Saturday, November 3, 2012

Opportunity for Election Forecasting

I have absolutely no expertise in election forecasting. But I have enough faith in the stability of human behavior, properly understood, to believe that election forecasting experts (not me!) should be able to accurately and confidently predict who will be president next year.

This not to say that such a forecast would merely consist of taking the leader in a large national poll, or the leader of a large poll conducted in Ohio. You have economic data, rallies, campaign contributions, early voting patterns, billions of tweets and facebook chatter, and many other glimpses at what people might do on Tuesday. I would think an expert could process all this and tell us who will win.

Theory 1: The Experts do Know
One view of the state of forecasting is that the experts have done exactly that, and President Obama will win. One reason I hesitate to embrace this view is that the experts I have seen (not a good sample, because I am not an expert on experts, either!) have not considered all that much data (I don't consider repeats of flawed data to be much data, regardless of how many times it is repeated).

Nate Silver, for instance, gives the President an 84% change primarily on the basis of state-level polling data, and gives essentially zero weight to the national polling data (see the middle of this long post), let alone any of the other sorts of data I mentioned. He may have the right weights -- he's the expert -- but an explanation of his dramatic departure from Bayesian good practice (ie, throwing out important data sources rather than down-weighting them on the basis of assessed flaws) is conspicuously absent.

Justin Wolfers looks at who people expect to win (which also points to victory for the President). That's an interesting data source, and he explains why he puts less weight on the usual polls. But what about all of that other data?

Intrade, meanwhile, puts the President's chances at 2/3 (ie, 2-to-1 odds of losing). That's pretty close, and not the confidence I am expecting from a good forecaster. (for confidence, look at silver's odds, which exceed 5-to-1). Either intrade refutes the theory that the experts know who will win, or intrade has done a poor job of aggregating information.

If the President wins by a non-trivial margin (ie, by more than Ohio, or by a 5+ percentage point margin in Ohio), it will tough to reject Theory 1. But that outcome would also support the theory that the experts got lucky.

Theory 2: It Really is Too Close to Call

Some elections, such as Bush v Gore, are genuinely close and I don't expect the experts to know those outcomes ahead of time (or even the day after the election). But they should know that they don't know. Silver, and maybe not even Wolfers, are saying that they don't know. Other experts are calling it a tossup, but I am worried that they work for TV networks who want viewers.

If either candidate wins by a non-trivial margin, that casts a lot of doubt on Theory 2 unless we have good reason to believe that the world changed significantly in the last hours before the election.

Theory 3: The Experts aren't Doing a Very Good Job

If the President loses by a significant margin, Theory 3 has to be the favored theory, which is an opportunity for an ambitious election forecaster to get it right next time.

1 comment:

TGGP said...

I thought Chicago folks believe in EMH. InTrade isn't as liquid as some might desire, I've heard Europeans recommend BetFair.