In election forecasting, most structural models embrace political-economy fundamentals. In Germany, structural models, although not plentiful, have been pursued for some time. These predictions differ from polling-inspired forecasts that, along with markets, comprise the trifecta of election-forecasting methods, in their commitment to causal explanation. The first German foray was the political-economy model of Jérôme, Jérôme-Speziari, and Lewis-Beck (Reference Jérôme, Jérôme-Speziari and Lewis-Beck1998), which emphasized the importance of economic conditions (see also Jérôme, Jérôme-Speziari, and Lewis-Beck Reference Jérôme, Jérôme-Speziari and Lewis-Beck2013, Reference Jérôme, Jérôme-Speziari and Lewis-Beck2017). Another prominent model is Norpoth and Gschwend’s (Reference Norpoth and Gschwend2003, Reference Norpoth and Gschwend2010, Reference Norpoth and Gschwend2017) Chancellor Model, which focuses on Chancellor-candidate popularity. Together, these pioneering studies address the role of political and economic dynamics in a campaign.
In this contribution, we break new ground in structural election forecasting. First, we spurn the dominant-government-versus-opposition paradigm and predict the performance of individual blocs. Although this approach is not new to the structural approach per se (Jérôme, Jérôme-Speziari, and Lewis-Beck Reference Jérôme, Jérôme-Speziari and Lewis-Beck2013; Mongrain Reference Mongrain2021; Stoetzer et al. Reference Stoetzer, Neunhoeffer, Gschwend and Sternberg2019), our effort eschews any public-opinion measures. Second, and more innovatively, we focus exclusively on the undermined dimension of a country’s political history. Our model posits that vote share as a function of longer term—if not enduring—political structure variables have potency in predicting. These variables tap profound political events, such as German reunification, “grand-coalition” governance, party dominance in the Länder, and past party strength. This constellation of factors constitutes what we call a political-history model. Such a stance may seem bold until we recall that Germany represents a contemporary beacon of political stability. We demonstrate that this approach yields relatively sound estimates for the Christian Democratic Union/Christian Social Union (CDU/CSU), Social Democratic Party (SPD), and All Others across 18 German Federal Elections over six decades.
Using the German political-history model for its first ex-ante prediction—the September 2021 Federal Election—this approach forecasts that the CDU/CSU will win 33.8% of the list vote (Zweitstimme) and the SPD will win 16.8%. For All Others, our model estimates the vote share at 49.4%. Thus, our political-history model forecasts that the incumbent grand-coalition government has a reasonable chance of remaining a viable governing alternative.
THE MODEL
This section discusses the theoretical foundations of our political-history model and describes the research strategy underpinning our empirical analysis.
Theory
Our forecasting model adopts the de Tocqueville (1856, cited by Ratcliffe Reference Ratcliffe2016) perspective: “History is a gallery of pictures in which there are few originals and many copies.” There is a long pedigree in political science that highlights the role of history and path dependence (Pierson Reference Pierson2000). Indeed, some traditional election-forecasting models incorporated historical patterns—such as midterm gains (Keilis-Borok and Lichtman Reference Keilis-Borok and Lichtman1981) and longevity in power (Abramowitz Reference Abramowitz2008)—to account for incumbent performance. We depart from existing scholarship in election forecasting because of our sole reliance on historical–political indicators. Critics may charge that this is a daring step in a world in which electoral volatility is prevalent (Gallagher, Laver, and Mair Reference Gallagher, Laver and Mair2011). Thus, there is legitimate concern about the historical model favoring stability over change. Nonetheless, short-term forces (e.g., candidate popularity) also may be unstable, resulting in surprise forecasting errors. The 2017 German Federal Election represents a case in point when the Chancellor Model (Norpoth and Gschwend Reference Norpoth and Gschwend2017) failed to forecast that Angela Merkel would remain in office due to the “Schulz-Effekt,” in which the opposition leader Martin Schulz’s favorability, whose popularity was high in the polls at the time of that forecast, had substantially waned by polling day. The key point is that this well-traveled model—with public opinion at its core, yielding a high R² and a low root mean square error (RMSE)—fell short due to short-term fluctuations.
