Explaining Our Power Rankings

2026-10-15 · power-rankings · Elo · methodology · validation

Readers asked why our power rankings still count last season eight games in. We replayed 13 seasons to find out what predicts the rest of a season: the record so far is the weakest option, dropping last season makes the rankings worse, and this season's 5v5 expected goals make them better. Here is the model, and what we changed.

Figures computed 2026-10-15, against the models as they stood that day. We change them — see the methodology page for where they are now.

Our power rankings went out on Wednesday, and the replies had a theme: eight games into a new season, why does last season still count? It is a fair question. Before answering it, it is worth agreeing on a few things about hockey that most fans would accept, whatever they think of our rankings, because the answer follows from them.

TL;DR

• A record is not a team, and a win is not proof. The better team wins 61% of the games between two clubs, and the best team in the league loses about 26 games a year.
• The rating measures strength, and we test it by what happens next. We score it by the points each team earns over the rest of the season, replayed over past seasons.
• Ten games in, the record so far has a rank correlation of 0.331 with the rest of the season. Our rating now has 0.575.
• Dropping last season makes it worse at every point we tested, even when the alternative chooses its own weights.
• What was missing: this season’s 5v5 expected goals, which settle faster than wins. Goaltending, finishing, special teams and shot attempts were tested too; none helped.

Where we can start

Each comes with what the results say: 12 full regular seasons from 2011-12 to 2025-26, leaving out the shortened ones, 15,252 games in all.

• A record is not a team. It is what happened in a handful of games, against one schedule, with one set of bounces. Of the teams in the top eight by points percentage after ten games, about half (52 of 96) were still in the top eight at the end of the season. One in five teams from the bottom eight (19 of 96) finished in the top half.
• Beating a team does not make you better than it. In games between two clubs, the one that finished the season with more points won 61% of the time. Even with a wide gap between them, like a 60% team against a 45% team, the better club won only 71%.
• Even the best teams lose a lot. The league’s best team each season won 68% of its games on average, and never more than 79%. It still failed to win about 26 games a year.
• Most games are close. Of all games, 43% are decided by one goal or go past regulation, and the home team wins 54%. A single game is a coin weighted only a little.
• So there are two kinds of ranking. Standings tell you what has happened. A power ranking should tell you how good a team is, which means looking past the record to the evidence underneath it, and that can be checked against what happens next. Ours is the second kind.

How power rankings are usually made

Most weekly power rankings are polls. ESPN’s comes from a panel of commentators, analysts, reporters and editors who each rank all 32 teams, and the ballots are combined. NHL.com’s Super 16 has been a points-based vote among its staff. A poll is a measure of informed opinion, and the standings are the natural place for that opinion to start, adjusted up or down for what the voters have seen.

Model-based rankings do a different job: they try to predict what comes next, which means not over-reacting to a week. They lean on more than this season. FiveThirtyEight’s NHL Elo started every team at 70% of its rating from the end of the previous season, the same carry-over we use. MoneyPuck’s rankings come from a predictive model built on wins, the quality of chances for and against, and goaltending. The Athletic’s team ratings are built from player values projected over the past three seasons.

Ours is the second kind. The rest of this post is why, and what we found when we checked.

What readers asked

Several replies argued that last season should not count at all: rosters turn over, and a team that is 6-2-0 should rank above a team that is 3-3-1. Most of the questions were about three clubs: the Leafs in 24th, the Canadiens in 4th and the Oilers, 5-0-1 and 11th.

We checked, and ranking teams by their record so far is the least accurate option we tested, all the way to the halfway mark of a season, the last point we checked. But the replies also found a real weakness. Our rating learned about this season only through results, and results are the slowest thing to tell you who is good. We have changed that.

What the power rating is for

We built the rating to answer one question: how strong is this team? Strength cannot be seen directly, so the fair test of any strength estimate is what happens next: stronger teams should earn more points from here on. That makes it checkable. A ranking that only describes what has happened is a standings table, and the NHL already publishes one.

The rating has four parts:

1. Last season, through Elo: each club starts the season at its final rating from the year before, pulled most of the way back to average.
2. This season’s results: Elo moves after every game, based on the result and the opponent, and that movement counts in full.
3. Roster strength: the projected wins above replacement of the players on each club’s NHL depth chart today. This is how the rating knows about trades and signings that Elo has never seen.
4. This season’s 5v5 expected-goal differential per game, new this week and explained below.

The weights on last season and roster strength were fitted on 296 team-seasons to predict final standings points.

How much did last season count?

We split every team’s rating into its parts, as it was when the graphic went out and as it is after this week’s changes. Measured by how far each part spreads teams apart (standard deviation, in rating points):

So yes: eight games in, last season was doing more of the sorting than this one, and more than it should have. Now this season’s results and expected goals together count for more than last season does. The three clubs readers asked about, now:

Was that wrong?

The way to find out is to go back and see what would have worked. We replayed every regular season from 2011-12 to 2025-26 with complete data (13 of them), stopped each one after 5, 10, 20 and 41 games, ranked the teams several ways, and checked each ranking against the points teams actually earned for the rest of that season. The score is a rank correlation: 1.0 would be perfect.

“Our rating now” adds this season’s 5v5 expected goals, explained below.

Three things stand out.

