Obama still in front
How can one reliably predict the outcome of an election? A new research focus at LMU’s Center for Advanced Studies reviews forecasting methods – and their predictions for the 2012 presidential election in the US.
Photo: ddp images
This round clearly went to Barack Obama, on points. Pursuing the boxing metaphor, one could say that in the third televised debate in the US presidential campaign, the incumbent pinned challenger Mitt Romney to the ropes, giving him little chance to land a clean punch. The first encounter in the series was a very different affair. Afterwards, commentators acidly observed that the President had spent more time inspecting his shoes than projecting a sense of self-confidence and optimism over the airwaves. And perhaps predictably, a slew of opinion polls suddenly saw Romney in the lead less than a month before Election Day. In any case, it now appears that Obama has a real fight on his hands. And with hurricane Sandy putting in an unscheduled and dramatic appearance in the final stages of the campaign, the President is now called upon to exercise leadership by taking decisive action in response to the crisis.
In comparison with this challenge, “winning” a TV debate is the kind of ephemeral campaign success that “usually has very little effect on the final result,” says Andreas Graefe of the Institute of Communication Science and Media Research at LMU. As Graefe, who studies forecasting methods, points out: “The media usually overreact.”
Moreover, as Election Day nears, the candidates’ ability to turn an unforeseen event into a quantifiable advantage decreases dramatically. Even the Lehman crash in 2008, which sparked the current global financial crisis and undoubtedly favored the challenger that year, was not the decisive factor in catapulting Obama into the White House.
Only 0.7% off
But what is the crucial element that determines the outcome of an election? And more importantly, how can one identify it in advance? These two questions form the core of a new research focus at LMU’s Center for Advanced Studies (CAS). Graefe, who is leading the project, has recently done a great deal of work on data derived from American presidential elections. For the past four years, he has been directly involved in the forecasting portal PollyVote.com. In fact, he is now one of four leaders of the project. In the last two elections PollyVote managed to predict the final results quite accurately. Its forecast was off by only 0.7% in 2008 and by just 0.3% in 2004. At the time of writing, PollyVote sees Obama with 50.9% and his Republican challenger Romney with 49.1% of the vote. But … these figures are based on the assumption that the two leading candidates divide the popular vote between them; it takes no account of the many third-party candidates. Interestingly, the gap between the main contenders has narrowed by only a tenth of a percentage point in recent weeks.
For the research focus, the US election is simply one, albeit highly topical, test case for how to assess the present state of the science of prediction. The project’s eventual aim is to extend the methodological basis of PollyVote to make it applicable to the federal election in Germany next year, which involves a much more complex roster of competing parties.
Beards and ballots
So what exactly is PollyVote? Contrary to what one might think, it is not a single, highly complex forecasting method at all. Instead, it functions by combining the results of some two dozen different assessment models, based on the use of diverse parameters. The data sources include five online services that analyze and aggregate the results of opinion polls as well as a forecasting market, where participants place bets on the outcome of the election. In addition, PollyVote integrates more than a dozen well-known econometric models, which estimate the impact of factors that seem likely to influence the election result, such as the state of the economy or the popularity of the incumbent. Finally, PollyVote also incorporates the results of regular surveys of academic political experts.
The architects of PollyVote have also developed their own models. One of these gauges the impact of selected aspects of candidates’ biographies - traumatic experiences during childhood or level of education, for instance. Even apparently trivial things, like the tone of a candidate’s voice, whether he has a beard or she wears glasses, appear to have an impact on how voters decide. When Graefe applied this tool to the biographies of the candidates in modern American presidential elections, the model correctly predicted the results in 27 of the 29 contests. Another polling model asks voters to name the candidate whom they believe is best equipped to handle the most pressing political and economic problems. And finally, PollyVote is also experimenting with an approach to polling in which the respondent is asked how he expects to the election to turn out, rather than how he intends to vote.
However, the real strength of PollyVote lies in the combination of the single forecasts, which is done in the simplest possible way. Rather than applying complicated weighting procedure to the individual results, PollyVote simply calculates the plain old average for the whole sample. The analysts first pool forecasts obtained by analogous methods and then averages the means obtained for each of these samples. As Graefe admits, “That not only sounds very naive, it is incredibly naïve.” While the simplicity of PollyVote’s approach may seem counterintuitive, “combining forecasts is one of the most effective ways to generate accurate forecasts.” Combining allows one to include diverse information incorporated in individual forecasts and tends to cancel out their systematic errors. In his contribution to the research focus at the CAS, he plans to demonstrate that simple forecasting methods can be just as effective as, and in many cases better than, complex ones. Graefe himself has a clear preference for simple procedures. “Research on forecasting shows that simple methods often lead to more accurate predictions. After all, voters themselves use pretty straightforward calculations in reaching their decisions.” math / PH