Our forecasting model adopts the de Tocqueville perspective: “History is a gallery of pictures in which there are few originals and many copies.”
The German political history model rests on the following four variables:
$$ {\displaystyle \begin{array}{l}\hskip-0.5pc \mathrm{Bloc}\ \mathrm{support}=\mathrm{f}\;(\mathrm{Short}\hbox{-} \mathrm{Term}\ \mathrm{Partisanship}\\ {}\hskip4.9pc + \mathrm{Grand}\ \mathrm{Coalition}\\ {}\hskip4.9pc + \mathrm{Bloc}\ \mathrm{Number}\ \mathrm{of}\ \mathrm{Land}\ \mathrm{Minister}\ \mathrm{Presidents}\\ {}\hskip4.9pc + \mathrm{German}\ \mathrm{Reunification}+\mathrm{e})\end{array}} $$
Our initial presumption holds that support is shaped by past manifestations of support. The theoretical underpinning is partisanship—that is, citizens form enduring loyalties to blocs that influence their voting behavior (Campbell et al. Reference Campbell, Converse, Miller and Stokes1960). We adopt the notion of short-term partisanship to account for the competing force of partisan dealignment (Dalton and Wattenberg Reference Dalton and Wattenberg2000), which we measure by taking the aggregate vote share of the bloc in the previous general election.
Our second measure captures the type of governing coalition. Countless studies have identified an electoral cost to governing (Cuzan Reference Cuzan2015; Norpoth Reference Norpoth, Norpoth, Lafay and Lewis-Beck1991). Germany today, however, shows relative stability in terms of governance—only a few parties have served in government. Unsurprisingly, the correlations between vote share and governing status for the CDU/CSU (r = −0.101) and the SPD (r = −0.224) are relatively modest. We argue that the constellation of governments has more significance. Specifically, we assume that a grand coalition, with the CDU/CSU and the SPD governing together, is a critical factor. This alliance means that both blocs must compromise significantly. Some voters, therefore, will abandon the grand coalition, providing an opportunity for smaller blocs. Empirically, we capture this effect with a simple dummy variable measuring whether a grand coalition is in place before the election.
Third, given the German federal system, there is potential for governing status at the local (i.e., land) level to be significant (see Jérôme et al. Reference Jérôme, Jérôme-Speziari, Mongrain and Nadeau2021 for a similar intuition in the United States). The traditional notion of governing argues that blocs in power at multiple levels of governance tend to suffer a potential double whammy (or blessing) at the ballot box. At the land level, the land minister president’s bloc is the most prominent actor in the local government; therefore, the costs (and benefits) of office primarily flow to it. Thus, we measure the number of land minister presidents that each bloc controls six months before the election.
Fourth, we model German reunification—a phenomenon constituting a historical “super shock”—encompassing four dimensions. The first dimension is a supply-side shock, when the Party of Democratic Socialism (PDS)—predecessor of today’s Left Party—entered the German party system and added a fifth, socialist party to the political spectrum. The second dimension is a demand-side shock, with a new segment of citizens unaccustomed to democratic elections and party competition joining the electorate. This demand-side shock coincided with the partisan dealignment beginning among voters in West Germany in the 1980s and became even more prevalent in East Germany (Dalton Reference Dalton2014). Consequently, this demand-side shock built on the mechanism of declining voters in the West losing ties with established political actors and fewer voters in the East establishing these ties. The third dimension is a sociocultural shock: reunification implied striving for “inner unity” among citizens from two fundamentally different political cultures (Dalton and Weldon Reference Dalton and Weldon2010). The fourth dimension is an economic shock, including reconstruction and privatization, labor-market integration, and currency conversion. These shocks have brought about long-lasting change, still noticeable today in German voting behavior (Arzheimer Reference Arzheimer, Schoen and Wessels2016). We measure this with a dichotomous variable capturing elections held from 1990 onwards.
