September 24, 2018

# SUR vs. ESPN’s FPI

The model I’ve developed and have been using for win probabilities didn’t have a great bowl season.  6-6 would probably earn a Poulan Weed-Eater Independence Bowl date (for the old timers out there).

It should be noted that when I talk about probabilities I’m not predicting a win (or loss).  I’m assigning a probability of a team winning.  Whether that team ultimately wins or not depends on whether they perform to expected levels.

Assigning a 61.9% win probability  also means that there’s a 38.1% probability that team will lose. Over the long term the results of a good model will fit those ranges – 61.9% (or close to that) will win and 38.1% will lose.

Still, it’s interesting to compare model results, to measure myself so to speak.  While my 6-6 record is less than sterling ESPNs FPI came in at 3-9 in the same 12 games.  In the three games that SUR and FPI disagreed on the winner SUR correctly picked all 3 (in green below).

It’s an infinitesimal sample size (the 12 games represent less than one half of 1% of the games the model is based on) that means nothing in the big scheme of things, other than I’ve had a decent start and believe I’m on the right trail, with a long way to go.

The biggest single game difference between my model and ESPNs FPI was our differing views on Georgia Tech vs. Mississippi State.  As anyone who reads this site knows, my model puts a lot of weight on yards and yards per play, two things that Tech is very good at and my model probably over rates Tech because of this.  However,  on this occasion my model was closer to being correct.

The Bowl season is chaotic and very unpredictable, in general.  After favorites won 75% of the time in the regular season underdogs have won 18 of 35 games (1 was a pick ’em, so there was no “favorite”) and I’m sure the ESPN FPI number is more accurate during the season.  I believe mine will be, too because there are more clear cut winners in the regular season (and the sample size is much larger).

My goal is to expand the model to include more games in the 2015 season by automating the process and am working through ideas to make that happen.  One idea I’m looking at is to provide probabilities for Power 5 Conference teams when they play other Power 5 Conference teams.  This will provide a larger sample, while focusing my time on games that people care the most about.  Not many people come to this site to figure out if Western Kentucky is going to beat Florida Atlantic.

Tomorrow, I’ll post my probabilities on the Oregon/Ohio State Championship Game.