I have done my share of speculating about "what if" and wondering how a close loss to a good team costs your team more ranking spots than the same close loss costs another team.
I have also posted about good losses and questionable wins before. If you've been following along, you'll recall that I was trying to quantify losing points for a loss vs just getting zero points. I had it this season, but when I tried to run it it broke my spreadsheet. Now I finally think I have it figured, so I'll explain and then show you what the "updated" system came up with.
What is Happening
If we leave each FBS loss as zero points, we're saying that losing to Georgia is the same as losing to UMass. I'm sure we can all agree that the 2 losses are not really the same: everyone should lose to UGA, no one should lose to the Minutemen. How do we make it so that you lose zero points for losing to #1 and a full point for losing to #130, especially when the rankings between the top and the bottom shift every week?
We do the same thing for losses as we're doing for wins. If you beat #1, you get 1 full point, and if you beat #130 you get 0.0 points. The value of each ranking step in between is 0.007751 (1/(130-1) works; that's 1/(n-1) in order to start at 1 and end at 0).
If we take the inverse, that is losing to #1 costs you zero and losing to #130 costs you a full point, with graduated steps between of 0.007751.
When I ran this the first time today, it gave me UGA at #92. I'm going to let you in on a little secret: UGA isn't the 92nd best team in CFB. I'll let you in on another, less well known secret: in science, stats, and math we actually do have an "eye test" though we call it the "sniff test." The sniff test is used when you achieve a result and need a quick take on the potential correctness of that result. For example, if we do a study over time and our results give back negative time, the study does not pass the sniff test. Like meat that's gone bad, it smells wrong.
UGA at #92 smells wrong, regardless of what my system thinks. When I looked at the underlying data, all I could find was that Georgia's opponents have overall good records, but that their 1st level points were negative. I thought this was odd until I saw Oregon at #1. Then I realized 2 things. 1, the 3rd level points had far too much weight, and 2, the "average" good team plays generally bad to very bad teams. A good example of this would be Arkansas. The Piggies are 7-3 with 2 decent wins (A&M and MissSt), an FCS win, and 4 wins against meh teams.
For the ranks to pass the sniff test, I removed the 3rd level points entirely, and I reduced the influence of 2nd level points to an unmodified level. I could add the 3rd level points back in once I figure the proper modifier; the issue is that there are so many teams that contribute to the 3rd level it will overwhelm 1st and 2nd level points without the modifier and I don't know what the modifier should be. I'm comfortable with 2 levels because at 10 games played we get 100 teams contributing to the 2nd level.
How Did it Change?
Let's take a look:
Nothing super surprising except maybe OkSt and MSU?
What do you think? Keep the current calculation, or use the new one that accounts for relative rankings for wins and losses?
I'm partial to the new way.
No comments:
Post a Comment