How to Improve Political Forecasts

Aubrey Clayton in Nautilus:

Every election we complain about horse-race coverage and every election we stay glued to it all the same. The problem with this kind of coverage is not that it’s unimportant. Who’s leading a particular race and by how much can matter a great deal to someone trying to decide which candidate to volunteer time or money to, or even whether it’s worth learning the policy positions of a fringe candidate with no chance of winning. The problem is the well-documented incentive for the media to make every race seem closer than it is, what Erika Fry in the Columbia Journalism Review called “rooting for the race.”

Those who simply want to know how the candidates stack up and who seek a more objective, quantitatively rigorous alternative to punditry have in recent years turned to such data journalism sources as Nate Silver’s FiveThirtyEight. But even they are feeling burned by 2016. Silver, who scored on both Obama victories, had Trump as a distant longshot in the primary and Clinton as a 71 percent favorite to make the winner’s circle in the general. Data-centered political coverage, as a whole, took a serious blow. Silver issued a partial mea culpa for the primary forecasts. But he has steadfastly defended the general election forecast with the arguments that (1) his model had given Trump better chances than nearly anyone else had, and (2) something having a 29 percent chance of happening doesn’t mean it’s impossible. Maybe the underdog got lucky. The latter defense, though, revealed a destabilizing truth about the nature of such forecasts that took many readers by surprise: Since all the forecasts are probabilistic, meaning they’re not strictly wrong in the event of an upset victory, then in what sense had they ever been right?

More here.