for where they are now.
- Carolina 111.6 points, Vancouver 79.1 — a 32.5-point spread across the league.
- Game-to-game luck alone gives the average team an 80% range 21.8 points wide — two-thirds of the whole league's spread. All 32 teams' ranges overlap the 16th-placed team's.
- The real uncertainty is about half as wide again. Our measured error implies an 80% range closer to 32 points, because the simulation cannot see goaltending, injuries or a model that is simply wrong about a team.
- 14 of 32 teams have playoff odds between 30% and 70%, and 21 of 31 adjacent gaps in the table are under a single point.
- The model still beats the honest baseline — RMSE 12.50 against 15.12 for simply carrying last season's points forward, over 296 team-seasons. A real edge, and a modest one.
Our model thinks Carolina will win the most points in 2026-27, and Vancouver the fewest. That is the sentence a preseason projection is supposed to produce, and if you only wanted the ranking you can stop reading and go look at the live table.
The more useful sentence is the one almost nobody prints next to their projection: we have measured how wrong this is likely to be, and the answer is wrong enough that most of the table should not be read as a ranking at all.
Both halves of that are our own published numbers. Here is the projection, and here is exactly how much to discount it.
The projection
Ten thousand simulated seasons, 84 games each, built from team Elo blended with projected roster WAR — every player projected forward individually and summed onto the depth chart we expect each team to dress. Snapshot of 27 September 2026, two days before the opener.
| # | Team | Points | 80% range | Playoffs |
|---|---|---|---|---|
| 1 | Carolina | 111.6 | 101–122 | 98% |
| 2 | Colorado | 106.9 | 96–118 | 96% |
| 3 | Vegas | 101.2 | 91–112 | 89% |
| 4 | Washington | 101.0 | 90–112 | 78% |
| 5 | Tampa Bay | 99.7 | 89–111 | 73% |
| 6 | Montréal | 99.5 | 89–110 | 73% |
| 7 | Ottawa | 99.1 | 88–110 | 71% |
| 8 | Edmonton | 97.6 | 87–109 | 78% |
| …ranks 9 through 24 span 8.7 points… | ||||
| 25 | Philadelphia | 87.8 | 77–99 | 22% |
| 26 | NY Rangers | 85.6 | 75–97 | 15% |
| 27 | Boston | 85.6 | 75–96 | 15% |
| 28 | Toronto | 84.1 | 73–95 | 12% |
| 29 | Seattle | 82.8 | 72–94 | 17% |
| 30 | Calgary | 82.1 | 71–93 | 15% |
| 31 | Chicago | 79.6 | 69–90 | 7% |
| 32 | Vancouver | 79.1 | 68–90 | 8% |
The live version — which will have moved by the time you read this — is on the standings projection page.
Now the part that matters
A projection without an error bar is a horoscope with arithmetic. Ours has one, because we tested it the way we test everything: leave-one-season-out cross-validation across 296 team-seasons, scored against the baseline that actually threatens it.
| Method | RMSE, points |
|---|---|
| Elo + roster WAR blend (what we publish) | 12.50 |
| Projected roster WAR alone | 12.87 |
| Team Elo alone | 12.96 |
| Carry last season's points forward | 15.12 |
The baseline choice is the whole game here. Scoring a projection against the league average would flatter it enormously, because in a 32-team league with a hard cap the average and the persistence forecast are nearly the same number. The honest baseline is persistence: assume every team finishes where it finished last year, and do nothing else.
Against that, we are better by about 2.6 points of RMSE. That is a real edge. It is also a small one — a typical miss of twelve and a half points is most of the gap between a playoff team and a lottery team.
The band on the chart is the optimistic one
Here is the thing the chart does not tell you on its own. Each of the ten thousand simulated seasons plays out the schedule with every team's strength fixed — the dice are the individual games, not the teams. So the 80% range on each row is the spread you would get from bounces alone, if we had every team's true quality exactly right.
We do not. An RMSE of 12.50 corresponds to an 80% range roughly 32 points wide, against the 21.8 the simulation draws. Put the other way, game-to-game luck accounts for a bit under half of our real error. The rest is everything the simulation holds still: a goaltender having the year of his life, a top pair losing half a season, and plain misjudgement of how good a team is.
What even the optimistic band does to a table
The league spans 32.5 points from Carolina to Vancouver, and luck alone gives the typical team a range of 21.8. Those two numbers together are why the chart at the top of this post looks the way it does:
- Every team's range overlaps the 16th-placed team's. All 32.
- Only ten teams sit entirely below Carolina's range — Winnipeg, San Jose, Philadelphia, the Rangers, Boston, Toronto, Seattle, Calgary, Chicago and Vancouver. The other 21 have a plausible path to the league's best record, before you even widen the ranges to their honest size.
- Twenty-one of the thirty-one adjacent gaps are under a point. Ranks 9 through 17 — nine teams — are separated by 4.8 points in total.
- Fourteen teams are coin flips, with playoff odds between 30% and 70%.
So read it as three tiers, not as a list of 32. The model is genuinely confident that Carolina and Colorado are good and that Vancouver and Chicago are not. It is mildly confident about another handful. And about roughly two dozen teams it is telling you, in the only language it has, that it does not know.
A ranking hides that. Anyone can sort a column; the ordering of rank 12 from rank 17 here is noise dressed as knowledge.
Where this model is weakest, in our own words
Two published results from our own validation work bear directly on this projection, and both of them cut against it.
