Beta_Rank's 2022 Projections
I'll admit to viewing projections as a necessary evil. I find the in season data to be much more interesting and revealing about football, but this attitude certainly came back to bite me a bit last year; where the projection model performed poorly and the in season model did extremely well. So the projection model has been further revised and the fit improved. So lets run through what we ended up with.
2020 Data has been Smoothed
Last year was such a weird year in the data without the needed number of non-conference games to allow for a good rank order. You have CFB powers who recruit well having the worst years we've seen in ten years of data and very good G5 teams suddenly looking amazing. So what to do? I smoothed the data for 2020 that went into the model with 2019 and 2021 data. This helped clean up serious outlier problems from the 2020 season.
2021 Data does not Include the Non-Playoff Bowls
We got through the 2021 regular season virtually unscathed other than a couple of wonky Cal games against Arizona and USC, but the bowl season was a Covid crapshoot. The data was awful, with many teams down significant numbers of players and this had outsized effects on the final ranking if included; since these were incremental non-conference games. One of the main drivers that will cause differences with other 2022 projections is whether they included the bowls in their final 2021 rankings.
The Coaching and Development Variable(s) Dropped (Some) Outliers
One of the unique things about these projections is trying to assign some school specific effects that explain performance beyond recruiting and what you did last year. This really helps to fit the service academies in particular. This also helps us to understand the CFB production function a bit better:
And yes I think it's fair to think of it something like a Markov chain that would be extensible back to when teams first started and you only have C&D and Recruiting; if we had that data. The simple version of the statistical model looks something like this though:
You will notice that C&D dropped the temporal subscript and that is a function of how we have to approach the data. Coaching and Development and Recruiting are HIGHLY colinear so you can't just look to pick up the residual in a regression without C&D and assign it to coaching. That would be kind of dumb considering how much of the C&D variation would be picked up by Recruiting. So we have to run school specific controls that look to account for how much schools add/subtract from their scores each year outside of Recruiting and Returning. It's problematic because you can't fit a school/year version of this and hope to have any kind of useful signal. You have to bite the bullet and accept you will capture the most recent window of C&D. Down the road I will make this a rolling window where we run on the most recent 5 years worth of data only. I dropped the worst year for each school in the model to remove outlier "catastrophe" years that are only explained by things outside the model; usually to do with injuries, coaching changes, etc...
Recruiting or Coaching and Development?
The nice thing here is that you get an idea of how much each drive CFB production. I feel like the CFB twitterverse has devolved into binaries on this where "Stars Matter" vs. "Stars Don't Matter" mostly talk past each other in uninteresting ways. A positive person would look at the results here and say "Wow they are both right." I am not that person, they are both wrong because they took it to extremes. The largest positive contribution from recruiting in the model is Alabama who gets a +1.73 added in for their score (not all model contributions are positive) but the largest C&D contribution is Navy who gets +1.51 added to their score. I like to point out to people that assume the secret to Nick Saban's success is recruiting that they are missing a big part of the picture. The secret to Nick Saban's success is everything. He leaves no stone unturned. Alabama grades out at #4 in C&D, something you often miss in some of the, in my opinion, very poor attempts to measure C&D through the NFL draft or just looking at recruiting rankings vs. final rankings. You can't catch Alabama with Recruiting or with C&D alone; he's already doing both better than you.
What Matters in the Projections?
To give you all some simple business rules to look at the projections that are USUALLY true in the results
Recruiting and C&D > Prior Year Performance
Returning Production matters more for offense (and that means...)
Defense is more stable as a predictor YoY regardless of returning production
So that is about it. I have some personal reservations here. I think LSU is high, but the defense was pretty good last year and the recruiting is still very good. I'd personally have them in the top 20. Oregon is getting a good size bump from recent recruiting efforts, but for the most part they have a pretty talented roster for Dan Lanning to take over. Coastal is way too low, suffering from having a decent G5 defense, low recruiting rankings, and still a short track record of success.