Introducing Massey-Peabody College Football Ratings
For the first year, Massey-Peabody Analytics will now provide weekly college football ratings and picks. This post will explain the basics of the rating system. Right now, all our college football content will be posted on the blog, but we should have a new section on the site in a week or two that will house all our CFB ratings and picks.
The framework we used to develop the ratings is similar to the one we use to rate NFL teams. Our model is built from play-level data. We use only four different statistics (rushing, play efficiency, play success, and score efficiency), but the strength of these is that we contextualize them in a number of ways:
- *We adjust for home field.
- *We weight plays based on leverage. The weights are asymmetrical. They effectively de-weight garbage time plays, but do not give additional weight to late & close situations, as plays in the 4th quarter of a tie game are not any more predictive of future performance than plays at the start of the game.
- *We norm each stat for opponent quality.
- *We norm for down/distance/field position. A 1-yard run on 1st-and-goal at the 2-yard line is worth more than a 1-yard run on 2nd-and-10 at midfield. Yards become more difficult to gain the closer an offense is to the opposing team’s goal line.
- *We tailor our stats to be predictive, rather than explanatory. While we like how other smart analytic types (such as Football Outsiders) use the framework of expected points gained/lost on a play level (this is WAY better than using yards/play), we think it’s a great way to explain performance but not the best way to predict future performance. We account for the nonlinearities of yards gained on a play level, and create statistics that are less noisy. The predictive value of four carries of 20 yards each exceeds the predictive value of one 80-yard rush.
- *Since team performance exhibits non-stationarity, we weight recent performance more heavily.
The approach we take is Bayesian. We estimate priors for each team-season-statistic for both offense and defense using previous season’s performance as well as recruiting, returning starters, and other publicly available data. As more games are played, the weights on the priors decrease, but they still have a positive weight throughout the season.