From today to polling day in Germany
Abstract. The next Bundestag election is due by 25 March 2029, so today's polls are not a forecast of polling day. Testing the British pull-back towards each party's last result on polls at ten distances before the five elections from 2009 to 2025, we found it made German forecasts worse at every distance, because change between Bundestag elections has been large and has mostly lasted. The polling-day forecast therefore keeps today's shares and widens the uncertainty with the time left, by a curve fitted to how far polls have missed at each distance.
Summary
Everything else in the Germany section answers the question "what if the election were held today?". The next Bundestag election is not today: at the latest it is due on 25 March 2029, 855 days after the most recent poll. This paper describes how we get from today to polling day.
In Britain we found that the older parties finish much closer to their previous result than their mid-term polls suggest, and we pull them part of the way back (From today to polling day: how far British polls move before a general election). We ran the same test on German polls, at ten fixed distances from one month to three years before each of the five Bundestag elections from 2009 to 2025. It failed. At every distance, pulling parties back towards their previous result made the forecast worse, because in Germany the change between elections has been large and has mostly lasted. So our polling-day forecast keeps today's shares as they stand and adds only uncertainty, which grows with the time left: polls this far out have missed each party by about 5.2 points by the time votes were cast.
What has changed since this paper was written
Updated 9 November 2026. The rest of the paper is as first published, except where a sentence below says it was corrected.
- Only elections from 2000 shape the forecast (28 September). Earlier elections stay in the long view and on the history pages but no longer feed any test, correction or error size. In Germany the tests now run on the seven elections from 2002 to 2025. The correction stays off (it was closer at only 2 of the 7) and the Wahlkreis swing stays uniform.
The question: today is not polling day
A poll taken two years before an election records how people say they would vote at that moment, with no campaign under way and no choice of Chancellor candidates in front of them. The literature on the "timeline" of elections finds that polls become steadily more informative as polling day approaches, and that early polls carry a good deal of movement that has gone by the time votes are cast (Wlezien et al., 2013; Erikson and Wlezien, 2012). Gelman and King (1993) framed the same puzzle for American presidential elections: campaign polls swing a great deal, yet results are predictable well in advance, because campaigns tend to bring voters back to where their circumstances would have put them.
Whether that pattern holds in a given country is an empirical question. The error we measure at final polls (Measuring German polling error, 2005–2025) answers a narrower one: how wrong were the last polls of a campaign? It says nothing about how far opinion moves in the years before. The polling-day forecast therefore needs its own evidence, and it is kept apart from the "today" forecast on the site.
The evidence: polls at fixed distances before five elections
Source
The evidence is every national Bundestag voting-intention poll in our series, 3,985 polls covering the parliaments that ended in the elections of 2009, 2013, 2017, 2021 and 2025. Where a pollster's own archive exists we use it: Forschungsgruppe Wahlen's Politbarometer releases, infratest dimap's Sonntagsfrage series and Allensbach's table from July 2010. Other pollsters come from Wikipedia's per-parliament poll tables. The parliament of 2005 to 2009 is thin: no English per-parliament table exists for it, so most of its polls are Forschungsgruppe Wahlen's and infratest dimap's. The parliament before 2005 is not covered at all, which is why 2005 is missing from this test although it is in the final-poll record. The German data: results, boundaries, polls and licences gives the sources in full.
Horizon points
For each of the five elections we average the polls at ten fixed distances before polling day: 30, 61, 91, 182, 273, 365, 548, 730, 913 and 1,095 days, from about one month to three years. Distances are counted back from the actual election date, so for 2025, an early election, they run back from 23 February 2025.
At distance \(d\) the window is the 60 days centred on the date \(d\) days before the election. Within it, each pollster's polls are averaged, and the pollsters' averages are then averaged with equal weight, so that a pollster publishing weekly does not outweigh one publishing monthly. The parties are the Union, the SPD, the AfD, the FDP, the Greens, the Linke and the BSW. Each record, a horizon point, holds the election, the distance, the party, the poll average \(\bar{y}_{p,t}(d)\), the party's result at the previous Bundestag election \(v_{p,t-1}\) and its result at this one \(v_{p,t}\).
