Wednesday, January 26, 2011

Evaluating with Bowl Results, Part II

Well, unfortunately A&M lost the Cotton Bowl to LSU, so I lost the competition against the BCS rankings. At least, I lost when only considering the Top 25. Overall, among bowl games with Top 25 BCS rankings (before the bowl games were factored in) my choices were 9-7. The BCS were 10-6, so my algorithm was only behind by one result. Looking at all bowl games, the algorithm fared much better, going 21-14, picking 60% of the games correctly. Given the erratic nature of college football, I'm pretty happy with this, but perhaps I can tweak the algorithm to do better!

I would still like to compare my results against the percentage of upsets over the whole season. If and when I do this, I'll comment on the results here.

Happy Off-Season!

Thursday, January 6, 2011

Evaluating with Bowl Results

I really enjoy watching bowl games. This year, I found an extra reason to want to pay attention: I can pay attention to the number of games that my ranker "guessed" correctly. Even better, I can compare these results to the BCS rankings.

Unfortunately, the BCS doesn't rank all the teams, but I can look at those games with BCS top 25 teams (ranked before the bowls) and compare those results to my ranker's guesses.

Last week my algorithm was doing pretty well! I was a few games up on the BCS rankings. The last few days, my teams have lost where I had them ranked higher, and now we're tied: 7 correct and 6 incorrect for both rankings, with three games to go. For two of these games, we agree: BC vs Nevada and Oregon vs Auburn. For LSU vs A&M, however, the story is different. My ranker believes in A&M, but LSU is nationally ranked higher.

That game happens tomorrow night. Let's go Aggies! :)

Overall, my ranker's doing pretty well: 17 correct picks, and 13 incorrect.

Why should we care only about bowls, though? Wouldn't it be best to count the number of upsets over the whole season? Yeah, I think so. I should do that... (Note, another improvement I should add for next season.)

Actually, it might be best to find the ranking that minimizes the number of upsets over the season. Given an entire season, what's the best way to find this? I have the suspicion that there is no (known) efficient algorithm for this. (Does anyone know whether this is NP-hard?)

Perhaps I should tinker with my parameters to try to minimize these upsets, but keep the same basic algorithm? Sounds like a good project for the future! :)

Until that time, I'm rooting for the Aggies!