The long layoff, varied levels of motivation, and (usually) close-to-even matchups make them more difficult to predict than you would think. Sample size is also a large issue (unavoidably so, of course); trying to rank every team on the basis of games against 1/10 of all possible opponents, most of which are within conference, is going to inevitably have a lot of noise in the signal.

For our bowl pick’em here, I based most of my picks purely on my set of Bradley-Terry rankings (and all four of the ones I went against the computer, I was wrong), and while running the numbers I saw that if the win probabilities I calculated are correct, I’d expect to get about 22-23 right on average. (I think the final tally was 21, going by the computer and not my picks – well within the realm of reasonable results.)

I do agree that the BCS computers could use some improvement (Billingsley starts with last season’s rankings as a “preseason ballot” and has all sorts of ad-hockery attempting to mimic human polls; none of the formulas except Colley’s are public, and Colley has the issue that beating a bad team is worse than not playing the game at all). But I wouldn’t necessarily use their bowl-game performance as evidence.