Intention of College Football Ratings

Retrodictive	Predictive
Tries to get best fit to current season's data	Tries to produce ratings useful for predicting future games
Determines which teams had the most impressive season	Determines which teams are strongest
Does not publish predictions	Might publish predications
Does not use data from previous season	Uses data from previous season early in year to avoid bad predictions from insufficient data
Early season ratings don't mean much because of lack of data (some systems don't publish early in the season for this reason)	Early season ratings are meaningful
Has no preseason ratings because there's no data	Might have preseason ratings
Might be based on win/loss record only	Systems serious about predictions usually use the margin of victory and maybe home field advantage
Judged by how well past results are fit. For example, what % of games were upsets based on the current ratings.	Judged by how well the predictions work. For example, what % of games were upsets based on the ratings from the day before the game.

Predictions are for entertainment purposes only.

Kenneth Massey's College Football Ranking Comparison now includes a final row giving the games per thousand that were upsets based on the current ratings. This "Ranking Violation %" is one way to compare retrodictive systems. See also On Ranking Violation Percentage by Jeff Bihl.

Eugene Potemkin has a better way of comparing retrodictive systems. (Be patient; this page comes from Russia!) This measure is also now included on Kenneth Massey's page as "Weighted Mistakes".

sum(games such that Rl > Rw) (Rl-Rw)*(2n-Rw-Rl)
-----------------------------------------------
                     1000

Parent Directory