Points Per Drive vs Power 5 Rankings

[ibeattetris]

I agree yet some on here want to throw out our P5 stats (UL) because they’re a “cupcake”. You also have teams like us who are very good on one side of the ball but not the other. Good luck factoring that in.

Yeah. That is just incredibly hard to do across the entirety of the NCAA in a way that is internally consistent with the stat. Should Alabama throw out their stats against Arkansas because they aren't good? Who determines what games don't count vs games that do count?

I don't disagree that having this weighted on SOS would add more information, but I'm trying to inform people on why AE 87 can't just causally add SOS data to this.

For people looking for crazy advanced stats, check out https://www.footballoutsiders.com/stats/fplus
F+ is just their combined stats. They have S&P and FEI. S&P is a metric based on "per play" metrics. FEI is based on "per drive" metrics. GT typically does really well in the FEI per drive metric (in fact, in 2014 they had to completely redo the metric because GT's was the best of all time by an incredibly large margin).

 

[smathis30]

I don't think people grasp how difficult factoring SOS would be to a stat like this. How do you determine the seeding for SOS? Do you use preseason rankings? Do you use current season data backfilled with last seasons data? Are you going to weight it on the strength of the opposing team def/off or just on the overall rank of the team? Once you begin to determine a SOS metric, you are now going to weigh the PPD up or down. Now how are we going to determine if the weight we determined was correct. What do we have to compare it against? Do we run this algorithm over previous years and then compare head to head of weighted PPD to see how well the weights compare? Or do we just look at the output and see if it compares well with other rankings?

I think I agree with AE here.

Its not that hard to do when you can run code and figure out how to weigh it and use linear regression. Depends on what you're aiming for as well. My rankings aim to guess against the spread as maximizing the total number of correct games is how my sheets gets the weights for all of its formulas

207-97 ATS in hindsight so far, 157-147 in predicting it overall, and my top 10 each week are shooting 35-25 so far which is what my main intent is. Bet on the games my spread is biggest from vegas's.

 

[AE 87]

Its not that hard to do when you can run code and figure out how to weigh it and use linear regression. Depends on what you're aiming for as well. My rankings aim to guess against the spread as maximizing the total number of correct games is how my sheets gets the weights for all of its formulas

207-97 ATS in hindsight so far, 157-147 in predicting it overall, and my top 10 each week are shooting 35-25 so far which is what my main intent is. Bet on the games my spread is biggest from vegas's.

Maybe you're right and App St really is a top 10 team.

 

[ibeattetris]

"when you can run code and figure out how to weigh it and use linear regression"
I think you lost a good portion of the audience here

I don't know if I think linear regression is an appropriate tool here unless we say PPDDiff is going to be used as a predictive measure (versus just a ranking measure). You definitely could attempt this though and probably see reasonable results. I am not sure if I agree though that PPDDiff by itself would be a good stat to use to determine who will win a given game, but I do think it gives clear impression on how well a team generally plays overall.

 

[smathis30]

Maybe you're right and App St really is a top 10 team.

Issue with App state is they have only played 4 games versus 6 for everyone else. Miscounting G5 teams when Wisconsin has played an easier schedule than Boise State for example, can cause disprortions as well. Nothings perfect, and with cfb having so little data at this point (unless you look at it play by play) you have to work with what you got. I stop using pre-season rankings after 4 games. Everything sorts itself out with SOS given time.

 

[AE 87]

Issue with App state is they have only played 4 games versus 6 for everyone else. Miscounting G5 teams when Wisconsin has played an easier schedule than Boise State for example, can cause disprortions as well. Nothings perfect, and with cfb having so little data at this point (unless you look at it play by play) you have to work with what you got. I stop using pre-season rankings after 4 games. Everything sorts itself out with SOS given time.

Yeah, my stat ignores Wisc for not playing enuf Pwr5 teams.

 

[iceeater1969]

"when you can run code and figure out how to weigh it and use linear regression"
I think you lost a good portion of the audience here

I don't know if I think linear regression is an appropriate tool here unless we say PPDDiff is going to be used as a predictive measure (versus just a ranking measure). You definitely could attempt this though and probably see reasonable results. I am not sure if I agree though that PPDDiff by itself would be a good stat to use to determine who will win a given game, but I do think it gives clear impression on how well a team generally plays overall.

When we did regression analysis using slide rules - u damn well were sure it was a good stat and wanted to measure how good. The prof had us run the cakes on one set of data. Net class he changed and we ran it again . And again- lesson learned it's a truth tool.