You may have heard of ESPN's advanced statistical ranking system, known as FPI. There is some debate about it, mostly because ESPN doesn't share many details about how it is calculated. At one point, FPI had Mississippi as the best team in the nation, and it is currently ranking Southern Cal as the sixth best team in the nation, despite their 3-3 record. I decided to investigate how accurate of a ranking system it is. The next paragraph is a brief explanation if you aren't familiar with FPI.
FPI assigns a point value to each FBS team, so it is able to rank all of them from 1-128. Each team receives a score related to what is theoretically an average FBS team. For example, MSU has a score of 14.7, which means that this system believes that if MSU played an average FBS team enough times, MSU would win by an average of 14.7 points.
Since it was way too much work to calculate every single result we have so far in college football, I limited the results below to SEC games only. FPI's record of predicting the winner of games and the winner of the spread is below. I used the spreads from Football Study Hall's weekly picks. FCS games were omitted, because FPI doesn't rate FCS schools (Sagarin rankings do). I added 3 points to whoever the home team is, because that is considered the standard amount to add.
In week one, the Georgia spread was nullified due to weather. During week two, FPI's prediction on the South Carolina vs. Kentucky game was the same as the spread. These games are omitted from the spread column, but not the game winner column.
|Week||Game Winner Picks||Spread Picks|
|Total||43-14 (75%)||29-26 (53%)|
75% accuracy in picking a game winner isn't bad, but it isn't anything exceptional. You should be able to get around 50% if you pick the winner randomly, and some of the games, such as MSU vs. Troy, don't require much skill to pick. 53% of correct spread picks isn't bad either, but that isn't good enough to warrant using FPI as a gambling tool. The best way to evaluate FPI would be to wait until the end of the year, and calculate how well it predicted every game, not just SEC ones.
It would be interesting to see how this ranking stacks up against other systems, like the Sagarin rankings, Bill Connelly's S&P rankings, or Ed Feng's The Power Rank. Also, some of the margins of victory FPI predicted were very close to the spread, so I would be interested to see how predictive it is when it disagrees with the spread by 5 or 10 points. Another issue is that the spread to a game isn't agreed on everywhere, so the results might be slightly different if I used another source to get the spreads. The results would also better reflect the quality of FPI if I calculated the result for every college game. For now, FPI is just a topic to debate about, not an exceptional projection system.