Relatively inefficient but perhaps unbiased (in a statistical sense), sparse new data is less trustworthy than relatively more efficient old data that is adjusted for compositional changes.
If you think about the determinants of a state's vote, the composition of the electorate is a huge factor (there would be no such thing as a "red state" or a "blue state" if it weren't). Nailing sub-groups in a state's electorate with old data is a pretty reliable way to deal with that portion of the variability in the outcome.
Economists (and people in general) are trained to worry a lot about bias. And that's very important, obviously. But when you're dealing with data points that trickle in, you really need to worry about the efficiency of an estimator too. Nate Silver does well because Nate Silver knows that.
(Motivated by Russ Roberts's curiosity [which yet again is followed by a weird Krugmania digression... why does this happen so much with their posts???] - HT Brad DeLong)