Two economists from Yahoo Labs predict the 2012 election
Steven Cherry: Hi, this is Steven Cherry for IEEE Spectrum’s “Techwise Conversations.”
Prediction is one of the great hobbies of the couch potato, along with second-guessing and Monday morning quarterbacking. In fact, if you think about it, it’s all prediction, whether it’s of the past, present, or future.
Any way you look at it, this is a shaping up to be a banner year for prediction.
The bettors beat the house for Super Bowl XLVI in February, when the New York Giants beat the three-point favorite New England Patriots. You can bet on the 2013 Super Bowl already, by the way. The Giants are 15 to 1 to repeat; Green Bay is favored at 6 to 1.
March Madness, the massive 68-team college basketball tournament, starts March 13th, another highlight of sports betting.
But the mother lode of predictions this year is the 2012 U.S. elections. Two billion dollars might easily be spent between now and November, so if you think it’s too soon to predict Super Bowl XLVI, you probably also think it’s too early to call the presidential race.
You’d be wrong.
President Obama will be reelected with 303 votes in the Electoral College, winning 26 states and the District of Columbia, including California and New York by wide margins and squeaking out a win in key battleground states Ohio, by 50.3 percent, and Pennsylvania, by just under 52 percent. So predict two economists at Yahoo Labs, Patrick Hummel and David Rothschild, who is my guest today.
He has a Ph.D. in applied economics from the Wharton School of Business at the University of Pennsylvania, where his dissertation involved creating forecasts just like this one. He’s been with Yahoo Labs, in New York City, since May of last year, and he joins us by phone from there.
David, welcome to the podcast.
David Rothschild: Thank you very much for having me.
Steven Cherry: David, maybe the most striking thing about your election model is that it doesn’t really know who the Republican candidate is—and almost doesn’t care. But before we get to that, maybe you could tell us about modeling in general and how yours was built.
David Rothschild: Sure. This is a fundamental model, and the basis of fundamental models are dropping out polls and prediction markets and thinking about fundamental data that’s available well before the election. And so this is based off such things as presidential approval ratings in mid-June, economic indicators, incumbency, ideological indicators, biographical details, and of course past election results. Those are the main categories, and what you’ve just read off was making some expectations on what those economic indicators and presidential approval will be later in the summer. But there’s really two main reasons to be making fundamental models like this. The first is that it does allow us to make fairly accurate predictions well ahead of the election, which is fun and interesting but also meaningful to those people involved in elections. And the second thing is that by making these fundamental models that take away polls and prediction markets, it lets us look and see how this fundamental data does correlate with election results. These are things that pundits knock back and forth on a regular basis, and here we can add a little bit of data and clarity to those discussions.
Steven Cherry: You mentioned “fairly accurate”—I mean, you have some measure, right? You applied your model to other elections.
David Rothschild: That’s correct. So this model in particular was calibrated based off the last 10 cycles. We’re looking at state-by-state elections, so we’re looking at 510 somewhat independent elections. Obviously there are national currents as well as idiosyncrasies between the states in any given election cycle, but especially for things like past election results you get to train it on a fairly large number of different elections, if you look at 10 cycles and the 51 different electoral college elections. And that’s what it’s based off of, but then we work to systematically drop data in order to continuously have an out-of-sample look at the data as well when we calibrate it.
Steven Cherry: We should note what the margin of error is.
David Rothschild: Sure. So what we’re looking at here—and the easiest way to think about it is, is that there’s a mean absolute error in the expected vote share—so the amount of the two-party vote share that either candidate will receive—it’s about three percentage points.
Steve Cherry: And there are a lot of state results—I guess, actually about 15 of them, including some big ones—within that, right?
David Rothschild: Sure. And in the way you read it off is fun, exciting, but if you look at the tables we provide, some of these states we’re giving a probability of victory within 40–60 percent, so there are at least three states here, Virginia, Ohio, and New Hampshire, which are within just a few percentage points of flipping over from one candidate to the other.
