We’re very nearly at the start of the 2021-22 NHL season, and as such, all of the very smart statistical minds in the public hockey sphere are releasing their season previews showing what their models project will happen this year. Other than the playoffs, this is one of my favourite times of the year, because it allows me to learn a bit more about statistical projections while getting my excitement level up for the upcoming season.
This year we’re going to be looking at 4 models:
Each model is constructed differently, with different levels of “open-source” information on how it’s created. I’m not going to go too in depth on how each model works, but I’ll try to give a high level description, and link to the in depth explanations where I can.
I’ll also summarize the projections in a table at the bottom of the post for quick consumption, if that’s your thing.
Truthfully, in writing this article I’m only now discovering that FiveThirtyEight does NHL projections. I’ve used them extensively for political projections at election time, but never for hockey.
Their model works in a basic format: give each team a relative strength, which they call an “Elo rating”. This Elo rating is then used to simulate who will win each match based on the following formula:
EloDiff is Team A’s pregame Elo rating minus Team B’s pregame Elo rating, along with some adjustments:
- A home-ice advantage adjustment adds 50 points to the home team. An EloDiff of 50 would be good for a 57.1 percent win probability, so this sets the home team as a slight favorite if the two teams were otherwise even. (Games at neutral sites have no home-ice adjustment.)
- A playoff adjustment that multiplies the adjusted EloDiff by 1.25 for playoff games, accounting for our finding that favorites tend to outperform underdogs by a wider margin in the playoffs than they do in the regular season
The pre-season Elo rating that we’re using in today’s post is determined as follows:
For our NHL forecast, teams retain 70 percent of their rating from the end of the previous season and are reverted 30 percent toward 1505. For example, the Toronto Maple Leafs ended the 2020-21 season with an Elo rating of 1541, so they start the 2021-22 season with an Elo rating of:
So, we know that the Leafs are about 31 points above league average to start the year, as projected in this model. How do they look overall? Here are their projected Atlantic Division standings:
As you can see, our Leafs sit uncomfortably near to and ever so slightly behind the Florida Panthers, who they should be battling all season long for that 3rd spot in the Division. The Lightning project to sit atop the division and share a three-way tie with the Avalanche and Golden Knights to win the Cup, which would be a staggering three in a row for them.
A family friend here at TLN, Dom is a former TLN contributor and is doing a great job at The Athletic now providing all kinds of interesting, albeit paywalled, content. The model works on “Game Score Value Added”, or GSVA, which is a predictive version of “Game Score”. Dom uses Game Score, a concept borrowed from NBA analytics, to show how well a play performed in a particular game. GSVA uses their past performances to predict how players will perform in future games.
Adding up the GSVA for each player gives each team a total strength. This is then used to simulate the season and figure out who will perform best. This model is a big fan of the Toronto Maple Leafs, like Dom is, and while we can assume he’s being as unbiased as possible in the creation of the model, it’s still hilariously awkward for him.
The Leafs project to have the second most points in the NHL next season, per this model, and a 14% chance at winning the Cup. A 1 in 4 chance to make the Cup final is… astronomical.
Trust me, Dom finds this just as crazy as you do. From his season preview piece on the Leafs:
Unfortunately the model, in all its sentient wisdom, has other plans because it would prefer I
look both biased and wrong every year. No, there isn’t a hidden modifier to make the Leafs look
better than they actually are because I, as a lifelong fan, know better. The team does not
deserve that benefit — the model just doesn’t know any better. There’s no coefficient for being
cursed, it just projects after all the inputs are put in that this is not only a good team, not only a
great team, but an elite one. That its time will come.
I am personally sick of the undeserved optimism. The model is child-like in its unrelenting
prediction that things will get better for this team. I am instead broken and defeated.
The third model is from a site often used in these Staturday columns, partially because of the useful and successful model developed by the twins from the Twin Cities, Josh and Luke, who are @evolvingwild on Twitter.
From their season preview post, here is how the model works:
We still use a roster-level system, project various metrics for every player for next season (notably xSPAR [expected Standings Points Above Replacement, an extension of their Wins Above Replacement model] and our own “game score” metric), determine the roster for every team (based on CapFriendly’s depth charts and research for when injured players might come back), aggregate those individual projections into team ratings for forwards / defensemen / goalies, and compute game probabilities for every game for the ’21-22 schedule. We then run a Monte Carlo simulation 50,000 times with all of that in place to project point totals for all teams in the league.
This is very similar to Dom’s model construction, so I won’t dig too deep into how it works. Let’s just look at the results:
In this model, despite being constructed in a way very similar to Dom Luszcyszyn’s, has the Leafs finishing with 105.6, good for just 3rd in their very strong division. Notice that they’d not only be 3rd in the division, but fifth in the entire NHL, which is still very strong. A note about why that is, from their preview post:
While these teams are all very good, they also get to beat up on three of the five worst teams in the league (Ottawa, Detroit, and Buffalo), which pads their total points overall.
Micah Blake McCurdy
The last model we’re going to look at comes from another hockey stats all-star, Dr. McCurdy. He runs the site hockeyviz.com, where he hosts lots of great visualizations and also yearly season predictions like what we’ll look at today.
Here is a breakdown on how the model works, from his season preview post:
I simulated the 2021-2022 NHL regular season a million times to estimate what is likely to happen. (I used a computer to help me.) To estimate the probability of the home team winning each game I used my prediction model, Magnus. Curious readers will find lots of detail following that link, but, very briefly:
10 I estimate the likely rosters for each team, starting from their contract list on October 5, 2021, and then weighting by past playing time. Injuries and suspensions are taken into account.
I estimate the individual impact of each player and head coach on:
- Shot generation and suppression, at even-strength and special teams, taking account of teammates, competition, and the effects of score, zone, coaching, and home-ice deployment;
- Rates of taking and drawing penalties
- Individual tendencies towards shooting or passing; and
- Ability to shoot or stop the puck, as appropriate.
These estimations are done using NHL-gathered input data from the past dozen regular seasons, suitably weighted; or, for a few important new players, using estimates from Hannah Stuart, a prospect expert with whom I have a long-standing partnership. These estimates include playing time as well as impact; the players chosen are ones that we expect to play substantial roles with their teams this season.
An adjustment for rest.
I removed some of the details from the above for succinctness, so please read the preview post if you want to know more.
Unlike the above models that give us predictions for standings points and Cup-win chances, Micah’s model gives us predictions for more granular elements of the team’s upcoming season. For instance, here is a chart showing how the model predicts each team’s shot generation will look this year:
However, our primary goal in this post is just what we have from the other models, standings points and playoff odds. Here’s how the Leafs stack up:
- Expected to lead the Atlantic division with 104 points
- 95% chance of making the playoffs
- 0.1% chance of getting the 1st overall pick
This is the second model we’ve looked at that has the Leafs slated to win their division, which is a really strong place for them to start.
Final Thoughts and Summary
Overall, I’m getting really excited to see if the Leafs can live up to expectations and win the division again this year, though obviously this year’s non-Canadian division is going to be much more difficult to win.
Here is a table summarizing the results from the different models. We’ll leave it at that for today, so I hope you all enjoy Thanksgiving!
|Model||Points (Place in Div.)||Playoff Chances||Cup Chances|
|Dom Luszcyszyn||109.3 (1st)||98%||14%|
|Micah Blake McCurdy||104 (1st)||95%||—|