The Union reaches 30 percent on Tuesday and Wednesday, 28 percent on Friday, and in the following week it lands between 24 and 26 percent. That is the current publication rhythm of the survey institutes. In view of this flood of polls on the federal election, even voters interested in politics quickly lose track of things. And some may have their doubts as to the extent to which such surveys can even provide a reliable look into the future.
The question “Which party would you vote if there were parliamentary elections next Sunday” does not even refer to the election date on September 26th. Not even in the future, actually. The question really only answers which party the survey participants prefer at the time of the survey – a snapshot. Nobody knows whether they will change their minds before the election date or whether the answer is right at all.
Even if the institutes reject this interpretation, many people still interpret the polls as a prediction of which party will do in the federal elections and how. “Today’s polls have only a limited connection with the election results in September,” says Thomas Gschwend, professor of quantitative methods in the social sciences at the University of Mannheim. The surveys are of good quality, but no forecast, just an interim result.
The SZ tries to take this problem into account in its dashboard and its graphics to choose from and to show the uncertainty of these surveys more strongly. The individual survey results take a back seat, instead we create an average that compensates for individual outliers but makes trends clearly visible over time.
The problem that the polls are not predictions remains, of course. Therefore, the SZ is creating additional offers in this election year to better classify forecasts and provide a different perspective on the possible outcome of the election, namely by simulating the Bundestag election thousands of times beforehand. The SZ is cooperating with this Second vote.org – a project that Gschwend started together with other researchers from the University of Mannheim in 2017. A website that wants to predict the election result six months before the election, with the help of political science explanations for voting behavior. For the federal election in 2021, Zweistimme.org is making its model of the SZ available for the election forecast.
How the prediction is made
The forecast looks at two levels: all of Germany and every single constituency.
For the prediction of the overall result, Zweistimme.org also works with the current polls. However, the researchers enrich them with historical results, as well as with an indication of which party is currently providing the chancellor. Because this data is available long before the election, a statement can be made at a very early stage as to where the parties are likely to end up, as this enables projections based on longer-term trends. The ups and downs in the polls during the election campaign are deliberately dampened.
But the composition of the Bundestag cannot be fully predicted with the second votes alone: There are no direct candidates who are elected in the individual constituencies. The survey institutes generally do not conduct any surveys locally, as that would be too time-consuming.
The first votes also play a role
Zweistimme.org tries to make predictions at this level as well. The basic assumption for this is that the parties lose or gain shares in their first votes in the constituencies in proportion to the national trend.
Because this forecast of the constituencies also requires data about the candidates, which has not yet been provided by the Federal Returning Officer, it will probably only be published from the beginning of September.
An example: In the 2017 election, the SPD got 20.5 percent of the second vote – let’s assume that the polls currently show only 16 percent, i.e. minus 4.5 percentage points. In other words: a loss of 22 percent compared to the election result four years ago. For each constituency, the model subtracts this value from the constituency result of the previous election.
The researchers combine these values with previously explored explanatory approaches for voting behavior from political science, taking factors such as the number of candidates in the constituency, the incumbent bonus, the location of the constituency in West or East Germany, the list position of the candidates, whether they have run before , your gender and any academic degree you may have.
A complex mathematical process from the field of machine learning uses these variables from the elections from 2009 to 2017 to calculate a comprehensive model for the election. It describes what the relationship between the variables and a possible election result could look like: The survey results are thus combined with other explanatory variables, the computer model tries to find an optimal mathematical way of how the variables relate to an election result.
From model to probability
Because models simplify reality and reduce it to a few variables, the researchers also use simulations to generate variance in the forecasts. The model is used to repeat the Bundestag election 9,000 times, and the result of each round is noted. The more often a party has achieved a majority in these thousands of repetitions, the more certain the result.
The following graphic shows the distribution of the simulation results for the most likely coalitions. For each coalition, the curve shows the proportion of seats where most of the simulation results land. The more of these results are above the 50 percent mark, the more likely this coalition will have a majority in the model.
From these simulation results, in turn, probabilities can be calculated – for very different questions in this Bundestag election: How often does a candidate get the most votes in a constituency? How often does a coalition get a majority?
And this data gives us even more information: Because the first votes also result in possible overhang mandates, the model can give an indication of how big the next Bundestag could be. In view of the debates about the new electoral law, this is essential information.
The combination of the first and second vote model ultimately results in a prediction of which party will receive how many seats in the Bundestag. This in turn can be converted into shares, and is therefore – unlike the Sunday questions – a prognosis of how the general election in September could turn out: