For those just joining, Amos is a statistical model created to predict the outcome of each NFL game. Amos uses team level statistics from 2005 to 2017 to learn what wins games, and uses that information to assign likelihood of the home team winning the game.
Amos is not alone the NFL prediction space; there are a number of sources for game predictions. I have identified Microsoft’s Bing Predicts (Bing), ESPN’s Football Power Index (FPI) and Nate Silver’s Elo (Elo) as benchmarks for comparison to Amos this year. The selection is due to the statistical approach to each of these prediction methods, which provides similar, but not exact, grounds for comparison.
Before viewing the model data below, it’s important to first point out the methodology used to track NFL season performance and its limitations. When a model predicts a home team win likelihood above 50%, this is counted as a prediction that the home team will win. Conversely, when a model predicts a home team win likelihood of below 50%, this is counted as a prediction that the away team will win. The models therefore increase performance by correctly predicting whether a home or away team will win and decrease performance by incorrectly predicting whether a home or away team will win. This is called a model’s classification accuracy. While this allows us to track season performance, it does not allow us to credit models that identify ‘close’ games (e.g., predicting a home team win likelihood of 51% and the away team wins) or penalize models for wildly off the mark predictions (e.g., predicting a home team win of 95% and the away team wins). In these instances, classification error may be more appropriate. Nonetheless, just as ‘close’ only counts in horseshoes and hand grenades, NFL teams don’t make it to the Super Bowl by almost winning games. In similar spirit, we’ve chose to use classification accuracy.
The below graphs keep track of three key pieces of information. First, cumulative season performance is tracked. This is represents the number of correct predictions by the number of games played. Second, performance is broken down by week. Similar to cumulative season performance, weekly performance represents the number of correct predictions by the number of games played in a given week. Finally, each models’ weekly predictions are presented. These predictions are each models’ assigned likelihood of a home team win.
Have thoughts on the predictions? See a missing game? Leave a comment below or send us an email at TrevorBischoff@gmail.com
See Amos’ predictions from the 2017 NFL season.