2015 College Football Model

By Michael Gertz
Thursday, September 3, 2015

Prior to 2015, our predictive team ratings model treated all teams equally coming into each season. In an effort to make the ratings as accurate as possible, and more useful early in the season, we have overhauled our ratings model for the 2015 season.

Predictive Ratings

Like before, the Predictive Team Ratings are based on an optimization process to estimate the most likely combination of all team ratings together, using game scores and locations to rate each team's quality. But beginning with the 2015 season, they now come into the season with a different initial assumption. Rather than assuming any rating is equally likely for any team, the model now attempts to predict each team's quality as accurately as possible even before the season begins.

To rate teams coming into the season, we ran a multivariate regression to analyze which variables have predictive value before a season has begun. While a team's rating from the previous season was expectedly a big factor, the result was that the team's recruiting class rankings (from Rivals.com) also provided a similar level of predictive power. Ultimately, the model uses a weighted average of a team's rating the prior year and its recruiting class rankings from each of the 4 most recent classes.

Before the season, each team is assigned an expected team rating. Based on how accurately those preseason ratings predicted end of season ratings in previous seasons, a standard deviation is established to estimate odds that a certain team's rating will end the season a given distance from its preseason rating. Therefore as the season progresses, the model now multiplies the odds that each team's rating varies from their preseason rating by a certain amount with the odds that each game's outcome would occur given the ratings of the teams involved. Prior to 2015, only game odds were multiplied and not team rating odds in the process of optimizing team ratings.

W-L Resume Ratings

The Won-Lost Resume Ratings model remains the same as in its debut in the 2014 season, but its framework more closely relates to the new predictive model than the old one. It quite simply works the same way as the new predictive model but while ignoring any data other than which team won each game.

That means that the W-L resume ratings don't use any preseason information, and only use the odds that each team would win a game rather than the odds than they would win by a certain amount of points. Whereas the predictive model comes into the season with a normal distribution assigned to each team's rating with the mean being their preseason rating, the resume model begins the season with every team being assigned a normal distribution that has a mean of 0 and a much larger standard deviation than the predictive model.

The resume ratings are not intended to predict game outcomes or judge team quality with accuracy like the predictive ratings. They are merely meant to be used to compare to the College Football Playoff selection committee rankings. The resume ratings essentially answer the question "How impressive is a team's Won-Lost record if you ignore any information but W-L records themselves for fairness's sake?". They sum up the component of the CFP selection that is purely a reward for winning tough games rather than just actually being a good team.

For a more in-depth discussion of what each of the ratings models represent in relation to the College Football Playoff selection committee, see our previous article on Optimizing College Playoff Selection