For years, readers have asked us to add NHL predictions to the host of forecasts we run here at FiveThirtyEight. And clearly, hockey’s absence wasn’t out of any aversion to the sport on our part — editor-in-chief Nate Silver, myself and many other writers here have been pretty obsessed with it, in fact. So after soft-launching a beta version of hockey Elo ratings last year, we are proud to finally roll out a full-fledged NHL forecast interactive for the upcoming 2021-22 campaign.
In the spirit of our earliest Elo explainer — for the NFL, published in September 2014 — here’s a list of Frequently Asked Questions you might have about the new forecast and what it sees in the season ahead:
How does it work?
The new NHL forecast model uses Elo rati —
Really, Elo ratings? Again?
Yes. We use the Elo rating system a lot here at FiveThirtyEight, but that’s because it’s a really useful, adaptable framework for tracking the quality of a team’s play over time. There’s a reason we like it so much!
Olympians, surgeons and even toddlers have used this technique to improve their focus
OK, fine. So what’s hockey Elo like?
Hockey Elo works basically the same as Elo in other sports. (Like I said, it’s adaptable!) Each team is assigned a rating, and the league average is roughly centered at 1500. Those ratings update after each game, allowing us to adapt to teams’ changing performances. And one of the great features of Elo is that the difference between any two teams’ ratings — with a few adjustments — can be turned into win probabilities for a game between those teams.
What are the adjustments?
Since our initial NHL model is fairly basic (compared with some of our other, more elaborate Elo models), it adjusts only for home-ice advantage — which is worth 50 extra points of Elo for the home team in hockey — and the fact that big favorites are more likely to win in the playoffs, Montreal’s opponents in last year’s postseason notwithstanding. Home-ice advantage is zeroed out for games played on neutral ice (i.e., the 2020 postseason bubble), but when it is invoked, it is the same for every team regardless of attendance. (So the teams starting this season in reduced capacity arenas — Montreal and Vancouver — still get the full adjustment.)1
related: The Seattle Kraken Have An Expansion Blueprint. Can They Follow Vegas And Build A Winner? Read more. »
The NHL gives out points for losing. Does that matter?
It does in our season forecast, which simulates the rest of the schedule 50,000 times and tracks where every team ends up in the standings. In losses that go into overtime — both in real life and our simulated future contests — a team will get the Bettman Point. However, one of the interesting findings in the research we conducted for NHL Elo ratings was that how a team won or lost — in regulation, overtime or the dreaded shootout — doesn’t seem to matter when measuring a team’s quality. When it comes to predicting a team’s future games, all one-goal contests are created equal, even if they happen in a glorified skills competition and the loser winds up receiving different numbers of points.
What are some of Elo’s lessons in hockey, compared with other sports?
A big one is that any given hockey game’s result doesn’t budge the ratings that much. In the NHL, Elo’s “K-factor” — which regulates how quickly it updates after each game — is just 6, which is just slightly higher than in MLB (where K is 3) and much lower than the K-factor of 20 found in both the NFL and NBA models or the 28 found in the WNBA model. (This is in keeping with other research that shows the role of luck is much larger in baseball and hockey than in football and basketball.) This is Elo’s way of being skeptical that the results of any one game really tells us much about the relative quality of the two teams involved.
The range of team Elo ratings in the NHL is pretty narrow, relative to other sports. Last season, hockey Elo ratings ranged from 1596 (the Tampa Bay Lightning) to 1420 (the Detroit Red Wings). That 176-point gap between best and worst is a narrower range than we saw in MLB (178), the NBA (465) or the NFL (476) in their most recent completed seasons.2
What happens to the ratings between seasons?
After each season, we found that reverting every team’s Elo 30 percent of the way toward a rating of 1505 (basically league average)3 is best for predicting future games. In terms of year-to-year upheaval, that puts the NHL somewhere between the NBA (whose ratings are reverted by 25 percent between seasons) and the NFL and MLB (which revert Elo by 33 percent), and well below the WNBA (which reverts by 50 percent).
Perhaps in future iterations of the model, we’ll also bake in Vegas over/unders to help make the preseason ratings more accurate (which is what we do in the NFL, for instance). But for now, that means the pecking order of teams going into 2021-22 is the same as it was at the end of 2020-21, just with the ratings more flattened out.
But not every team had a rating last year! How does the model handle new teams like the Seattle Kraken?
Great question. For almost the entirety of NHL history, expansion teams were, as a general rule, very bad. (Like, very bad.) That all changed with the 2017-18 Vegas Golden Knights — who were not only actually good, but came within three victories of winning the Stanley Cup. There are many reasons why Vegas was so successful right out of the gate, but one had to do with being the first expansion team of the NHL’s salary-cap era (which began in 2005-06). Because of other teams’ financial constraints and some clever dealing, the Knights were able to field a roster with a surprising amount of talent by debut-franchise standards.
So what do we do with Seattle? By most accounts, other teams were a lot smarter about handling the Kraken’s expansion process than they were with Vegas, and as a result Seattle had its eye on a more methodical build than the Golden Knights did. Still, the betting markets see the Kraken as an average team at worst going into 2021-22. For Elo’s purposes, we didn’t want to necessarily overreact — and overfit — to Vegas’s instant success, but it’s plain that the days of terrible expansion teams are over in a modern, salary-capped NHL that charges $650 million expansion fees. Averaging together a bunch of different prospective data points that could suggest how Seattle might do,4 we arrived at a very rough guess of 1490 for the Kraken’s initial Elo (compared with the 1380 we set for all pre-Golden Knights expansion efforts).5 That number is sure to change, however, as the season goes on and we learn more about this new team.
Who does Elo like this season?
Here’s our initial round of projections going into the season:
Just like most of last season, we essentially have three co-favorites at the top for 2021-22: the Colorado Avalanche, the Golden Knights and the defending-champion Lightning. Each claims a 12 percent chance to hoist Lord Stanley’s Mug. But there’s plenty of the NHL’s trademark parity as well, with a 64 percent chance that a team other than one of those three wins it all. And as for the Kraken? We give them 45 percent odds to make the playoffs (and even a 1 percent chance to win the Cup) in their inaugural season.
So now that there’s a hockey model, when does the curling forecast come out?
Hmm, come to think of it, the Winter Olympics are right around the corner … 🤔🤔🤔