Steven Cherry: And if somebody runs as a third-party candidate, a serious third party candidate—I mean, for example, if Ron Paul ran and might get 10 percent of the vote, that might throw everything out, right?
David Rothschild: I wouldn’t say it would throw everything out. It is something that we’ve looked at. It’s something we forget, but there have been three serious third-party challengers in the last 11 cycles, I guess. We had Wallace in ’68 and Anderson in ’80 and Perot in kind of ’92 and ’96, so it is something that we’ve seen before, though it is hard to tell exactly at this stage which candidates it affected the most. But it is something that the data has seen before.
Steven Cherry: Now your model does assume that Mitt Romney is Obama’s candidate [but] only for the purpose of picking the Republican candidate’s home state, I guess?
David Rothschild: That’s correct. It doesn’t make very much of a difference in this model, and I think that’s one of the main interesting findings of this, is that, quite frankly, you can make a fairly accurate prediction pretty far away from the election and do it without even knowing the candidates. Now, this is not to say that the candidates and campaigns don’t make a difference; as you mentioned, probably well over a billion dollars will probably be spent by each side, so first of all we’re talking about the net effect of the campaign. So if one candidate spent over a billion dollars and the other candidate wasn’t able to equal that, you’re likely to see some major impact. So we’re talking about the net effect, and there’s also some error, and this error, a lot of it is idiosyncratic between the states in a given year. But still, these are things that are affected by the campaigns and candidates, and that’s really where you can think their impact is made.
Steven Cherry: I guess the biggest variable in the model is the economy, but even there the model looks at it differently from the way the pundits do.
David Rothschild: Right. So, I say there’s kind of two interesting things—and it kind of speaks to your last question as well—is that we look at the economy trends rather than levels. And this is something we took a very extensive look at, and we feel very confident stating it. And it makes sense; the American public is rather myopic. They’re not likely to focus on what the state of the economy was several years ago. To know what really is good or bad is something they have a grasp on, but still, it’s a very short-term memory. And so, movement within the first and second quarters are really the key things to be considering, and I emphasize first and second quarters because we actually gain no extra significance by putting in the growth rates of the economy in the third quarter, which is something that’s not released until mid-October, which seems to have no influence as correlated with predicting the elections.
Steven Cherry: So if the economy continues to improve at least through June, Obama could win by an even bigger margin than the model predicts?
David Rothschild: That’s true. So we actually—just to keep it extremely conservative—assumed what would happen if the presidential approval rate held steady and the economy grew at what would be an average election-year pace. But that’s likely to be a conservative estimate at this point because all economic indicators are moving forward in a positive manner.
Steven Cherry: So besides the economy—and we’ve mentioned the home state as well—what are some of the other big predictors?
David Rothschild: So the big predictor, of course, one is past election results, which is one of the best ways to predict a given election. But also incumbency, which is also massively important—well, actually, in all levels of our government. There is a huge incumbency advantage, and that is coming from a lot of different factors including money, name recognition, and of course it is an indication of past election victories. Presidential approval is something we put in this model that actually is something that runs fairly closely with economic indicators. So if you want to predict presidential approval based off economic indicators, about 50 percent of the variability of presidential approval is described by the economic indicators. And then finally we throw in some state-by-state ideological indicators, such as the percentage of Democrats in the lower house of the legislature, as well as using the rankings by different outside groups of senators to indicate how liberal or conservative a given state is.
Steven Cherry: You know, David, which party controls Congress in 2012 might be at least as important as what party has the presidency. Did you look at the congressional and Senate races?
