It's funny, I would assume pg - as in "a plan for spam" pg - would be supremely interested in running some kind of Bayesian predictive model over YC candidates. Maybe the subtext of that quote was that he tried it and it didn't give any useful data.
An interesting conjecture is that numerically predicting startup success (defined as a high standard deviation of return on investment) might actually be impossible because any venture risky enough to get those kinds of results would fall outside the acceptable error bars of the predictive model. The equivalent in spam filtering would be if you wanted a system to show you only messages that were 99% likely to be spam, but still not spam.
I'm not sure if that actually makes any sense, someone feel free to jump in and tear it to shreds.