Hey everyone! After a busy summer, I have made my way back from the desert in time for the return of the NFL tomorrow night. As usual, my first order of business is to blithely act like I can predict who will win the week 1 games based on team performance from last year.
The model is very straightforward. The first game last year (2013), for example, was the Baltimore Ravens traveling to Denver to face the Broncos. In 2012, Baltimore won 10 games and outscored their opponents by 54 points over the course of the season. Denver won 13 games and had a 192 point advantage. You could use either wins or point differential as a predictor of strength (or both, I suppose), and we want to know which team is more likely to win. I like to use point differential. As it turns out, Denver won. As you can see in the link above to last year’s post, that’s what the model predicted. In fact, the model picked the correct team 11 times out of 16, which isn’t too shabby.
What’s odd about the model is that only the home team’s strength is statistically significant – the away team’s performance last year has “nothing” to do with the outcome statistically speaking (actually, if you use point differential the away team’s weight is about half the home team, and if you use wins it’s about a third). Incredibly, if you only use the away team’s strength, it isn’t statistically significant. Apparently you can get pretty far just by knowing one team’s ability, roughly speaking, and who’s playing at home. This has been true as long as I’ve been running this regression. So to make this year’s predictions, I’m going to use a model based on the home and away teams’ point differential last year and an intercept that reflects home field advantage. In theory the away team doesn’t matter, but I just find that a little hard to believe in a practical sense.
For example, tomorrow night we have Green Bay traveling to Seattle. Normally you would say that this should be a close game; the Packers tend to be pretty good and Aaron Rodgers isn’t the kind of guy to fold just because he’s on the road. But all the model cares about is that Seattle outscored their opponents by 186 points last year while the Packers were outscored by 11. Here’s a little table with the games and the home team’s predicted win probability.
|home||away||home win prob|
The last thing to mention is the Hilton Supercontest. I use it as one of a couple benchmarks to see how well my models do. This first week isn’t really part of ‘the model’, and it isn’t meant to be used to pick against the spread. But in the interest of getting a full season of picks in, here are the five games where the model prediction differs most from the lines, according to my eyes: Dallas +4.5, Carolina +2, Miami +5, Oakland +5.5, and Buffalo +7. The Cowboys and Dolphins are home underdogs that the model thinks should be tiny favorites (less than home field advantage, but not getting over a field goal). Carolina is a road underdog that the model thinks should be a favorite (the Bucs only have a 43% chance to win). And the Raiders and Bills are road underdogs as appropriate, but the model thinks they should be closer to the home field number of +3 instead of at 5.5 and 7. Take those with a big grain of salt.
Whoo! Back to football tomorrow!