A party with no previous result cannot be pulled back towards it, so its points are left out of the test: the AfD in 2013 and the BSW in 2025. That leaves 280 horizon points: five or six parties at each of ten distances at each of five elections. Each window has between 7 and 77 polls, and between 2 and 13 pollsters. The low figures belong to 2009, where most distances rest on two pollsters.
The prior
The alternative to a poll taken years out is to assume nothing has changed since the last election. We call the previous result the prior, because it is what one would predict with no polls at all. A forecast for polling day can then be framed as a choice of how far to move from the prior towards today's polls, and the data can say how far that has been worth moving at each distance.
What the data show
The table sets out the results. "Polls' miss" is the average absolute difference between the poll average at that distance and the result, per party. "Previous result's miss" is the same for the prior, which does not depend on the distance. "Pull-back's miss" is the model described below, each election predicted from a curve fitted to the other four. "RMSE" is the root-mean-square miss of the polls as they stand, which is what we use to set the uncertainty.
| Days before polling day | Elections | Polls' miss | Previous result's miss | Pull-back's miss | RMSE |
|---|---|---|---|---|---|
| 30 | 5 | 1.43 | 5.11 | 1.48 | 1.84 |
| 61 | 5 | 2.16 | 5.11 | 2.24 | 2.93 |
| 91 | 5 | 2.45 | 5.11 | 2.54 | 3.39 |
| 182 | 5 | 2.78 | 5.11 | 2.88 | 4.01 |
| 273 | 5 | 2.93 | 5.11 | 3.13 | 4.08 |
| 365 | 5 | 2.94 | 5.11 | 3.18 | 4.07 |
| 548 | 5 | 3.20 | 5.11 | 3.57 | 4.01 |
| 730 | 5 | 4.53 | 5.11 | 4.60 | 5.40 |
| 913 | 5 | 4.42 | 5.11 | 4.52 | 5.48 |
| 1095 | 5 | 4.52 | 5.11 | 4.62 | 5.59 |
Two findings stand out.
The polls beat the previous result at every distance, even three years out. The previous result missed by 5.11 points per party on average. The polls missed by 1.43 at one month, between 2.2 and 3.2 from two months to eighteen months, and about 4.5 at two to three years. In Britain, from about six months out the polls missed the older parties' results by about as much as the previous result did, and at several distances by more. In Germany the previous result is a worse guide than even a three-year-old poll.
Change between elections has been large, and it has lasted. The prior misses by more than five points per party because German elections have moved a long way. Some examples from the horizon points:
- 2013. The FDP had won 14.6% in 2009. A year before the 2013 election it was polling 4.5%; it finished at 4.8% and left the Bundestag. Pulling it back towards 14.6% would have been badly wrong.
- 2017. The AfD had won 4.7% in 2013. A year out it polled 13.4%; it won 12.6%. The Union, on 41.5% in 2013, polled 32.7% a year out and won 32.9%.
- 2021. The Union had won 32.9% in 2017. A year out it polled 35.9%, and it finished at 24.1%. The SPD polled 15.8% a year out and won 25.7%, against 20.5% last time. This is the one election at which the previous result was a useful anchor, because the parties' mid-term positions reversed.
- 2025. The AfD had won 10.3% in 2021. A year out it polled 18.2%; it won 20.8%. The SPD fell from 25.7% to 16.4%, and a year out it was already polling 15.5%.
In Britain the pull-back works because mid-term surges and slumps of the older parties have mostly been temporary. In Germany, over these five elections, they mostly were not. The party system has changed: new parties have entered the Bundestag (the AfD in 2017) or come close (the BSW in 2025), the FDP has left and returned, and the Union and the SPD, which shared more than two thirds of the vote in 2005, have lost combined ground at four of the five elections since. A mid-term poll in that setting has usually been a better guide to the next result than the last result was.
