This post was most recently updated on June 28th, 2021. For the 2021-22 NHL Season and subsequent seasons, it may not be accurate. Always view our NHL Playoff Probability page for the most current link to how our model works.
For starters, this should probably be viewed as a beta version of our model. There are changes we’ll be making before next fall, most prominently splitting Gelo up into an offensive and defensive component, and thereby changing the scale such that the gaps in Gelo mean something more than “5/23 of the projected goal margin on a neutral rink.” We may also change what exact outputs we output. The basic structure will likely remain the same, though, and the basic structure is this, in FAQ form:
What is Gelo?
The “G” before “elo” just signifies that this is a version of an elo model that is purely goal-based. If you’re unfamiliar with elo, the gist of it and its imitators is that each competitor in the system has a score, those scores can be used to calculate win probabilities between competitors, and the more unexpected a result, the more each competitor’s score changes post-result. Elo itself originated in chess, and we’re largely familiar with it because it’s a basis for many of FiveThirtyEight’s models. In chess, there is no margin of victory, so the original elo is based solely on wins and losses. In sports, margin of victory sometimes matters, and so we incorporate it into ours.
How does Gelo respond to individual games?
Like elo, Gelo adjusts based on how unexpected a result is. In addition to reacting to the winner and loser, though, it also reacts to the margin by which a result occurs, with diminishing returns for bigger blowouts.
How does Gelo adjust during the offseason?
Before each season, Gelo incorporates a team’s end-of-season rating from the previous year, the average rating across the system (2.85, or roughly the average number of goals a team scores in an NHL game), and the Vegas point total for the team, with an adjustment to that last piece based on the division in which the team plays, that division’s average rating at the end of the prior season, and the number of games that will be played against said division in a given year (i.e., this usually isn’t a huge adjustment but this past offseason it was significant because of the unique structure of this season).
Does Gelo account for home-ice advantage?
Yes. Gelo adds 4.5 percentage points of win probability to the home team, based on the home teams’ win percentage from the beginning of the 2015-16 season through the end of the 2019-20 season. Theoretically, this could lead to a win probability being over 100%, but we have never seen anything particularly close to that in our backtesting, so we’ll wait to refine it until the offseason.
How accurate is Gelo?
In our tests, beginning with the 2015-16 season (for starting Gelo’s, we used only the Vegas point total), Gelo accurately predicted 57.0% of winners. It was slightly overconfident—it expected itself to predict 58.6% of winners. We’ll be looking to align those numbers this offseason.
How confident is Gelo at its most confident?
Gelo’s biggest favorites over the five seasons preceding this one were the Boston Bruins on February 15th, 2020, when they played the Detroit Red Wings in Boston. The Bruins were 84.8% favorites to win that night, which equates to a -559 moneyline with no vig.
How do your simulations work?
We use Gelo to simulate the remainder of the season 10,000 times each time we run it. In those simulations, Gelo is adjusting for a team’s results as each simulated season goes on. Teams get hot, teams get cold, and rather than assume each team will stay the same as the season goes on, Gelo adjusts to the simulated results, with each simulation its own parallel universe.
Does Gelo adjust to personnel changes or injuries?
Gelo does not adjust to anything but wins, losses, and margin of victory/defeat. This is a shortcoming of Gelo, but we’re a long way from fixing it. It’s probably going to stick around into the future. Take that for what you will.
Can I bet with Gelo?
You can use Gelo to bet, but we wouldn’t recommend it. We haven’t tested it against betting odds, and while it might provide some futures value, as futures markets—in our experience, at least—tend to be less sound than individual game markets, we’re also very late in the year at the point this is published, and are thus not going to encourage using it for such purposes (though we will, for what it’s worth, consider using it ourselves).
Does Gelo account for the uncertainty this season that arises from teams not playing teams outside their division before the playoffs?
Nope. As an elo-based model, it only reacts to individual results, and doesn’t carry teams’ division-mates with said teams as they rise and fall in the playoffs. Should it? Arguments can be made in each direction. In the end, the end-of-season Gelo is a small enough part of the next season’s starting Gelo that it isn’t going to make a huge difference.
How do I know how accurate Gelo has been in the past?
All we’re giving you right now is those 57.0% and 58.6% numbers from above, but we’ll have a bigger data set next year of past predictions when we backtest Gelo over the offseason. We were admittedly rushing to get it out this year, and that was one of the things that fell by the wayside.
