According to an op-ed in today's New York Times, a couple of physicists (J. Richard Gott and Wes Colley) have discovered that in "swing states" in US presidential elections, the median of the results of polls conducted in the weeks prior to the election is a good predictor of how a state will vote.
That's true -- but it's also incredibly obvious. Asking a sample of people how they're going to vote is apparently a good way of predicting how they're going to vote. As for why they take the median, it's because the median is more robust than the mean; there are some highly partisan or incompetent pollsters out there. The fact that it's newsworthy says something, although I'm not entirely sure what.
This was apparently something they knew in 2004.
As for the article's claim that according to the polls, Clinton would beat McCain but McCain would beat Obama, the conventional wisdom seems to be that that's precisely because Obama and McCain have been campaigning against each other and ignoring her for the last month or more. If Clinton were to somehow get back into the race (let's not get into how that would happen), then McCain would campaign against her, and she'd go down in the polls.
Of course, this is all analysis of the "if the election were held today" type, which it's not. I only support this type of analysis when it predicts that the Phillies will make the playoffs, which it does as of right now.