First off, apologies for not getting this out before week zero (but does week 0 really even count anyway)?1 I also missed the week 1 start yesterday, but from here on out you can expect model updates every Thursday for the upcoming weekend’s games.
About the Model
I developed this during last season (pre-Substack), and fine tuned it a bit his year. It is essentially based on linear regression modeling taking into account various metrics including rating differences between opponents - such as FPI, SRS and FEI - winning percentage, post-game win expectancy, whether or not a team is ranked in the AP top-25, point-spread (currently attained from DraftKings), and team Elo.2 These metrics are run through some linear regression modelling, plus a few statistical adjustments, and voilà! What results is a projected winner and point spread for each game. I first began testing out this model in the back-half of last season, and it did quite well. Hoping the adjustments I made leading up to this season improves it even more.3
Week 1 Predictions
Here are the model’s projections for week 1:
You may be asking what the “Excitement Index” is. Basically, this is a way of quantifying how exciting a game will be relative to the rest of the games that week. It takes into account how close a game is expected to be relative to the rest of the games that week. It also adjusts up 10% if the home team is predicted to win on an upset, and 15% if the away team is predicted to upset their foe on the road.
Going Forward
I will update the model and post predictions every Thursday for that week, and if necessary make an update before the bulk of the games start on Saturday. You can find this model at all times on my Substack homepage here: College Football Predictions. Don’t forget, you can also keep up with my College Football Elo Ratings here: College Football Ratings. If you missed my article discussing my Elo rating system, give it a read here: 2024 Pre-Season College Football Rankings
Enjoy the games this week and this season!
FSU fans would like to thin that it does not
I utilized CollegeFootballData.com’s R package CFBDfastR to collect many of these stats. Huge thanks go out to them and the valuable package they have put together!
Model statistics in development indicate improvements, lets hope that translates to the real world!
I’m pretty surprised by those predicted point spreads - they appear too extreme at a glance