MODEL

At its core, the political history model for the 2025 German federal election mirrors the model used for predicting the preceding 2021 election (see Quinlan, Schnaudt, and Lewis-Beck Reference Quinlan, Schnaudt and Lewis-Beck2022). However, “[f]orecasting models are never carved in stone” (Gschwend et al. Reference Gschwend, Müller, Munzert, Neunhoeffer and Stoetzer2022, 89). We are not ignorant that our model lacked accuracy in predicting the outcome of the 2021 election. Model tinkering has been explored in the election forecasting literature (e.g., Campbell Reference Campbell2004; Lewis-Beck and Tien Reference Lewis-Beck and Tien2008). Such adjustments can be reasonable, but they must be prudent and theory-driven and avoid overstretching (Lewis-Beck and Tien Reference Lewis-Beck and Tien2008).

For 2025, we adapt the PH model by adding a variable that captures the electoral influence of the 2015 “refugee crisis” as a potent political shock to electoral politics in Germany.

Theory

The PH model for Germany seeks to predict bloc support for three different political blocs—the Union, the SPD, and All Others—based on five different variables that capture influential and persistent political constellations and events throughout the country’s historical development. Thus, it takes the following form:

$$ {\displaystyle \begin{array}{l}\mathrm{Bloc}\ \mathrm{support}=\mathrm{f}\;\Big(\mathrm{Short}\hbox{-} \mathrm{Term}\ \mathrm{Partisanship}\\ {}\hskip10.2em +\mathrm{Grand}\hbox{-} \mathrm{Coalition}\ \mathrm{Governance}\\ {}\hskip10.2em +\mathrm{Bloc}\ \mathrm{Number}\ \mathrm{of}\ \mathrm{Land}\ \mathrm{Minister}\ \mathrm{Presidents}\\ {}\hskip10.2em +\mathrm{German}\ \mathrm{Reunification}\\ {}\hskip10.2em +``\mathrm{Refugee}\ \mathrm{Crisis}"+\mathrm{e}\Big).\end{array}} $$

The first variable, “short-term partisanship,” is based on the presumption that current support for political blocs is best understood as an extension of past support. Theoretically, this notion captures the tendency of many voters to form long-lasting bonds with specific parties or blocs (cf. Dalton Reference Dalton2016). We operationalize partisanship as each political bloc’s vote share at the preceding election.

Second, our model considers “grand-coalition governance” phases between the Union of CDU/CSU and the SPD as a political constellation with far-reaching implications. Theoretically, grand coalition governance implies that both blocs must cut down on their electoral promises. Consequently, phases of grand coalition governance between the two dominant players in the German political arena should come with more electoral costs for the Union and the SPD and thus be followed by electoral gains for smaller blocs at the next election. We operationalize grand coalition governance with a dummy variable reflecting the presence or absence of a grand coalition during the preceding legislative period.Footnote ⁴

Third, in a federal system such as Germany, the electoral fortunes of different political blocs at the federal level should be related to how powerful these blocs are at the level of the German Länder (Kayser and Leininger Reference Kayser and Leininger2017). As the most prominent and politically influential actors at the subnational level, it can be expected that governance costs (or benefits) will be primarily tangible for those blocs holding the post of Land minister-president. To capture such power dynamics of multilevel governance, our model includes a variable reflecting the total number of each political bloc’s Land minister presidents six months before the 2025 election.

Fourth, our model considers the long-term electoral repercussions of German reunification. Theoretically, this unique historical event can be understood as a “super shock” to the German political system, comprising a supply-side, a demand-side, a sociocultural, and an economic shock, with persistent effects on German voting behavior (for details, see Quinlan, Schnaudt, and Lewis-Beck Reference Quinlan, Schnaudt and Lewis-Beck2022, 92). Our model captures this super shock using a dummy variable reflecting elections before and after 1990.

