Tyler Cowen has an interesting interview with Nate Silver. Here's something Silver said that I rather like:
By the way, another thing about the Trump thing I’ve been thinking about is — so my early view, that Trump had a very low chance — not zero, but very low — of winning the nomination was not based on any formal model, per se. I wonder what if I had even like a fairly bad model instead?The good thing about building a statistical model is that it commits you to rules, right? Instead of just kind of saying, “Well, early polls aren’t very predictive and your prior is it currently probably won’t win, therefore, probably not.”It pins you down and says, “Well, OK, early polls aren’t predictive, but at what point do they become more predictive?” When Trump went from being at 25 percent in the polls to 35 percent after Paris and San Bernardino, how significant is that?To have an answer that is set up by an algorithm you designed ahead of time is actually maybe more helpful than people would think.The long way of saying this is that I’m not sure that I’m any better than the average pundit unless I have a model. The disciplining effect of a model, doing your thinking in advance, and setting up rules of evidence is probably quite important.