Hockey Elo Predictions

Hockey game predictions inspired by FiveThirtyEight’s methodology for their NFL predictions.

Being a Buffalonian (sorta) and given the Buffalo Sabres recent hot streak, I’ve developed a newfound interest in the sport of hockey. I’m also quite interested in data science, especially when it comes to prediction, and so the question emerged: how easy would it be to make a system (sorta) like FiveThirtyEight’s predictor for the NFL, but use it to predict the outcomes of NHL games? The answer: pretty easy. All it takes is some data and a rudimentary knowledge of the Elo ranking system.

The above is the formula for calculating the probability that team A will win a game. This can also be interpreted as the expected score for A.

The S term is the A’s actual score. This, so far as I can tell, isn’t necessarily set in stone, but sensible values for wins, losses, and ties could be 1.0, 0.0, and 0.5 respectively. These values are supposed to be between 0 and 1.

This is the fun part. Basically, using the probabilities you get out of your Elo calculations, you simulate a bunch of games (i.e. 100000) and then determine the mean outcomes. Mostly these values converge close to the values you get out of your Elo calculations.

On the FiveThirtyEight NFL Predictions methodology page they talk about a moderate bonus in the newly calculated ranking for Margin of Victory. This seems simple enough to implement. I didn’t think it would be too important for hockey given the games seem to generally be fairly close, but it’d be nice to see if a Margin of Victory multiplier improves the system’s accuracy. Also, Elo has no means of adjusting for factors outside of the outcome of games, such as certain players not playing, the acquiring of new players, etc. I’d have to think about it, but it’d be interesting to try to augment rankings by accounting for factors like the ones mentioned.

In the future it’d be cool to apply Elo elsewhere. One wonders how it’d perform in non-sports contexts. Also, FiveThirtyEight’s model seems to incorporate a team’s history well into the past. In my basic model I didn’t do this. I suppose I should test the model’s accuracy given more history and less history. I’ve thrown something together on glitch.com, in the future it’d be nice to polish the interface more; it’s super slow for the time being because I’m manually reading a file and then sending back calculations on every request.

Thanks for reading!

Related posts:

During Chanukah we must remember the victims of the Korban, השואה, the Holocaust. I share this tale from Professor Yaffa Eliach, Hasidic Tales of the Holocaust — told to me by her on Shabbos Chanukah…

What a month it has been. My plan with this letter was to start it with some alliteration based on “May”, the Magnificent Month of May, or the Marvelous Month of May but it all got just a bit too…

Super Bowl LV was an American football game. It played to determine the National Football League (NFL) champion for the 2020 NFL season. The game was televised nationally by CBS. The Weeknd will…