In honor of Election Day, we bring you a slight change from our usual programming.
There has been a lot of talk about the use of futures markets to predict elections. The granddaddy of election markets (in the US) is the Iowa Electronic Market. The one that gets the most attention these days is Intrade. I used to trade on the IEM during the primaries and made a decent return in just a few weeks, mainly by betting that people would overreact to news. (For example, when Huckabee (remember him?) spiked – I believe he was in the lead on IEM at one point – it was a pretty easy bet that at some point before the convention he would come down to zero.) But then I did a bet on the general election and forgot to close it out at the right time, so on balance I lost a few bucks. (The maximum you can put into IEM is $500, so we’re not talking big sums here.)
FiveThirtyEight.com takes a different approach: they take polling data as inputs, and then run multiple simulations of who will win each state election. A given survey has a midpoint (say, Obama 47 – McCain 45) and an error distribution around that midpoint. By doing repeated simulations, you estimate how often each person would win that election, based on expected variance around the midpoint. If you do this for all states at once, you get an estimate of what the electoral vote tally will be. I don’t know if they account for correlations in the error across different states – the fact that if McCain does 2 points better than expected in Pennsylvania, he is likely to do better than expected in Ohio, too (the two are not independent outcomes). They should take this into account. (I don’t know because I haven’t read the website other than the predictions.)
The problem with both of these approaches is that they take polling data as their inputs – so if there is a problem with the polls (the Bradley effect, for example), they will produce inaccurate results. Polling markets partially compensate for this, because they incorporate people’s expectations of how accurate the polls are. But given the prominence of the polls, I doubt they can correct for polling inaccuracy.
Not surprisingly, economists have developed predictive models for presidential elections based on economic conditions. Mark Thoma provides an excerpt of one (and a link to the whole paper) here. These are statistical models that compare election outcomes to various economic variables at the time of the election. The problem with these models is that presidential elections are overdetermined: the sample size is small enough that you can find many different series of data that seem to predict outcomes accurately, like the Washington Redskins predictor. All of these cute predictive models are based on the same fallacy: with hundreds of sports teams to choose from, and the thousands of ways you can slice the data, it would be remarkable if you didn’t find one that seemed to be a perfect predictor of presidential elections. Economic models are better (though not perfect), because they are based on variables that you would expect should have an impact on election results.
Happy Election Day.