System Design Update

I'm writing this now so that I can refer to it in a few weeks when my system actually starts to work:

I've been thinking a lot about my system and my original goal in setting this up. A few things occurred to me whilst I was pondering the purpose of my proposed prognosticator.

The overarching goal was to create a 100% unbiased, 100% transparent system, and then to use that system to propose a playoff filled with deserving teams. That goal is still in place, but I feel that over the past couple of seasons I've been fiddling with it too much.

I started with essentially the same system that Ohio and other states use for their high school playoff seedings (I use Ohio because I live here and they publish their methodology). The main difference is that the Ohio system has multiple tiers of schools, from D-I to D-VII, and awards win points based on what division your opponent is in. While the FBS division could probably be broken down this way, it isn't, and so we would end up with Alabama beating East Carolina being worth the same as UGA beating Clemson. I think that we can all agree that these wins are not equal. 

Also, losses in Ohio are simply treated as 0 points. I get it; they need to have a relatively simple system that can be explained to parents when little Johnny's team doesn't make the playoffs at 5-5. But to me, and I expect to a lot of you when a huge underdog wins, that losing team should maybe be downgraded a bit.

They also award points based on the opponents' opponents' records, which we call 2nd level points. I have that, too, and I've added another level (yes, opponents' opponents' opponents' records). More data is gooder.

Ohio modifies points based on number of games played. My system does that as well.

Lastly, Ohio also ignores margin of victory. Again, I get it, but we really should consider MoV in college ball, so I do that.

This season I am going back to basics, with a few tweaks:

1.  As usual, teams get x points for winning games. 
1. Instead of 1pt for a P5 team and 0.5 or 0.75 for a G5 team, I've added a gradient factor. Winning teams will earn the point differential between themselves and the team they beat. There are 130 teams in FBS this season, and there will be 130 rankings. The difference between 1 and 2 is 0.007692, inversely applied (#1 is worth 1pt, #2 worth 0.9923 and so on). So, if you beat #1 you get 1 point, and if you beat #130 you get 0.007692 point.
2. Instead of 0 points for a loss, you lose 0.007692 point for for every rank below #1 of the team you lose to. If  you lose to #1, it costs you nothing; if you lose to #2 it costs you 0.007692, and so on. Losing to #130 costs you 0.9923 point.
3. This was the best way I could come up with to approximate the divisional point differentials in the Ohio system. Upsets reward the winner and penalize the loser based on the ranking difference. 
4. As is usual and unique for my system, you lose a full point for beating an FCS school, and new for this year you lose 2 points for losing to an FCS school.
   
2.  Ohio has 1st and 2nd level points
1. I have 1st level points as described above.
2. I have 2nd level points that are the sum of the 1st level points of the team you beat.
3. I have 3rd level points which are basically the 2nd level points of the team you beat.
3. Margin of victory matters. I will use the College Football Reference SRS number. Because that number gets so big (and small, as in negative numbers), I will be modifying it by a factor that reduces its overall influence. In other words, MoV matters but not as much as winning and losing.
1. I won't know the exact modifier until my first week of real rankings because I don't know exactly how many points the system will generate for 1st-3rd level points.
   
4. Add up the 1st, 2nd, and 3rd level points for a master score. Add in the modified SRS number. Biggest number wins. That's it.

Things I've used that I won't be using: 

• My own SRS, as the iterative calculations were really killing spreadsheet performance and giving spurious results, plus I'm not sure I had the formula set correctly.
  
• Average ranking within the different levels. I think the sliding scale of points will accomplish the same thing.
• Any sort of weighting other than reducing the influence of the SRS.
  

As I mentioned in my earlier blog post, I have to use an iterative function on the rankings differences. Points make the ranks different, rank difference makes points change, points change make ranks change and so on. I found a good number of iterations with a point variance that gives repeatable results.

Anyway, this is how it will work starting week 4. For now, I am sticking with the weirdness that incorporates my fake pre-season numbers!