I admit defeat: Wonkblog’s June election model beat me

November 6, 2012

In April, I worked with three political scientists -- GW's John Sides, Yale's Seth Hill, and UCLA's Lynn Vavreck -- to build an election forecasting tool that didn't require anything more than economic growth through the first three quarters, the president's approval rating in June, and whether the incumbent president was on the ballot. 

The point of the model wasn't to maximize accuracy on Nov. 6. If we'd wanted to do that, we would've built a model that converted into a polling aggregator as we got closer to the election. Rather, it was to show that, in most cases, elections are broadly predictable months before they happen. 

In June, President Obama's approval rating averaged 46.3 percent. Economic growth from January 2012 to June 2012 was 0.81 percent (this is a non-annualized number, so it's different than what you typically see reported for economic growth; the number you're more used to is about 1.8 percent). And yes, Obama is the incumbent, and he's running for reelection. Those three pieces of information convinced our model that Obama had about a 67 percent chance of being reelected. 

I didn't buy it. If you go back and look at the column accompanying the model, I'm searching for reasons the model is wrong. Obama's odds of winning were, in the model's telling, simply too good in such a weak economy.

But UCLA's Lynn Vavreck told me to chill. “There’s this moment where you go, ‘Whoa, that’s a high number’,” she said. “ ‘Something must be wrong.’ But what it forces you to do is . . . to divorce yourself from contemporary context. You can’t go in thinking, ‘But, oh, this president is black,’ or ‘Gas prices are high,’ or ‘We just had the tea party.’ You have to strip all that away and say: ‘Incumbent parties, in growing economies, almost always win in contemporary American history.’ ”

Vavreck was right, and I was wrong. Using data available over the summer, Wonkblog's model accurately predicted where the fundamentals of the race would be today. Here's our prediction against the most recent predictions from Nate Silver, Sam Wang, and the major betting markets:

As you can see, it's less confident of an Obama victory than the models, but about as confident of an Obama victory as the betting markets. And, to repeat myself, the Wonkblog model is using data available months ago -- long before the debates and the final rounds of ads and all the rest of it. 

Which is all to say that the model did a far better job than I did predicting where we'd be today. It also did a better job than much of the media, which spent the election anointing a new "gamechanger" almost every week, despite the fact that the game never really seemed to change. In presidential elections, the fundamentals are strong.

The model's final test will come Wednesday (or as soon as we know the results of the election). The model predicted Obama would win 51.3 percent of the two-party vote, though it has a margin of error of about three percentage points. That's a bit higher than what we're seeing in the polls, but who knows? It's been right before. And if it wins, then I, for one, will welcome our new algorithm-based overlords.

Comments
Show Comments
Most Read Business
Next Story
Sarah Kliff · November 6, 2012