2024-2025 NHL Elo Ratings & Projections
Last updated Thu, Dec 5, 2024, 1:13 AM EST.
1 | Dallas StarsDAL | 1575 | 93.5% | 10.5% | |||||||
2 | Carolina HurricanesCAR | 1569 | 94.2% | 9.5% | |||||||
3 | Florida PanthersFLA | 1566 | 91.6% | 8.4% | |||||||
4 | Winnipeg JetsWPG | 1563 | 95.8% | 8.2% | |||||||
5 | Toronto Maple LeafsTOR | 1562 | 93.0% | 8.1% | |||||||
6 | New Jersey DevilsNJD | 1560 | 92.0% | 7.2% | |||||||
7 | Edmonton OilersEDM | 1560 | 82.6% | 6.8% | |||||||
8 | Minnesota WildMIN | 1550 | 95.5% | 6.1% | |||||||
9 | Vegas Golden KnightsVEG | 1546 | 92.2% | 5.6% | |||||||
10 | Washington CapitalsWSH | 1544 | 93.2% | 5.1% | |||||||
11 | Tampa Bay LightningTBL | 1544 | 77.9% | 4.5% | |||||||
12 | New York RangersNYR | 1539 | 70.3% | 3.7% | |||||||
13 | Vancouver CanucksVAN | 1533 | 78.9% | 3.6% | |||||||
14 | Los Angeles KingsLAK | 1532 | 79.2% | 3.3% | |||||||
15 | Colorado AvalancheCOL | 1534 | 64.0% | 3.0% | |||||||
16 | Boston BruinsBOS | 1519 | 52.2% | 1.8% | |||||||
17 | St. Louis BluesSTL | 1487 | 28.0% | 0.6% | |||||||
18 | Buffalo SabresBUF | 1489 | 22.5% | 0.5% | |||||||
19 | Ottawa SenatorsOTT | 1485 | 21.4% | 0.5% | |||||||
20 | Utah Hockey ClubUTA | 1481 | 21.3% | 0.4% | |||||||
21 | Calgary FlamesCGY | 1476 | 36.5% | 0.4% | |||||||
22 | New York IslandersNYI | 1481 | 19.2% | 0.4% | |||||||
23 | Philadelphia FlyersPHI | 1475 | 27.1% | 0.4% | |||||||
24 | Nashville PredatorsNSH | 1485 | 11.1% | 0.3% | |||||||
25 | Detroit Red WingsDET | 1477 | 18.3% | 0.3% | |||||||
26 | Seattle KrakenSEA | 1475 | 16.6% | 0.3% | |||||||
27 | Pittsburgh PenguinsPIT | 1474 | 15.5% | 0.3% | |||||||
28 | Columbus Blue JacketsCBJ | 1439 | 9.3% | 0.1% | |||||||
29 | Anaheim DucksANA | 1420 | 3.4% | 0.0% | |||||||
30 | Montreal CanadiensMTL | 1426 | 2.4% | 0.0% | |||||||
31 | Chicago BlackhawksCHI | 1411 | 0.6% | 0.0% | |||||||
32 | San Jose SharksSJS | 1393 | 0.8% | 0.0% |
Version History
v3.0 - current
Re-ran backtesting of model parameters with a bias towards the salary-cap era. This resulted in changes to several parameters, including increasing the K-factor from 6 to 8 and reducing home ice advantage from 50 to 42, which were found to improve predictive performance of the model at the game level. This in should improve the quality of the forecast, which is done via game-level monte carlo simulation.
v2.0
I added a Vegas totals bias to the teams' season starting elo. This differs from v1.0, which uses only last season's ending elo, with a small regression back to the mean. In this change, I mapped Vegas totals to elo ratings using a simple linear regression model, then I blended the v1 rating with the Vegas projected rating, at 65% Vegas-biased elo and 35% v1 elo. The most notable change is Boston going from first with a 20% cup chance, down to 4th at only 9%.
v1.0 - FiveThirtyEight's
This was an exact replica of FiveThirtyEight's NHL forecast.
What is this?
This is an elo forecast based on FiveThirtyEight's deprecated NHL Elo model and forecast. You can read about how the original version works on FiveThirtyEight's website. Model credit to @ryanabest and @neil_paine, with modifications as detailed in the version history above. This website is not affiliated with FiveThirtyEight or ABC News.
How does it work?
Each team is given an elo score, which is effectively their strength relative to other teams. At the start of the eason, this score is a blend of their previous season's ending elo, regressed to the mean, and an implied elo derived from the Vegas season point total projections. Before each game, we compute the probability of each team winning based on their elo ratings and other factors like home ice advantage. After each game is played, we update the teams' scores based based on who won, by how much, who was originally expected to win, etc.
To simulate the season, we go game by game, randomly picking a winner of each game weighted on their pre-game probabilities. We then update their elo rating as described above as if they really played the game, and continue to the next game. We do this for the whole season (including playoffs) tens of thousands of times, recording how each team did in each full season simulation.
We then average their results across simulated seasons to get the probability of making the playoffs or winning the cup, etc., which are presented in the table above. For example, if the Edmonton Oilers won the cup 6,000 times in 50,000 simulations, they would show a 12% chance of winning the cup in the table above. If the Leafs won the cup 5,000 times, the table would show 10%, but we'd all know the true probability is still 0.
Is it accurate?
These projections have been shown to be exceptionally well-calibrated over the years when 538 was doing it, and last year when I did it. So across a season, they're quite reliable. When you factor in the simplicity of the modeling, the ROI on effort to results is quite frankly astounding. (As an aside, other projections websites go to mind-bending lengths incuding things like puck locations and player tracking, and achieve only marginally better results.)
However, there are shortfalls to the system, especially at the game level. It doesn't account for day-to-day changes like injuries, trades, suspensions, coaching changes, etc. When these things happen, it can take a few games for the elo score to capture the impact, and so any given game could be poorly calibrated in the short term. One way to think about this is if the Vegas money line is significantly different from the elo projection, it's not a value find but an indication that there's a significant factor the model doesn't know about.
How do I use this to gamble?
You don't. I'm just some random guy on the internet (see below) and this could go away at any time. It's made for entertainment and hockey obsession purposes only.
Why do this?
I'm a super fan of both hockey and elo ratings. So when I learned that FiveThirtyEight wasn't doing their forecasts last season, I shed a few tears. Ironically, a hockey injury gave me some extra time back in my week, which I used to build this.
Who made this?
I'm a software engineer, machine learning practioner and beer league hockey player. My day job is building software for Coastline, a driving education startup.