Summing players to a team is our worst axis. In the four-test validation series we measured team-summed player WAR against next season's standings on HockeyStats' published protocol and got an R² of 0.208 — against 0.269 for simply reusing last season's standings. On that test we lost to persistence outright.
That is not the same exercise as this projection, and the difference is worth stating precisely rather than waving at. That test summed last season's player WAR. This projection projects each player forward and sums onto a projected roster, which accounts for aging, arrivals and departures — and that version does beat persistence, 12.87 to 15.12. The roster work is doing real labour. But the weak joint is in the same place, and if this projection is badly wrong about a team, the player-summing half is where I would look first.
And we cannot project goalies. We have published that our goalie projection ranks goaltenders worse than a three-year weighted average does — R² 0.024 against 0.064 — because goalie value repeats from one season to the next at a correlation of about 0.209, roughly 4% of the variance. Nobody has solved this. But a team's season can turn on one goaltender having a year, and it is the largest single thing the simulated range leaves out.
What would make us wrong
Two things, in descending order of how much they worry me. Goaltending, for the reason above. And injuries — the projection dresses the depth chart we expect, and a team that loses eighty games from its top pair is a different team than the one we simulated.
One thing that sounds like it should be on that list is not. When we re-ran the backtest using the rosters teams actually iced, rather than the ones we expected before the season, the error fell only from 12.50 to 12.44. Knowing every trade, waiver claim and surprise rookie in advance would barely help. Roster churn is not what limits this model; judging how good the players are is.
None of this is a reason to throw the projection out. It is the reason to read the chart's ranges as a floor on the uncertainty, not as the uncertainty.
How to use it
Read the tiers, not the ranks. Treat any gap under a few points as no gap at all. When the question is who gets in, use the playoff odds rather than the point totals, because they fold in the schedule and the divisional format: Edmonton projects fewer points than Tampa Bay, 97.6 against 99.7, and is more likely to make the playoffs, 78% against 73%, because one of them has to finish ahead of Montréal and Ottawa in the Atlantic and the other does not. Just remember the odds share the ranges' blind spot — they are luck-only too, so anything between about 20% and 80% is closer to a coin flip than it looks. And if our number disagrees violently with your read on a team — we have Toronto 28th and Montréal 6th — the honest answer is that on a range that wide, both of us are inside the noise.
We will publish the same table in April with the actual results beside it, because a projection nobody scores afterwards is marketing. If the middle two dozen come out scrambled, that will be the model working as specified, not failing.
Projection snapshot of 27 September 2026: 10,000 simulations, 84-game season, team Elo blended with projected roster WAR on expected depth-chart lineups. Each simulation holds team strength fixed and draws game outcomes, so the 80% ranges and playoff odds reflect game-to-game luck only. The live projection refreshes daily and will differ from the figures here — this post is a dated record of the preseason state, not a live table; for current numbers see the projection page. Validation figures are leave-one-season-out cross-validation over 296 team-seasons, scored as RMSE in points against a persistence baseline; the ~32-point range is 2 × 1.28 × RMSE, assuming roughly normal errors. The 0.208 and 0.269 R² figures are from the WAR validation series, computed on HockeyStats' published pooled protocol against points-per-82. Goalie projection figures re-derived 13 September 2026.
Frequently Asked Questions
Who is projected to win the NHL in 2026-27?
As of 27 September 2026 our model projects Carolina first at 111.6 points, with Colorado second at 106.9 and Vegas third at 101.2. Vancouver projects last at 79.1. Those are expected points across 10,000 simulated 84-game seasons, and the live projection refreshes daily.
How accurate are NHL preseason standings projections?
Ours is accurate to an RMSE of about 12.5 points, measured by leave-one-season-out cross-validation over 296 team-seasons. The baseline that matters is persistence - simply carrying last season's points forward - which scores 15.12. That RMSE implies an honest 80% range about 32 points wide for any one team.
Why do so many NHL teams have similar projected point totals?
Because the league is tightly packed and the uncertainty is large. The spread from first to last is 32.5 points, while game-to-game luck alone gives a typical team an 80% range 21.8 points wide. Every team's range overlaps the 16th-placed team's, and 21 of the 31 gaps between adjacent teams are under a point.
Should you trust playoff odds or projected points?
For who makes the playoffs, the odds, because they account for the schedule and the divisional format: Edmonton projects fewer points than Tampa Bay (97.6 to 99.7) yet is more likely to make the playoffs (78% to 73%). But both come from simulations that hold team strength fixed, so they reflect game-to-game luck only and are more confident than the model's measured error justifies.
More from the Notebook
The Model That Earns Its Keep
Is our projection actually better than a simple average? We ran the honest test - forecasting one player's WAR next season, walk-forward across thirteen seasons - and beat the baseline that is supposed to be unbeatable. Then we ran it on goalies, where we lose.
Talent vs Production: Why the Box Score Misleads
Two wingers score 20 goals; only one of them will do it again. The gap between what a player produced and the process underneath it is the most useful idea in hockey analytics - and it is the split our expected-goals and finishing models are built to make.
How Good Is Our WAR Model? We Ran Four Tests and Lost Two
We tested our WAR model against HockeyStats and Evolving Hockey on the four questions an honest player-value model has to answer, matched to each benchmark's published protocol. Summed to a team it loses both tests - it forecasts next season worse than simply reusing last season's standings. Measured per player it wins both, repeating at 0.77 against 0.62 and 0.46.