German election forecasts
Getting the result right
Is there a simple way to correctly predict the result of the upcoming general election in Germany? LMU’s Andreas Graefe has helped to develop a method that has performed impressively in US elections. But will it work in a German setting?
Photo: Deutscher Bundestag / Marc-Steffen Unger
The German election campaign has taken to the highways and the debate surrounding the current election is becoming tenser. With a little over a month until Election Day, attention focuses on two related questions: Will the FDP, junior partner in the present coalition, be represented in the next Bundestag? And, if so, will the outgoing coalition government reclaim its power?
According to Andreas Graefe, the current answer to both questions would be yes. The young LMU researcher recently set up PollyVote.de, a forecasting portal. According to the website, support for the major parties is currently distributed as follows: CDU/CSU, 39.7%; SPD, 24.9. The Greens have 12.6%, and the Left Party is on 7.3%. This voting pattern would give the CDU/CSU-FDP coalition a small overall majority in the federal parliament – because, in contrast to many other polling agencies such as Emnid, Forsa or Infratest, PollyVote puts the FDP’s share of the vote at closer to 6% than 5%, the threshold required to win seats in parliament. (The forecast cited above dates from 21. August (14:00). The latest PollyVote prediction is available at http://pollyvote.ifkw.uni-muenchen.de/).
“That is a realistic estimate,” says Graefe, who is at LMU’s Department of Communication Science and Media Research. “The FDP has almost always done better on Election Day than in the opinion polls.” The Free Democrats tend to profit from what are called Leihstimmen, literally “borrowed votes”. Many CDU supporters vote for the FDP, their favored coalition partner, simply to ensure that the party gets over the 5% threshold. For example, in the state election in Lower Saxony earlier this year, the polls predicted around 5% for the FDP. In the end, they got 9.9% of the vote.
Pirates and AfD not onboard
PollyVote has just incorporated the results of its latest poll of experts into the forecasting model. This component combines the estimates of more than 100 professional observers of politics – political scientists and journalists. The most recent forecast once again sees the FDP safely in the next Bundestag. Moreover, it sees the same set of parties in the newly elected legislature as in the outgoing parliament. The Pirate Party and the recently constituted Alternative für Deutschland (AfD) are predicted to obtain less than 5% of the vote, leaving them on the outside looking in. In any case, the vast majority of experts remain convinced that Angela Merkel will win another term as Federal Chancellor.
The PollyVote portal is a newcomer in the German election forecasting scene. So why should anyone have more confidence in its methodology than in the techniques used by its well established rivals, the pollsters? “All forecasting methods have their strengths and weaknesses,” says forecaster Graefe. For this reason, PollyVote, which was developed in the USA, does not attempt to replace the established ones by taking an entirely novel or particularly sophisticated approach. Instead, PollyVote simply combines forecasts from existing forecasting methods. The more diverse the methods and the sources of information their forecasts draw upon, the better, as the systematic errors associated with each are more likely to cancel each other out, Graefe says.
Obama was always in front
This is borne out by PollyVote’s accurate prediction of the result of the Presidential Election in the US last November. The model not only consistently forecast that Barack Obama would be re-elected, but its estimate of the margin of victory diverged from the final value by just 0.9%. In the election of 2008 the difference was only 0.7% and for the 2004 election the error was even smaller, at 0.3%. Andreas Graefe has been involved in the development of PollyVote for several years now, and is currently one of the project leaders. In this respect, he can also be numbered among the winners of the 2012 election.
PollyVote’s predictions have not only been remarkably precise, they have also proven to be comparatively stable over the course of the election campaign. One of the features of the 2012 campaign in the US was that, at various points – after the first TV debate, for instance – many pollsters saw the challenger Mitt Romney out in front. PollyVote, however, was not among them. On average, over the final 150 days of the campaign, PollyVote’s forecast varied from the final result by no more than 0.3 percentage points. And for months before Election Day, the portal saw the incumbent ahead, while the media kept overstating the importance of campaign developments. Individual, apparently dramatic events “usually have little effect on the final results,” says Graefe. Even the Lehman crash, which precipitated the wider financial crisis, was not a decisive factor in the 2008 election.
Betting on the ballot
Notwithstanding these successes, given the obvious differences between the electoral systems in the US and Germany, is PollyVote’s forecasting methodology actually applicable to the election here in September? That is precisely the question that Graefe has been studying in the context of a Research Focus at LMU’s Center for Advanced Studies (CAS). It was not necessary to modify the general principle of the method, he says. The biggest change in the German variant has to do with one of the components that contribute to the final forecast. In the American case, PollyVote can draw on more than 20 statistical models that rely on different variables to predict the election outcome, such as the state of the economy, the popularity of the incumbent, or the length of time the incumbent party has been in power. For the German variant, only four such models are available.
Another component that goes into PollyVote.de are forecasts derived from so-called prediction markets, in which the participants place bets on the outcome of the election. In order to incorporate the results of traditional opinion polling (in which individual voters are asked to state their voting intentions), Graefe makes use of two services that aggregate the data from a whole series of such polls, thus increasing the reliability of the overall result. And finally, as mentioned above, PollyVote incorporates the results of regular consultations with expert observers.
The beauty of simplicity
The real trick, however, lies in how the forecasts from the various methods are combined: PollyVote calculates simple averages, dispensing with the need to use complicated weighting procedures. First, the results of methodologically related forecasts are pooled. Then, the resulting mean values are added together and averaged. That may sound too simple to be true (in every sense) but, for Graefe, this “mechanical” combination of predictions, with its avoidance of weightings, represents “the secret of the method’s success” and is “one of the most effective ways to obtain accurate forecasts”.
According to Graefe, research on forecasting has clearly demonstrated that the notion that complex problems can only be solved by complicated procedures is a misconception. Indeed, his real mission is to show that the simpler techniques can also be the more effective ones. The election in Germany gives him a new opportunity to test the thesis. And the result of this test, not the question of who wins and who forms the next government, is clearly the central issue for him.