It’s October! It’s my birthday! It’s time for hockey! IT’S TIME FOR NUMBERS!
I won’t bury the lede here: in a 1,000,000-trial Monte Carlo simulation, FSU sweeps about 23.5% of the time, UAH sweeps about 11.6% of the time, and the rest falls in between. My model is still a work in progress (mainly in the data-management side; the calculations are easy), but it’s pretty simple.
|New OT Involved||20%
How I get to these numbers – it starts with BELOW
BELOW stands for Bringing Elo to the WCHA. Elo-style rating systems work on a simple principle: you can go into a match with an estimate for the likelihood that contestant A will win, known as an expected value. If contestant A wins, they are rewarded with a jump in their rating commensurate with the expected value. If contestant A had a high expected value, they won’t receive much of a change.
Conversely, contestant B could pull off the upset, and an Elo rating system will reward them for that. I explained this moderately well in December, and I’ll be refining that soon. If you’re new here, this is where you should go.
How we got here – re-calculating BELOW
Ferris State, despite all the wonder that it did through the postseason, starts the 2016-17 season with a BELOW rating of 1518, or just a hair above an average baseline of 1500. Alabama-Huntsville comes in at 1407. Both teams have been regressed to the mean. by one-third; UAH, since they were farther from 1500 (1361) than the Bulldogs (1527), moved closer.
ABOVE is Adusting BELOW through Operative Value Experiments, the model that uses that Elo-style rating and iterates, time after time. This used to be pretty easy, actually: in the regular season, you would win, lose, or tie; now you can win, win in 5×5 overtime, win in 3×3 overtime, or win in a shootout. I used to assign a different value adjustment based on whether a non-tie was settled in OT, but I didn’t predict those on a macro level.
I actually am predicting 16 outcomes per weekend: two combinations of four results: win, overtime win, overtime loss, and loss by team A. I could spit that data out at you, because to ABOVE, it actually matters whether Game #001 is a UAH win even if it’s a given that Game #002 is an FSU win. It doesn’t matter to you.
A brief sidebar as to why it does matter to ABOVE
Think about it, though: if a team with a 1400 ranking beats a team with a 1500 ranking — pretty close to where Ferris-Huntsville is — a split will move the teams closer, but order matters. To our example: a 100-point differential means that team A is expected to win 64% of the time. An upset means that team B get 0.64 x 40 = 25.6 points, and the spread is now 1474 – 1426; this ELO Difference Calculator will show that the expected win for team A is now just 57%. But if the reverse happens, and FSU wins first, the gulf is wider between the two teams, and the value of an upset is greater. We saw this last season also with Ferris State opening.
In short, a split does not mean that ABOVE will calculate the same BELOW regardless of order.
Say, when do we get that model?
Well, there are two things at work:
- I’m flying back today after 12 nights in Iceland. I haven’t had a chance to work on it while I’ve been here.
- I’m not sure yet how I want to model the new overtime.
Now, the new OT is weird. How do you model 3×3? You can do it two ways:
- It’s a crapshoot, so model it as 50-50 and go on with life.
- Do some guessing based on GF/GA differential or something silly like that.
There’s also the question of how often teams will push to go into the extended overtime. My current assumption in ABOVE is, “It’ll be like last year,” but I kinda doubt that. A weaker team is unlikely to have the skaters to win in 3×3 or in a shootout, and they may see the prospect of getting three points greater than playing for just one. We don’t know yet! So I’m trying to go with what makes sense for now, and that’s just … not yet done.
Please leave a comment below or hit me on Twitter @wchaplayoffs. Today is a travel day for me — KEF to BOS — but I’ll see stuff as I have time.