Dependent Variable and Data
Using historical data (Quinlan, Schnaudt, and Lewis-Beck Reference Quinlan, Schnaudt and Lewis-Beck2021), we estimate the list vote share (Zweitstimme) of three political blocs: the CDU/CSU, the SPD, and All Others.Footnote 1 The CDU/CSU and the SPD have been dominant players, and all German Chancellors have come from these two blocs. However, the “elephant in the room” is the Free Democratic Party (FDP), which also has been a long-term player. The “poster child” of the “half” in the so-called two-and-a-half-party-system definition, the FDP is a challenging party to understand. Figure A4 in online appendix A demonstrates why: FDP support follows no discernable pattern. Consequently, we include it in our estimate of All Others.
We compiled election data from 1953 to 2017 from the German Federal Elections Commission (Bundeswahlleiter) and the regional election authorities of the German Länder. We used Seemingly Unrelated Regression (SUR) (Zellner Reference Zellner1962), which allows concurrent calculation of separate models for the CDU/CSU, the SPD, and All Others combined; this results in more efficient parameter estimates and within-sample estimates of vote share that total 100. We include summary statistics, variable operationalizations, and parameters used for the 2021 forecast in appendices A and B. Supplementary analyses are in online appendix C, including within-sample forecasts for each bloc by election (see Tables C1–C3).
MODEL PERFORMANCE 1953–2017
In table 1, the slope estimates tend to align with our theoretical suppositions. The coefficients for vote share in the previous election are in the expected positive direction for all blocs. However, only the coefficients for the SPD and All Others reach conventional levels of statistical significance (p<0.05). A grand coalition for the incumbent government also results in the expected patterns—the CDU/CSU and the SPD on average losing support, as evidenced by the negative coefficients, and All Others gaining votes. The dummy variable capturing the reunification shock also is statistically significant across all three models and in the expected directions: the CDU/CSU and the SPD losing votes in a unified Germany and All Others gaining support. The number of minister–president offices that a party holds has a statistically significant, positive effect on SPD support. For the CDU/CSU and All Others, the coefficient is positive but does not reach conventional levels of statistical significance.
Table 1 SUR Models Explaining Vote Share (Zweitstimme) of the CDU/CSU, the SPD, and All Others in 18 German Federal Elections (1953–2017)
Note: Entries are SUR unstandardized coefficients with standard errors in parentheses. +=p<0.1; *=p<0.05; **=p<0.01; ***=p<0.001.
A forecasting model can be evaluated on four key criteria: parsimony, replication, lead time, and accuracy (Lewis-Beck Reference Lewis-Beck2005). The model performs well on the first two—four independent variables easily constructed from publicly available data. On lead time, the model can be estimated at a nontrivial time (i.e., six months before the election). Naturally, the most critical component is accuracy. Our first metric that taps this is model fits as evidenced by the R2, which falls within a respectable range from 0.871 (the SPD) to 0.738 (the CDU/CSU). We also examine how our models fare compared to a naïve model using vote share in the previous election as the only predictor variable. This comparison shows that our models achieve a better fit and smaller errors (see online appendix C, table C5). However, fits provide only the initial chapter of the story. More important, how well does the political-history model predict vote shares?
Table 1 shows that the SPD model has the smallest within-sample mean absolute error (MAE) (2.150), followed by the CDU/CSU model (2.402); the model for All Others exhibits the largest MAE (3.308). The within-sample estimates suggest that the model mostly determines the winners; in terms of CDU/CSU versus SPD contests, the model predicts the winner correctly in 15 of 18 contests (approximately 83%). More encouraging, we observe that general out-of-sample error—as measured by the RMSE—is only slightly greater than the average within-sample MAE for all three blocs. On average, an out-of-sample forecast for the SPD or the CDU/CSU would have about the same level of error (i.e., 2.493 for the SPD and 2.688 for the CDU/CSU). For the All Others model, the RMSE is 3.978.Footnote 2
For robustness, we apply three out-of-sample diagnostic tests. Regarding the jackknife method, in which each election in turn is removed from the analysis and the equations reestimated with the remaining observations, the MAE for all three blocs is slightly lower than the RMSE values (i.e., 2.168 for the SPD, 2.411 for the CDU/CSU, and 3.329 for All Others). For the rigorous one-step-ahead test—which involves estimating the model on a time series up to a particular year and predicting the next election from that base—the CDU/CSU forecasts contain, on average, almost one point more error than the SPD (i.e., 3.209 for the CDU/CSU, 2.389 for the SPD). Nevertheless, the CDU/CSU and the SPD one-step-ahead forecasts are better than the All Others forecast (i.e., MAE one-step-ahead = 3.870).Footnote 3
For our third diagnostic test, we report the largest MAE from both the jackknife and the one-step-ahead procedures. The results reinforce the picture with the SPD model having the smallest largest error on average, followed by the CDU/CSU and then All Others. In summary, we deduce that the political-history model is promising as a prognosis tool. What does it forecast for the upcoming contest in September 2021?
