I agreed to give a talk in December for Jared, and this is what I came up with:

Election Forecasting: How We Succeeded Brilliantly, Failed Miserably, or Landed Somewhere in Between

Andrew Gelman, Department of Statistics and Department of Political Science, Columbia University

Several months before the election we worked with The Economist magazine to build a presidential election forecasting model combining national polls, state polls, and political and economic fundamentals. This talk will go over how the forecast worked, our struggles in evaluating and improving it, and more general challenges of communicating data-based forecasts. For some background, see this article.

Actually, the title is misleading. Our model could fail miserably (for example, if Joe Biden wins Alabama, which we say has less than a 1% chance of happening) or it could land somewhere in between (for example, if Biden wins the electoral college but with just 51% of the popular vote, which is at the edge of our forecast interval) but it can’t really succeed brilliantly. Even if our model “correctly predicts 49 states” or whatever, that’s as much luck as anything else, as our estimates have margins of error. That’s one reason why, many years ago, my colleague and I decided not to put more effort into election forecasting: it’s a game where you can’t win big but you can lose big.

Anyway, I’ll be able to say more about all this in a couple weeks.