Stats First 2017 PPDvsPWR5 Rankings

Sideways

Helluva Engineer
Messages
1,589
Okay, so once again, this is a calculated stat based on games against power 5 opponents. I only calculate it for teams which have played more than 2 Pwr 5 opponents.

In the past I have ranked teams based on Diff PPD, their Offensive PPD minus their Defensive PPD. Motivated by the Domination Ratio thread, I've also ranked them based on the Ratio of OPPD/DPPD (again only Pwr5 opponents).

Sorted by Differential PPD
View attachment 2987

Sorted by Ratio of PPD

View attachment 2988

By this stat, our Offense is ranked #13 and our D is ranked #23.

AE, man you are making my head hurt with all this data. I didn't go to Tech and college algebra was as high as I got in math. Just interpret it and tell us what you think. Oh wait, I see you did later. Kirvonte better get well soon. OK I can understand that.
 

AE 87

Helluva Engineer
Messages
13,015
Okay, I've updated the PPDvPWR5 rankings for yesterday's games. Here is a comparison of the top 25 from AP, Coaches, PPDvPWR5 (Off PPD - Def PPD), and YPPvPwr5 (Off YPP / Def YPP).

upload_2017-10-22_20-36-8.png


The statistical rankings prefer Washington and tOSU over TCU and PSU. They also like GT. UCF and USF aren't ranked in my statistical rankings because they haven't played 3+ Pwr5 opponents.
 

AE 87

Helluva Engineer
Messages
13,015
Okay, I've updated the PPDvPWR5 rankings for yesterday's games. Here is a comparison of the top 25 from AP, Coaches, PPDvPWR5 (Off PPD - Def PPD), and YPPvPwr5 (Off YPP / Def YPP).

View attachment 3022

The statistical rankings prefer Washington and tOSU over TCU and PSU. They also like GT. UCF and USF aren't ranked in my statistical rankings because they haven't played 3+ Pwr5 opponents.

Okay, there's an error with this. I'm working to fix it.
 

AE 87

Helluva Engineer
Messages
13,015
@AE 87 can you update your score predictions?

upload_2017-10-23_12-58-47.png


Again, these predictions assume 12 drives/game and are calculated by averaging OPPD and Opponent DPPD. So, it's a very raw "prediction."

What makes GT difficult to predict is how bipolar our D can be. The first half vs Wake and the second half versus TN were awful, while much of the rest of the time we've been really pretty decent.
 

ibeattetris

Helluva Engineer
Messages
3,545
During the discussion in that thread some, raised the question of whether a yards/play ratio might be a better stat than the yards/game ratio. I began to calculate that using only games against Pwr5 opponents. Limiting the opponents to Pwr5 helps adjust for strength of schedule some, but I think that we get closer to comparability when teams play at least 3 Pwr5 opponents.
I think there is a reason the football outside F/+ combines the FEI and the S&P. Per drive and per play metrics add different dimensions to the discussion and help paint a better overall picture of an offense.
 

AE 87

Helluva Engineer
Messages
13,015
I think there is a reason the football outside F/+ combines the FEI and the S&P. Per drive and per play metrics add different dimensions to the discussion and help paint a better overall picture of an offense.

Any idea how to do that with these raw stats? Also do you have an opinion/argument on whether subtracting defensive ypp or ppd from offensive or dividing is a better way to normalize?
 

smathis30

Ramblin' Wreck
Messages
732
Any idea how to do that with these raw stats? Also do you have an opinion/argument on whether subtracting defensive ypp or ppd from offensive or dividing is a better way to normalize?

they do that to help normalize for SoS. for my personal formulas I multiply the raw score by 8%*number of standard deviations the avg opponent is, which would be difficult for you to do unless you have the games separated. Also out of curiousity, is the PPD the yearly point per drive avaerage or the average of every PPD per game?
 

AE 87

Helluva Engineer
Messages
13,015
they do that to help normalize for SoS. for my personal formulas I multiply the raw score by 8%*number of standard deviations the avg opponent is, which would be difficult for you to do unless you have the games separated. Also out of curiousity, is the PPD the yearly point per drive avaerage or the average of every PPD per game?

IIuc, football outsiders already normalizes for SoS for both their FEI and S&P+. They combine them for their "official" rankings in Fplus because they think it gives them a better assessment of actual strength of the units. I also think they use/recommend fplus for prediction, but I could be mistaken.

You're right, that I can't adjust for SoS because I don't look at games or drives individually. I use cfbstats. I download the Rushing Offense, Passing Offense, FGs, Punting, Fourth Down Conversions, and Turnover margin tables against Power 5 opponents.

Drives = sum(Rush TD's, Pass TD's, FG Atts, Punts, Failed 4th, Lost Turnover)
Points = sum(7*(Rush TDs + Pass TDs), 3*FGs Made)

Comparable calculation for D.

So, end of half drives and safeties don't get counted as drives. Also, Defensive and Special Team Scores don't get counted for points.

While not sophisticated, I think that only counting power5 games and only ranking teams that played at least 3 does an alright job in adjusting for strength of schedule. I posted the table comparing my PPDvPwr5 ranking with the AP and Coaches as at least making an argument on its behalf.

The question I was asking in the post to which you responded was whether there was a way of combining raw yards/play and points/drive stats to improve this poor-man's predictor.
 

ibeattetris

Helluva Engineer
Messages
3,545
Any idea how to do that with these raw stats? Also do you have an opinion/argument on whether subtracting defensive ypp or ppd from offensive or dividing is a better way to normalize?
With raw stats? I don't think so. Unless you tried normalizing then applying different weights to the stats based on some performance metric. How weights are chosen is probably the most interesting and unique part of the process. This can be done with manual tweaking or through a machine learning process (there are quite a few tailored to this). I'd start with just taking the average rank and ordering the average rank to start.

Subtraction vs division. Depends on what you're trying to express, and maybe both are important. 3ypp Off 1ypp Def vs 6ypp Off 4ypp Def. The ratio changes when you divide, but the differential remains the same. That is another metric where I think tracking both and seeing which is a better predictor would work.
 
Top