2021 FORECAST
Figure 1 summarizes our 2021 forecast based on the German political-history model, which estimates that the CDU/CSU will win 33.8% of the list vote. For the SPD, it forecasts a vote share of 16.8%—which, if borne out, would be an all-time low in the party’s history.
Figure 1 Political-History Model Forecast of 2021 German Federal Election with 95% Confidence Intervals (Top) and Forecast Vote-Share Change Compared with 2017 Federal Election (Bottom)
For the third remaining bloc of All Others, our model predicts that they will win 49.4%. Briefly, our political- history model forecasts that there is potential for the grand coalition between the CDU/CSU and the SPD to remain a viable governing option, with the two blocs forecast to achieve 50.6%. Naturally, with uncertainty around this estimate, it could be that a grand coalition of the CDU/CSU and the SPD may fall slightly short of a majority. This best estimate declares that it will be a very close contest as to whether they can reach the majority winning post. However, if political history is a reasonable predictor, these two blocs combined are still “in the game” to govern with a majority.
REFLECTIONS
Presenting a forecasting model that focuses solely on a country’s political history represents a significant departure. However, the solid framework of the contemporary German electorate presents an ideal test case. Indeed, our model—which focuses on partisanship, historic shocks, grand-coalition governance, and political-power configurations in the German Länder—passes muster on the traditional criteria for evaluating an election forecast over 18 previous elections. For the 2021 contest, our model predicts the CDU/CSU at 33.8%, the SPD at a historic low of 16.8%, and All Others at 49.4%. Thus, the model suggests the real possibility of another grand coalition—a prognosis clearly at odds with the current public mood in Germany, at least as measured by the polls.
For the 2021 contest, our model predicts the CDU/CSU at 33.8%, the SPD at a historic low of 16.8%, and All Others at 49.4%.
We recognize that we are breaking new ground using only historical patterns as an electoral-forecasting tool, but we are not blind to the limitations of this strategy. We appreciate that our model is better at predicting the performance of the two larger blocs, with more significant variability in the prophecy of the All Others bloc. At the time of writing in June 2021, vote-intention polls stand in stark disagreement with the political-history forecast. They suggest historic lows for both the CDU/CSU and the SPD, with the Sonntagsfrage putting the CDU/CSU at 24% and All Others at 60% (wahlrecht.de 2021). Why is this? Perhaps it is due to the uniqueness of the COVID-19 pandemic; or that Angela Merkel, Chancellor since 2005, is not contesting this election; or, conceivably, the leadership tension within the CDU/CSU between Markus Söder (CSU) and Chancellor candidate Armin Laschet (CDU) is the root cause. Moreover, for the first time, the Green Party has nominated a Chancellor candidate—Annalena Baerbock. In summary, current conditions may seem to favor volatility over stability. In the face of these apparently contrary winds of change, a final result in the 2021 German Federal Election that comes close to the political-history forecast will support the notion that this approach can serve as a powerful addition to the structural-modeling toolbox of election forecasting. In conclusion, and channeling Frank Sinatra: “If I can make it there, I can make it anywhere” (Levy Reference Levy2008, 12).
At the time of this writing in June 2021, vote-intention polls stand in stark disagreement with the political-history forecast.
Supplementary Materials
To view supplementary material for this article, please visit http://dx.doi.org/10.1017/S1049096521000998.
DATA AVAILABILITY STATEMENT
Research documentation and data that support the findings of this study are openly available in the PS: Political Science & Politics Dataverse: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/CZLJS6.