David Rothschild: So—well, I can come at that with a few things. So we haven’t released yet, but we do have, a fundamental model which is running parallel to this which will come out shortly for the senatorial races, but I can say that the Senate is heavily likely to go to the Republicans at this point. And this is something that we can look at through a bunch of different lenses, including prediction market data—which on these aggregate questions is fairly accurate this far out even—in which I can say the prediction markets are giving the Democrats about a 20 percent chance and the Republicans about an 80 percent chance of controlling the Senate. This is despite the fact that Democrats have more senators going into this election. But due to the fact that they have a lot more senators up for reelection, and a lot of them who won in the 2006 landslide, they won some fairly Republican districts, really putting the Democratic majority at risk this year.
Steven Cherry: Very interesting. So twice we’ve mentioned prediction markets, and they were sort of the big thing in handicapping the presidential elections of a few years ago. And in fact, prediction markets are a big area of research for Yahoo Labs, so it’s pretty interesting that you didn’t use them, and maybe you can just remind our listeners of what prediction markets are first.
David Rothschild: Sure, definitely. So prediction markets are markets in which you can buy or sell contracts that expire at either zero or one, depending on the outcome of the event. And so if you imagine a prediction market for the 2012 election, you could buy a contract for, say, Mitt Romney to win the election which you can purchase at any price but will be worth one dollar if he wins, will be worth zero dollars if he loses. And the price that that is selling at any given time is a very strong indication of the probability of victory of the candidates. And so this is something we’ve also been using extensively. And what I can say is that the fundamental model provides a very strong state-by-state elections predictions this far out; there’s not much liquidity yet in the predictions market when it comes to looking at state-by-state outcomes. There is a mass amount of liquidity in the national election and in the GOP primary—we have been talking about that a lot on our blogs—so that’s something we do follow closely as well, and it’s something that we put a lot of weight on as we get closer to Election Day. So as we look at our state-by-state predictions, we’re going to start with fundamental models this far—nine months—out from Election Day, but as we approach 75 to 90 days away from the election, we’ll start weighing prediction markets more heavily and then prediction markets and polls as we get closer to the convention, and then finally at some point, probably about 45 days out, we don’t use the fundamental models any more in our predictions, and they’re based mainly on prediction markets and a little bit on polls.
Steven Cherry: And so, just to be clear, “liquidity” really has to do with how much betting is being done, and without it a single bet can really shift the market.
David Rothschild: That’s right.
Steven Cherry: David, I think election watchers would note that this is the first presidential election since the Supreme Court decided the Citizens United case, and that change in politics has the potential to make this election different from every one that preceded it.
David Rothschild: We’re not too worried about it. I mean, there’s been a lot of dramatic changes in elections and electioneering in the last couple of cycles, and these things come and they go, and one of the things that we value heavily with prediction markets per se is that they are able to incorporate those changes as well, so that’s why when we do get closer to the election we will be resting more heavily on them. If there is some sort of seismic shift this election with the models, then the fundamental models get changed in the future years. That’s something, that they have to be based on historical data, so we understand that as a limitation. That’s one of the reasons why we really do focus more on prediction markets as we get closer and there’s enough liquidity.
Steven Cherry: Well, David, I think one of the most fun things about predictions is that they give us a framework for looking at the future as it unfolds, and I know you said you’ll keep updating your predictions throughout the election season on your blog, which we’ll link to from the podcast page so our listeners can follow along too. Thanks a lot for doing this work and for joining us today.
David Rothschild: Thank you. It was really my pleasure being here.
Steven Cherry: We’ve been speaking with the economist David Rothschild, of Yahoo Labs, about how Barack Obama won reelection in 2012, at least, according to his detailed prediction model. For IEEE Spectrum’s “Techwise Conversations,” I’m Steven Cherry.
Announcer: “Techwise Conversations” is sponsored by National Instruments.
This interview was recorded 27 February 2012.
Audio engineer: Francesco Ferorelli
Follow us on Twitter @techwisepodcast
NOTE: Transcripts are created for the convenience of our readers and listeners and may not perfectly match their associated interviews and narratives. The authoritative record of IEEE Spectrum’s audio programming is the audio version.