• The record so far is the weakest, at every point we tested. After ten games it correlates 0.331 with what comes next, and it does not catch up with the others even at the halfway mark. Ten games is not enough hockey for wins and losses to separate a good team from a lucky one.
• Last season is not dead weight. Elo, which carries last season in, beats the record so far by a wide margin. Rosters do turn over, but most of a team comes back.
• Roster strength adds to Elo, which is why the rating blends them.

We also tested the replies’ proposal directly: keep roster strength and this season’s expected goals, and drop last season, or drop Elo altogether. To be fair to the alternatives, each was allowed its own best weights for roster strength and expected goals.

Dropping last season makes the rating worse at every point we tested, including the halfway mark. So the replies are wrong about the cure, but they spotted the symptom: the rating’s only way to learn from this season was through results, and early results are noisy.

What we changed

Expected goals measure the chances a team creates and allows, not whether they went in, and they settle much faster than wins. Each team’s 5v5 expected-goal differential per game this season, against the league average, now counts at 40 rating points per goal of differential. That weight came from the same replay; anything from 25 to 50 does about as well. The same thing with actual goal differential added nothing, because Elo already holds the results.

Checking the rating against the replay turned up a mistake of ours, too. The weight on last season was fitted to the starting rating our standings projection uses, but the power rankings applied it to a less regressed version of last season’s rating. In effect they gave last season about one and a half times the weight our own fit supports. The replies were right that last season counted for too much, if not that it should count for nothing. The power rankings now use the fitted weight, and the replay above was run on the corrected rating.

We fixed one more thing while we were in there. Roster strength summed the projected wins of every goalie on a depth chart, and each goalie’s projection assumes his own workload, so a club carrying three goalies was credited with all three. Goalies now share one net: the season’s games go to the projected starter first, then the next.

What the changes do to the ranking, against the graphic we posted on October 14. The posted ranks count the games through October 13; the ranks now also count October 14, when the Canadiens were the only one of these teams to play:

Power rankings and the standings projection

One reader asked how the Oilers could be well down our power rankings and near the top of our standings projection. The two answer different questions. The power ranking is how strong a team is from here on, against an average schedule. The standings projection is where it finishes: that same strength played over its actual remaining schedule, plus the points it has already won. This is the same split FiveThirtyEight, MoneyPuck and The Athletic use: a strength rating, and a forecast built on top of it.

Both now use the same rating, so the only differences are the schedule and the points in the bank. For the Oilers:

A fast start is banked even if we do not expect the pace to continue, which is why a team can project higher than it ranks. Until this week the two models did not quite share a rating, which made the gap look bigger than it should have; that is part of what we fixed.

What we tested and left out

We put five other numbers through the same replay, each on top of the new rating, each across a range of weights. None earned a place:

Goaltending and finishing are already in the rating, because they decide games and Elo learns from games. What the replay says is that a hot start in either does not predict the rest of the season beyond the wins it has already banked.

The teams people asked about

The Leafs were 24th at 3-3-1. Last season is what holds them down, and this season has not lifted them: they have 45.9% of the 5v5 expected goals in their games, and have allowed 21 goals on 24.4 expected, so their goaltending has saved them about three. Their record is roughly what their play has earned. They are 25th now.

The Canadiens were 4th at 3-2-1, mostly on last season and the roster, and lost in Washington the next night. Their play has been better than the record says: 54.7% of the 5v5 expected goals, with 21 goals scored on 24.8 expected and 24 allowed on 19.5. They are 7th now.

The Oilers were 11th at 5-0-1, winning on finishing: 28 goals on 15.6 expected, with an even split of 5v5 chances (50.7%). Scoring twelve more than expected in six games is a great stretch, and the rating credits the wins through Elo. It does not assume the finishing continues. They are 8th now.

One more, because it shows the other side of the same coin. The Rangers are 6-2-0 and were 21st. They have scored 21 goals on 23.1 expected and allowed 13 on 24.4 expected, with 48.9% of the 5v5 expected goals. That is a team being carried by its goaltending. They are 18th now.

Limitations

• The replay scores each method against one season’s remaining points, which are noisy themselves. A correlation of 0.575 is not 1.0, and no ranking ten games in will be.
• Roster strength still comes from September projections. When this season’s WAR is published, after about 20 games per team, the roster part will use it.
• 5v5 expected goals leave out special teams. A team that wins on the power play gets that credit through results only.

The current rankings are on the power rankings page.

Data: Hockey Alchemy expected-goals model and team Elo. Replay: rest-of-season standings points, regular seasons 2011-12 to 2025-26, cuts by the median team’s games played. Rankings and team figures as of October 15, 2026, through the games of October 14.

More from the Notebook

How We Calculate NHL Power Rankings: The Elo System Behind Hockey Alchemy

A transparent look at the Elo rating system that powers our NHL power rankings, game predictions, and Stanley Cup odds. K=12, home ice = 50 points, 30% regression to the mean — and three different ways we account for margin of victory.

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.

Inside the NHL EDGE Goal Visualizer: How We Scrape, Clean and Use Tracking Data

The NHL publishes about fourteen seconds of player tracking for every goal, ten frames a second. We hold 26,237 of them. How we fetch it, the four bugs we shipped handling it - none of which ever errored - what it is good for, and the one thing it can never do.

Related: All Posts Methodology Glossary