Here is a great site that tracks the season-long performance of the various NFL prediction models. It also summarizes their weekly predictions. (Thanks to Andy for the link). Model accuracy records from previous years are also available. I will add its link to the right so it can be referred to as the season progresses.
There are many competing statistical models, but it looks like the vast majority don't do very well. Fortunately, it appears the efficiency model used on this site ranks among the very best of the models. So far for 2007, the NFL Stats efficiency model would rank 3rd out of 61. (And if you nitpick, you would notice that the only two models with better records have either attempted predictions for fewer weeks or have possibly cherry-picked games to predict.)
The records for 2006 were strikingly lower. It appears to confirm it was a difficult year to predict with many upsets and with a historically weak average home field advantage. With its 62.56% accuracy rate, the NFL Stats efficiency model would have ranked #1 out of 58 (by a whisker).
Interestingly, there appears to be a lot of churn among the best prediction models. The best models one year aren't usually among the best in other years. This may indicate many of them are benefiting from a significant amount of luck. With so many different models, there are bound to be some that do well not because of predictive insight, but due to randomness. The models with true predictive ability would be the ones that rank near the top from year-to-year.
You can show your support for NFL Stats by sending an email recommending its inclusion in the rankings. [Edit: Never mind. He can't include me unless I predict scores as well.]
- Home
- predictions
- NFL Prediction Tracking
NFL Prediction Tracking
By
Brian Burke
published on 11/03/2007
in
modeling,
predictions
Subscribe to:
Post Comments (Atom)
I wrote an e-mail but was told that you could not be included on the list because you only predict winners and not results. :(
You could always run your stats through a linear regression model to predict the result.
I do have a linear model for prediction point differentials already that I ran for most of 2006. I think it works as well as any other, which doesn't say very much. Scoring in the NFL is so irregular, non-normal, and non-linear. Unlike baseball the points are not in increments of one. There are a lot more 3 point wins or 7 point wins than 1 point wins, which blows the assumptions required for linear regression. Plus, there is always the "trash-time" effect that would be so hard to quantify. I do have some ideas for approaches to overcome these problems, but I'll probably have to work on it in the off-season.