Current ABOVE Goal Margin Multipliers
When I calculate BELOW, I use the following goal-margin point multipliers:
- Win in a shootout: 10 points
- Win in 3v3 overtime: 20 points
- Win in 5v5 overtime: 30 points
- Win by one goal in regulation: 40 points
- Win by two goals: 50 points
- Win by three goals: 60 points
For the regulation wins, I deduct any empty-net goals from the margin, including Friday night’s 5-2 Ferris State win over Bemidji State. This keeps things fairer in my mind.
I think that it should be obvious that any game going past 60:00 minutes should have fewer BELOW points awarded than a game decided in regulation — if you have to go longer, you’re not demonstrably better than the other team. This is the same argument that I made in previous seasons about overtime games: winning in overtime got you 0.75 in actual value, not 1.0.
A review of the 2016-17 outcomes to-date
Now that we’re 73% of the way into the season, I thought that I’d look at matters. Here are when games have been decided so far:
- Shootout: 8
- 3v3 overtime: 7
- 5v5 overtime: 8
- One goal: 30
- Two goals: 22
- Three goals: 27
Why did I decide to spend a couple of hours looking at this? I was running some fake simulations on Sunday night to test some new wrinkles in my by-series script and saw this:
For teams 200 BELOW points apart or more, the weaker team never can get to the two- or three-goal margin. (You can see that a handful of one-goal wins by the weaker team on the first night gets a handful of two-goal wins.)
Is this borne out by what we see in reality? After all, BELOW is designed to be an estimate of reality, and if it’s not doing that job, I have to make an adjustment.
The overtime results, in order, by bracket
Let’s consider all 102 results and the differential between them in each of the outcomes. I’ll highlight any that are interesting. Negative point differentials mean that the weaker team won. I’ve ordered these from lowest to highest.
- Shootouts: -364, -184, -101, -89, -58, +92, +233, +301
There doesn’t seem to be any real order to that set.
- 3v3 overtime: -335, -224, -89, -62, +20, +53, +257
There’s not much order there, too — it seems like getting to the new, funky overtimes gets us to randomness. That fits with how ABOVE is currently modeling these event.
- 5v5 overtime: -89, -38, -38, +56, +69, +70, +172, +239
Overall, it looks like the BELOW-favorite is a better pick in overtime, but it’s not by much. Overtime appears to be as much of a toss-up as ever, which reinforces the concepts that BELOW is an estimate and improbable things do happen.
Mean and Median BELOW through 102 games
Now I want to note that the average differential of BELOW so far this season is 128 points. The median — as many results lower as higher —is 110. That’s pretty consistent with the spread that we’ve seen in the standings to-date. Before we go on to regulation wins, let’s carry on to a simulation of a 110-point margin through ABOVE:
The BELOW-underdog wins a respectable amount of time, but this one is pretty heavily-weighted to the BELOW-favorite. But what I want to note are the line for Home: they did not win by more than two goals on the first night, and they barely won by 3+ on the second night. Does this match what we see?
- One-goal wins: -237, -154, -143, -141, -134, -119, -111, -84, -59, -47, -46, +4, +9, +21, +22, +35, +38, +38,+108, +113, +114, +143, +187, +233, +237, +272, +297, +354
The one-goal wins are strongly skewed to the stronger teams, and there are 12 games where the margin is smaller than the median, and the stronger team won eight of those. That seems to validate — with a small sample size — that the model is working. This is without doing anything like root-mean-square or other statistical fun that I will not bore you with (other than to show results).
- Two-goal wins: -387, -192, -164, -143, -114, -39, -33, +34, +41, +65, +87, +90, +98, +107, +124, +145, +159, +165, +215, +204, +209
Of the 22 two-goal wins, 15 were won by teams favored by BELOW, with eight of those by teams with a differential greater than +100. This is pretty much what you would expect. There have been five big upsets, but that’s going to happen.
- Three-goal wins: -118, -98, -48, -23, -20, 0, +4, +17, +46, +64, +66, +71, +80, +89, +98, +103, +116, +130, +148, +156, +172, +193, +194, +251, +265, +286
Only one team — Lake Superior over Minnesota State in Week 6 — with a differential greater than -100 won a three-goal game. Ferris State (-98) defeated the Mavs in Week 11. In fact, just five BELOW underdogs won by three or more goals.
In general, these results are what we would expect. Only that Laker win would be missed by ABOVE, and for me, a 1% improbability is okay in my book. The improbable is often possible!
Is the 3+ differential worth it?
In our highest BELOW differential-turned-blowout, Northern Michigan at Bemidji State, the Beavers got all of 10 points. Reducing it would give the Beavers just eight. Remember: BELOW assigns points based on who wins and what their expected value of winning was. If NMU had won that game, they would’ve been +50, not -10.
Here’s a table of BELOWs with and without the 3+ differential multiplier:
2016-17 Week 16 BELOW, Revisited Again
Above are the 2016-17 BELOW rankings after Week 12 of WCHA conference play (102 of 140 games played). The left column has three-goal multipliers; the right stops at two goals.
In eight of the ten cases, BELOW reverts to the mean a bit, as expected; the outliers are Lake Superior and Minnesota State, who were both involved in the three-goal upsets. Alaska and Lake Superior flip places.
In short: having the 3+ goal multiplier doesn’t make a huge change, but I’m happy to keep it in there out of momentum if nothing else. If I do drop it, I’ll probably go to something like 5-10-20-30-40. That recalculates everything and will generally draw teams closer in the ratings, as you can see above. I will consider this after the season unless there’s some huge reason to make the change.