Yglesias's metaphysical question is: "better" in what objective or measurable way?

I have to agree with @mattyglesias here. There's certainly a model, even in the "Now-cast". Silver models individual polster's bias and "deskews" the polls. The Polls-plus model incorporates those deskewed polls, and adds in a model of voter behavior based on economics, another based on the current president's approval rating, and another about the random process of voter opinions between now and the election. The question is, "are these scientific models?" Is there any reasonable experiment we can do that would invalidate the model? As with a lot of Bayesian approaches to the social sciences, the answer is "no." Silver (and Wang, and ...) are proposing a model that predicts a probability. Today (Aug 3) Nate Silver's Polls-plus model predicts that Hillary Clinton will win with 66.5% probability, and Sam Wang predicts that Hillary Clinton will win with 80% probability. If Clinton wins have we learned anything about which model was "better?" How about if Clinton loses? Then the Bayesians will point out that they didn't predict that Clinton would win, they predicted that Clinton would win only 66.5% (or 80%) of the time, and this is just the other 33.5%. If we wait a century so that we have a sample of 25 such predictions of the model, of which 10 predictions were in the 60-70% range, but 9 out of 10 (i.e., 90%) of those predictions were "right" have we learned anything? Does it matter that I will be dead by the time we can finish that experiment? If the 10 predictions in the 60-70% range turn out to be right more or less than 60-70% of the time will my descendants reject the model or say "oh, well one of those elections was Clinton-Trump, and we have to use a slightly different model when one of the candidates has a personality disorder"? This is classic searching for patterns in some pretty cool math that produces white noise. Like my horoscope, it doesn't mean anything, or have any real predictive value, but it makes me feel better (or worse, depending on the election) to look it up every day.