Wednesday, November 25, 2009

Revised Power Rankings

I’ve noted my struggles with my current methodology in my Power Rankings that allows teams with multiple losses to be ranked higher than they ought to be based on having strong schedule strength and other data points.

In particular Oklahoma troubles me. With 5 losses they come in at 11th in my poll because of their high national rankings in Scoring Offense and Defense, and strong SOS. However, since they have lost virtually every game against a decent opponent, their SOS is helping them in my rankings while it has hurt them in their overall record.

I’m not alone in this conundrum – Sagarin ranks Oklahoma 7th in his predictor system.

Reader “Robert” points out in Oklahoma’s case their national stats are skewed by their games against Idaho State and Tulsa, in which the Sooners won by a combined score of 109-0, impacting both their national ranks in Scoring Offense and Defense. He also suggested I find a way to make multiple losses count more against teams.

I may have found a method to do just that.

Right now I count the Win-Loss record as a mere percentage – a record of 8-2, for instance, is a 0.80 score. What I have done here is use a multiplier for losses that works like this –

W/ (W+(L*L)) =W/L Value

Here’s how that works practically –

For an 11-0 team

11/ (11+(0*0) = 1.0 (the highest value, and unchanged from old formula).

For a 10-1 team

10/ (10+(1*1)) = 0.9091 (the same value, again, as the old formula as 1 times 1 equals 1)

For a 9-2 team

9/ (9+(2*2))= 0.6923 (here is the first change, where a 9-2 was 0.8182 before, it is 0.6923 now)

And, for a 6-5 team

6/ (6+(5*5)) = 0.1935 (A very BIG difference here, instead of the old 0.5455 the score is now 0.1935).

Using this methodology we get the following top 25 –

1 Alabama
2 Florida
4 Texas
5 Cincinnati
6 Ohio St.
7 Penn St.
8 Oregon
9 Virginia Tech
10 Boise St.
11 Georgia Tech
12 Pittsburgh
13 Oklahoma St.
14 Iowa
15 North Carolina
16 Clemson
17 Miami (FL)
18 Nebraska
19 Oklahoma
20 Arkansas
21 Mississippi
22 Wisconsin
23 Houston
24 Texas Tech
25 Oregon St.

To me, this “looks” better the last poll, even with Oklahoma at 19th.

Let me hear what you think.


Ethan said...

I am continually impressed by your work. That new list seems exactly to be my perception of how teams might fair against each other at a neutral site. I do wonder if the newest change (squaring the number of losses), might over-exacerbate the role those losses may play in a teams total power rankings. This would make it an order of magnitude more complicated, but what if you combined SOS with MOV, which may even out the huge MOV's against teams like Tulsa, and highlight MOVs against tougher teams. Just some thoughts. Thanks for your work!

Kevin said...

Definitely an improvement over the previous poll, Penn State at #7 is a reach. You're telling me on a neutral field against Ga Tech or Oregon to bet on Penn State? It might be out of the scope of what you're trying to accomplish here, but I think to get a more accurate picture you need to account for the SD and SO based on opponent quality. For example I discount a lot of what Penn State's defense has done because of who they're playing. Maybe weight the conference the opponent is from and adjust for points allowed - Allowing 3 points to LSU has to count more than holding Eastern Illinois to that same value.

Jams said...

I like the idea of weighing opponents individually rather than collectively, as Ethan said above.

Another possibility: something that's been done by Ed Gunther (TheNationalChampionshipIssue) and David Wunderlich (Year2 at TeamSpeedKills), which is to use the percent of opponent's average points scored/allowed, rather than the raw stats like scoring offense and scoring defense.

Example: let's say (and I'm just making these stats up) Texas is allowing, on average, 13 points per game against teams who collectively average 30 points per game. And TCU is allowing 12 ppg against teams who average 25 ppg.
Which one is a bigger accomplishment?

Also, I want to point something out: you are tweaking the formula so that it looks more like the human polls, or at least your opinion. Isn't the point of a statistical ranking to eliminate opinion? If you're going to tweak the numbers so that they match your opinion, why not just use your opinion instead of the stats?

Amos said...

I agree a lot with Jams. I really like Ed Gunther's work, though his power rankings he just uses the MoV after his form of SO and SD (which he discusses automatically adjusts itself for SoS). I think something combining his rankings on SO and SD with a little bit of what you're using could be pretty interesting.

I think it might be important to (and maybe you've done this) also run the new methodology on past seasons to see if it still passes the eye test, or is just working for this season's, where results aren't coming out the way we expect them too.

Robes said...

Not related to this post but Texas just gave up it's 32nd point of the night = no championship.

Henry Louis Gomez said...

Robes, I was thinking the same thing. Not a championship defense as we've seen statistically.