Fifth and last, our model incorporates the 2015 “refugee crisis” as a more recent political event with potent and still unfolding implications for electoral politics in Germany. Ever since former chancellor Angela Merkel’s (CDU) famous proclamation “We can do this” in August 2015, the “refugee crisis” has been a recurring epitome of political contestation. In essence, the “refugee crisis” has initiated nothing less than a reconfiguration of the German party system, whose electoral implications have crystallized more clearly over the last decade. On the supply side, these implications pertain to an increasing polarization between political parties on the immigration issue, including the normalization of far-right positions (cf. Valentim Reference Valentim2024) as well as the split-up of existing parties (e.g., Die Linke) and the emergence of new electoral competitors (e.g., BSW) for, among other things, immigration-related reasons (cf. Thomeczek Reference Thomeczek2024, 537). On the demand side, the implications include a stronger inclination for immigration-based issue voting (cf. Schnaudt and Stecker Reference Schnaudt and Stecker2022). The refugee crisis thus constitutes another critical juncture in Germany’s political history. Our model captures this new period in Germany’s political history with a dichotomous variable for elections held before and after 2015.

MODEL PERFORMANCE 1953–2021

Table 1 details the models on which our central forecast is based. The coefficients broadly are in sync with our theoretical suppositions. We see that the coefficients for short-term partisanship (i.e., vote share in the previous election) exhibit the anticipated positive direction and achieve statistical significance at conventional levels (p<0.05) across all blocs. The dichotomous variable capturing the Reunification shock is statistically significant across all three models and in the presumed directions—the Union and SPD losing support ceteris paribus in a unified Germany and All Others acquiring support. The migrant crisis shock is strongly associated with the support of the Union and All Others, with the support of the Union significantly less on average and the support of All Others significantly higher postcrisis. However, there is no notable association with SPD support. When the outgoing government is a Grand Coalition, we also observe the expected patterns evidenced by the negative coefficients for the Union and SPD and the positive coefficient for All Others. However, none attains statistical significance at p<0.05. The number of minister-president offices a party occupies is positively associated with SPD support, although the pattern is marginally statistically significant at p<0.1. For the Union and others, the coefficient is negative but does not reach conventional levels of statistical significance. The lack of statistical significance for these latter coefficients essentially derives from a problem of high collinearity, which makes it difficult to sort out the weight of the relative variable effects.Footnote ⁹ This does not, however, undermine the quality of the theoretical specifications embodied in the equations. Nor does it prevent prediction of the election outcome, our forecasting goal, as the model fit for each equation is high (cf. C. Lewis-Beck and Lewis-Beck Reference Lewis-Beck and Lewis-Beck2015, chaps. 3 and 4).

Table 1 Central Political History (PH) Forecast Model: Seemingly Unrelated Regression models explaining vote share (Zweitstimme) of the Union, SPD, and All Others in 19 German Federal Elections 1953–2021

Note: Entries are unstandardized coefficients of seemingly unrelated regression with standard errors in parentheses. One-step-ahead analysis conducted on elections from 1994 onward (~42% of data).

⁺p<0.1; *p<0.05; **p<0.01; ***p<0.001

There are four main standards to consider in classifying a model’s prediction capacity: parsimony, replication, lead time, and accuracy. The PH models have five consistent variables for each bloc, and the data are readily available and replicable. The PH forecast can be made six months in advance, meaning the projections will be nontrivial.Footnote ¹⁰ But the Shangri-la of forecasting is accuracy. There are several means of investigating this. First, the model’s fit to the data. They are solid—an R² of 0.84 for the Union model, 0.81 for the SPD estimation, and 0.91 for the All Others bloc. Second, we test whether the model performs better than a naïve model with merely previous vote share as the sole predictor. This test shows that the PH models achieve a better fit and more minor forecast errors (see Table C5, Appendix C). Third, we examine the within-sample mean absolute error (MAE), which treats all errors equally and provides a yardstick of the conventional prognostication error. The MAE for all three blocs—2.244 for the Union, 2.576 for the SPD, and 3.013 for All Others—is lower than the corresponding values for root mean square error (RMSE). Fourth, we assess the RMSE, which is a stricter test of average error because it gives more weight to the more significant errors from the model and is seen as a typical yardstick for how much error we might expect.

The Union RMSE of 2.6 shows that, on average, the PH model estimates the share of the Union vote within 2.6 points of the official result. The RMSE increases to 3.571 for All Others, whereas it comes in at 3.123 for the SPD. Finally, we decipher how often the model correctly predicts the plurality winner of the vote and, thus, who will be the most likely actor to lead the government. Encouragingly, the within-sample analysis correctly predicts which bloc—the Union or the SPD—will win the most votes on 84% of occasions.

