football

Since I started closely following college football in 2012, I’ve always been irked by an immense amount of bias when talking about different teams. Even after the support shown for one’s own team, in this increasingly factionalized world of college football (CFB), fans are nearly required to support their division and root for the success of their conference. As much as I love my Buckeyes, I’m a fan of all Football Bowl Subdivision (DI-FBS) football, from the top tier Alabama vs Tennessee arms race all the way down to the obscure Georgia Southern vs Louisiana Lafayette–”the whole hog” as some in the industry call it. It is the combination of my frustration with the level of bias shown towards traditional power house teams (“blue bloods”) and my fandom for the entire sport that I decided to create a model to generate power ratings for each of the 131 FBS football teams (interchangeably referred to as a “computer poll”) using game results from all games that involve at least one FBS team. 

The Design

The key attribute I wanted from my model was a way to establish a baseline across all the teams. The model is not meant to be a definitive power rating and is definitely not meant to beat the spread. It answers the question no one had, but I went ahead and answered it anyway: “If aliens from outer space started watching college football without knowledge of anything else, where might they rate the teams?” In other words, this is the anti-AP poll. If one were to hypothetically plot how much bias a model has on the x-axis and how complex it is on the y-axis, it might look something like this:

Some models lean into bias, using metrics like recruiting rankings, preseason polls, and conference adjustments. Others lean into the variance of on-field results, but try to account for strength of schedule or cap win margin. If the AP poll were treated like a model, it would have the highest bias of any model out there. My model, Only Points Matter, is designed to be just the opposite; as little bias as possible, even at the expense of predictive accuracy, to create a baseline. 

The Model

The model itself is fairly straightforward: generate a set of power ratings for each team such that they minimize the total error generated between the projected spreads and the adjusted actual win margin throughout the season up to this point. Three things to note:

1. “Adjusted actual win margin” is the win margin adjusted for home field advantage (a value which is decided by again minimizing error.) 
2. “Error” is the squared difference between the projection and the actual.
3. Each game gets a weight assigned to it to tell the model how important it is to minimize the error associated with that record.
   • The first component in the weight is how many games the two teams combined have played against FBS teams, which is meant to significantly reduce the impact of games played against lower division schools (DI-FCS). 
   • The second component is a decay factor based on how many weeks ago the game was played. This allows the model to pick up on teams improving over time and not hold games from the beginning of the season against them too much (another value which is decided by again minimizing error.) 

That’s it. There isn’t a strength of schedule adjustment, a recruiting adjustment, injury adjustment, conference adjustment, there isn’t even a win-loss adjustment. This is purposefully designed to be an incredibly naive model, again just to create a baseline upon which caveats and story lines can be applied. 

It’s a bunch of math going on under the hood (code is in my GitHub repository if you are that curious) but this setup leads to the model effectively trying to balance the ratings of all the teams at the same time.

An Example

Take for example the end of the 2022 regular season, Texas has a rating of about 28, good enough for 7th place. “Wait a minute,” I hear you say, “Texas was 8-4! How are they #7?!” That’s a great question! We can look at two games Texas played that result in this: they trounced Oklahoma 49-0, and did the same to Kansas a few weeks later, 55-14. 

With Texas at 28.7 and Oklahoma at 16.3, the projected margin between the two on a neutral field would be about 12.5 points. The actual location-adjusted margin (LAM) was 46 points. The further down Texas’ rating hypothetically goes, the more error is generated by this game. The same phenomenon plays out against Kansas. All of the teams are connected, and the goal of the model is to find the minimum in the messy system of equations. For Texas, the result is a diagram like this, showing how each game contributes to their overall error, and why the minimum is found where it is (marked by the black line.) 

As a neat aside, the variability of a team can be thought of as the sum of error at its minimum. If a team wins each week by exactly what the model predicts, their total error will be 0. If they have blowout wins as well as blowout losses, the total error will be much, much larger. 

Is Texas actually the 7th best team in the country? Probably not! But to explain why not, you’ll have to use logic that goes beyond “they lost to abc team by x points.” I will leave that as an exercise for the reader.