Sagarin says

[gumby013]

Has Sagarin published, or do we know, any of the details of what goes into the new Dimin_Curve method other than what is posted on his page with the rankings? (See below.) I did find an interesting hypothesis for what the Dimin_Curve method represents (see farther below), but I don't know whether it is valid.

Superficially, based on the little that I know of the two methods, I would vote for using the RATING synthesis of the Predictor and Dimin_Curve. I would like to know more about what goes into each, but I suspect Sagarin keeps that pretty confidential. I wonder if Gumby has any insight from his work with his rating system??

If I had to make a guess at what he is doing with the Dimin_Curve algorithm, I would say he is using a diminishing return curve to weight score differentials. Basically, it would give less incremental weight to scores as the differential increased. After all, there really isn't much difference between a 70 point blowout, and a 77 point blowout.

Example Weighting: (not saying this is by any means accurate, just an example)

Point Differential - Weight

0 - 1.00
7 - 1.25
14 - 1.45
21 - 1.60
28 - 1.70
35 - 1.75
42 - 1.775

That's my best guess. When applied to ELO principles it rewards teams for winning comfortably, but doesn't give them a lot of extra juice for piling on points.

[thebootfitter]

If I had to make a guess at what he is doing with the Dimin_Curve algorithm, I would say he is using a diminishing return curve to weight score differentials. Basically, it would give less incremental weight to scores as the differential increased. After all, there really isn't much difference between a 70 point blowout, and a 77 point blowout.
...
That's my best guess. When applied to ELO principles it rewards teams for winning comfortably, but doesn't give them a lot of extra juice for piling on points.

That makes sense to me. Especially if Predictor truly only uses the score margins with no factor to diminish the value of larger point margins. In some ways, such a metric would be preferable, methinks.

[runtheoption]

If I had to make a guess at what he is doing with the Dimin_Curve algorithm, I would say he is using a diminishing return curve to weight score differentials. Basically, it would give less incremental weight to scores as the differential increased. After all, there really isn't much difference between a 70 point blowout, and a 77 point blowout.

Example Weighting: (not saying this is by any means accurate, just an example)

Point Differential - Weight

0 - 1.00
7 - 1.25
14 - 1.45
21 - 1.60
28 - 1.70
35 - 1.75
42 - 1.775

That's my best guess. When applied to ELO principles it rewards teams for winning comfortably, but doesn't give them a lot of extra juice for piling on points.

That makes sense to me. Especially if Predictor truly only uses the score margins with no factor to diminish the value of larger point margins. In some ways, such a metric would be preferable, methinks.

English please, guys. CAS reads this thread.

[TAILG8R]

English please, guys. CAS looks at this thread.

FIFY - 10 chars

[missingnumber7]

We use the Predictor here not the mash-up that is the published order. Though as of mid-year 2013 the published order is now a mash-up of the Dimin_Curve and Predictor rather than the Predictor and ELO (BCS B***S***). So it's theoretically better than it was before when it willfully included a clearly inferior method of evaluating teams.

Were you sitting by the fire in your smoking jacket holding a pipe thinking in a British snooty accent when you came up with that last line?

[bisonaudit]

Were you sitting by the fire in your smoking jacket holding a pipe thinking in a British snooty accent when you came up with that last line?

If I were that guy I wouldn't be trying so hard.

[thebootfitter]

If I were that guy I wouldn't be trying so hard.

If you were that guy, I can't imagine that you'd be on Bisonville talking about Bison football.

[thebootfitter]

Well... I suspect that anyone outside of the few of us who responded will trust our judgment regarding what metric we use in our respective analyses. Gumby didn't really comment on which metric he'd like to see, but unless there's any future dissent, I think we can safely use Sagarin's RATING for internal consistency going forward instead of the Predictor.

I'm still trying to finish up my analysis comparing each week's actual game results against Sagarin. I had been using Predictor only, but I may see if I can easily include the composite rating as an alternative view. Not sure if I mentioned on here, but somebody over at AGS supplied me with the Sagarin results from each week throughout the year.

One thing in particular that I want to isolate is the prediction skill for just the limited number of FCS-FBS match ups that occur each year. I also want to try to determine the impact of treating all sub-DI opponents equally. Those are the two biggest criticisms that I hear when attempting to apply Sagarin ratings to FCS teams. Particularly regarding where they would rank among FBS teams.

[thebootfitter]

English please, guys. CAS reads this thread.

D'oh, sorry CAS and RTO.

One gud number better than one gud number and one bad number mixed. But two gud numbers mixed maybe even better.

[RadioBison]

Did anyone use any of Sagarin's methods in a bowl picking contest this year? I used my knock off system in the Bisonville contest, and it did pretty good until New Year's day. Then, everything fell apart. A couple guys picked just about every game the same as I did, but the confidence levels were all chosen differently.