Our NHL Model is…
*drum roll*
…back.
The Stanley Cup Playoffs begin tonight, and with them, we roll back out Gelo, our predictive model which gives:
- Individual-game probabilities
- Playoff-long probabilities
- Expected numbers of goals scored (new this postseason)
We first published this model last year, and if you’re reading this and just want to see the probabilities and projections themselves, you can find those here. If you want to know how it works, well, that’s what we’re here to do.
It’s a simple model, as far as these sorts of things go, with only one input during the season (the score of every game) and one input during the offseason (the Vegas projected point total for each team). Here are the details:
Gelo – Our Goal-Based Elo Foundation
Elo is a ratings system which originated in chess. It’s the basis for many FiveThirtyEight models of this sort (more on FiveThirtyEight below—if you’re already thinking, “Did these guys copy FiveThirtyEight?” the answer is yes and no), which is where we learned about it. Elo is conveniently simple—in its original form, each competitor has a score entering each competition, and following it they gain or lose points on or from that score based on whether they won or lost and how surprising that win or loss was, with their opponent losing or gaining a corresponding number of points from their own Elo score. If there’s a wide gap between two competitors, the favorite can only gain a small number of points with a win, while the underdog can gain a large number. It’s a self-adjusting model in this way, reacting in the ways one would hope to surprising and unsurprising results.
We added a G to the front for two reasons: First, “Gelo” can be pronounced “Jello,” which is useful for speaking about the thing out loud. Second, we’re using goals as our only live input. Goals starts with the letter ‘G.’
We made one other big change to Elo: Rather than basing its postgame changes solely on wins and losses, we base those off of both whether a team won or lost and the team’s over or underperformance relative to Gelo’s expected margin. It’s possible, then (rare, but possible) for a team’s Gelo score to worsen even after a win, if the win was underwhelming enough.
FiveThirtyEight Note
Ok, so this is, um, very similar to FiveThirtyEight’s NHL model, which prompts a very obvious question, which we’ve already asked above: Did The Barking Crow copy FiveThirtyEight?
We didn’t copy FiveThirtyEight’s formula. We actually published our model first, hustling it out right before the end of last season to avoid this exact question, knowing ESPN was about to start broadcasting NHL games again and, since ESPN and FiveThirtyEight are both owned by ABC, a FiveThirtyEight NHL model was likely. Our models have similar approaches but different results. FiveThirtyEight views the Colorado Avalanche as the Stanley Cup favorites, giving them a 20% chance of winning, double the 10% Gelo assigns Colorado today (both models have the Florida Panthers 14% likely to win, but that’s mostly coincidence). Our model is more elastic than theirs, with our optimization leading us to build Gelo to react more quickly to surprising results (I’m not familiar with the specifics of their model, but our latest optimization covered results from the start of the 2015-16 season through the end of last year’s Finals). We also, I assume, use different starting points. Which brings us to our next subheader:
How We Handle the Offseason
A lot happens in the NHL offseason, and it’s difficult to quantify, which is why a large part of our offseason adjustment to each team’s Gelo comes from that team’s projected point total from the Las Vegas betting markets. There are other things as well—a bit of regression to the mean—but most of it comes from what bettors and oddsmakers expect to happen (in other words, using Gelo to bet on NHL preseason point totals is a largely circular exercise).
Ogelo and Dgelo
We have two subcomponents to Gelo, encompassing offense and defense. These are changed by each game’s Gelo change, moving depending on whether a team’s over or underperformance relative to expectations came more on the defensive or offensive end (or if they overperformed on one end and underperformed on the other). They move in correspondence with the opponent’s, just as Gelo does (each team’s single-game Ogelo change is equal to their opponent’s Dgelo change). We use these to give expected goal totals for each team in each game (theoretically allowing us to bet over/unders in addition to the moneyline and the puck line, though we’ll have more on that below).
Ogelo and Dgelo are where the model gets its scale, which is to say: For each game, if you average the visitor’s Ogelo with the home team’s Dgelo, then do the same with the home team’s Ogelo and the visitor’s Dgelo, then sum the results…you’ve got your expected total. (The Gelo gap, for whatever it’s worth, is roughly one-fifth the expected goal margin, for reasons of…us not taking the time to figure out how to scale it otherwise. Put that on the list for next year, or later.)
In the offseason, Ogelo and Dgelo move equally, each bettering or worsening by half the bettering or worsening of a team’s Gelo from the end of the previous season.
Simulations
In addition to forecasting results in single games, Gelo runs simulations of the remainder of the season. 10,000 simulations at a time, it uses random numbers to play out various results and see what happens. These simulations are run “hot,” meaning ratings change according to the various results which occur within each simulation. In some simulations, the Panthers remain rated the best team in hockey throughout the Stanley Cup Finals. In others, the Avalanche, or the Wild, or someone further down like the Oilers takes the crown.
Miscellany
Home-ice advantage exists, our findings indicate there’s no real widening that occurs as totals increase (a one-goal favorite in a hypothetical game with somehow an 11-goal over/under is just as likely to win as a one-goal favorite in a hypothetical game with somehow a 1-goal over/under).
Betting
We wouldn’t recommend using this to bet on games, as it’s untested and single-game betting markets are notoriously efficient. We think it can be useful for futures, where markets are not at all as efficient, but we haven’t tested that yet either, so if you do it, be cautious and—as always—don’t bet more than you’re willing to lose. We’re planning to publish hockey bets daily from here through the end of the playoffs, as a test. We’ll see how it goes.