Within-sample estimates give us a firm clue about the precision of a model and its fit with the data. Still, they are known to be optimistic. Out-of-sample estimates are more concrete tests, as they involve predictions by excluding data about the contest in question, either temporally or spatially, thus mimicking the situation predictors contend with. To test the veracity of our models, we apply the jackknife method. This standard out-of-sample method involves leaving out one observation from the data and then producing a forecast of the excluded case based on the remaining information. On this basis, we report two diagnostics. The first is the MAE for each of the three blocs (2.279 for the Union, 2.629 for the SPD, and 3.015 for All Others). Under this scenario, the model primarily sorts out the winners between the Union and SPD, correctly predicting the winner in 14 of 19 contests (~74%).

The one-step-ahead method is a firmer out-of-sample test, as it measures prophetic precision by generating forecasts sequentially in chronological order, relying entirely on data that would have been accessible to the forecaster when making an ex ante prediction. As we might expect, the MAE for these specifications is significantly higher than other specifications for the Union and the All Others bloc, coming in at 3.712 and 4.255, respectively. The largest MAE errors are 4.857 for the Union and 4.695 for All Others. Encouraging for the SPD, the MAE is 2.476, even smaller than the MAE under the jackknife procedure, and the largest MAE comes in at 2.605. From this analysis, we deduce that the updated PH model has some promise as a prognosis tool.

PH FORECAST FOR WAHL ‘25

Figure 1 displays the 2025 PH forecast for the German Federal election and suggests that the result will be much closer between the two dominant parties than is currently anticipated by the polls, with the Union and the SPD forecast tied at 26% each. For the SPD, the PH prognosis is substantially ahead of current opinion poll projections, with recent polls clocking support for the party at around 15%. All Others are projected to win 48%. And although the PH model point estimate anticipates a gain in vote share for the Union vis-à-vis its 2021 outing, it is much less than the polls predict. Thus, the 2025 PH model forecast indicates that the battle to lead the next government may be closer than most commentary assumes (granted, the AfD will not be tasked to lead the government even if they should win the most votes). The point estimates from the PH model suggest that a grand coalition between the Union and SPD will likely achieve a parliamentary majority at 52%. That said, the error bands around our estimates, as represented by the RMSE, mean we could be off by roughly three percentage points, one way or another, so this isn’t a sure bet.

Figure 1 Political-History (PH) Model Forecast of 2025 German Federal Election Zweitstimme Vote Share with 95% Confidence Intervals (Top) and Forecast Zweitstimme Vote-Share Change Compared with 2021 Federal Election (Bottom).

REFLECTIONS

The 2021 German elections were a low watermark for academic election forecasts, with all the scientific models falling short. The 2025 German federal election presents an opportunity to recapture lost ground. That said, the contextual circumstances surrounding Wahl ’25 suggest this contest could also prove difficult to prognosticate accurately. Nevertheless, we have taken on the challenge to forecast the 2025 election with a political history (PH) model, with an eye toward theoretical and empirical improvement. We altered the PH model by incorporating the unfolding shock that the 2015 “refugee crisis” has delivered. We demonstrate that this modified PH model shows promise in predicting the performance of the Union, the SPD, and All Others in 19 German elections, correctly forecasting which party will emerge with a plurality of the vote three-quarters of the time and estimating each bloc performance on average to within 3–4 percentage points of the eventual result. This is no mean feat, given that the PH model relies on structural and institutional dynamics and does not focus on public opinion in the lead-up to the poll. For 2025, this modified PH model, our central forecast, suggests a cliffhanger contest, with the point estimates indicating both the Union and SPD coming in at 26% each and All Others at 48%. A grand coalition seems on the horizon if this central forecast is borne out.

We recognize that 2025 is a tough test for the PH forecast endeavor. Critics may also challenge our modification to the PH model (including the migrant crisis as a potent shock to the system) as premature. A more general approach for incorporating the electoral influence of crises into our PH model could be to adopt a notion of “Poly Crisis,” incorporating several emergencies of the past 17 years ranging from the GFC, the migrant crisis, COVID, the energy crisis, and the cost-of-living crisis into our PH assumptions (see Table C7, Appendix C). A prognosis on this basis would suggest a significantly better result for the Union at the expense of All Others. (However, all potential constellations of the PH model project a considerably better result for the SPD than all the polls.) Wahl ’25 provides a valuable opening to evaluate how recent historical developments can be methodically integrated into ex ante prognosis models.