The pull-back curve and its test
Form
We tested the same model as in Britain. The projected polling-day share is a weighted average of today's share and the previous result:
where \(y_p\) is today's share, \(d\) the days to polling day, \(f\) the floor (the share of the change since the last election that counts however far out we are) and \(\tau\) a time constant in days. The weight is 1 on polling day and falls towards \(f\) as the distance grows. One difference from Britain: there the curve is fitted to the Conservatives, Labour and the Liberal Democrats only, and the insurgent parties are carried forward as polled. In Germany it is fitted to every party with a previous result.
Fitting
The two parameters are chosen by least squares over all 280 horizon points, searching \(f\) from 0 to 1 in steps of 0.01 and \(\tau\) from 5 to 600 days in steps of 5:
The best fit on all five elections is \(f = 0.65\) and \(\tau = 425\) days. That curve counts about 98% of the change since the last election at one month, 93% at three months, 88% at six months, 80% at one year and 69% at two and a half years. It is a much weaker pull than Britain's, where about half the change counts from three months out.
The fit is unstable. Refitted with each election left out in turn, the floor ranges from 0.52 to 0.83 and the time constant from 125 to 600 days. A curve whose parameters move that much when one election is dropped is describing the elections it was fitted on, not a regularity that carries over.
The test
As in Britain, each election is predicted with the curve fitted to the other four only (Stone, 1974), and the pull-back is kept only if, averaged over every distance, it beats the polls as they stand by more than 0.01 points per party. It does not. It was worse than the polls on average at every one of the ten distances, by between 0.05 points (30 days) and 0.37 points (548 days).
Election by election, the pull-back helped at two of the five at each distance, but not always the same two. It helped in 2021 at every distance, because the Union's and the SPD's mid-term positions reversed. It helped in 2025 up to six months out and in 2013 from nine months out. It made 2009 and 2017 worse at every distance. That is not a pattern one could use in advance.
Because the test fails, the site sets the floor to 1: the polling-day shares are today's shares. The fitted curve is shown on the election-day page for reference, labelled as not used.
We did not test a version restricted to the two traditionally largest parties, the Union and the SPD, on the British model. It is a natural next step. The 2021 example suggests it might sometimes help; the 2017 and 2025 examples, where both parties' mid-term positions held or went further, suggest it would often not.
Drift: uncertainty that grows with the time left
The pull-back fails, but the uncertainty it was meant to address is real. A poll three years out misses by about three times as much as a poll one month out. The polling-day forecast keeps that uncertainty and calls it drift.
The drift curve
Drift is measured by the root-mean-square miss of the polls as they stand at each distance, the RMSE column of the table above. It grows with distance, but more slowly than in proportion, and a logarithmic curve fits it:
fitted by ordinary least squares over the ten distances. It gives 2.1 points at one month, 3.8 at six months, 4.4 at one year and 5.1 at two years. At 855 days it is 5.2 points. Two details come from the code. Under 30 days the drift is scaled down in proportion to the days left, so that it reaches zero on polling day, when only the final-poll error remains. And it is never less than half a point.
The same drift applies to every party. In Britain the insurgents get a larger drift than the older parties; in Germany we have not split the parties, and with five or six parties per election and five elections, a split would rest on very little.
The miss at each distance, election by election, shows how much the drift varies:
| Days out | 2009 | 2013 | 2017 | 2021 | 2025 |
|---|---|---|---|---|---|
| 30 | 0.97 | 1.24 | 2.25 | 1.42 | 1.28 |
| 91 | 1.61 | 1.62 | 3.05 | 3.63 | 2.33 |
| 182 | 1.42 | 1.97 | 3.80 | 4.28 | 2.45 |
| 365 | 2.13 | 2.89 | 1.65 | 5.80 | 2.20 |
| 730 | 4.34 | 4.84 | 4.34 | 5.54 | 3.59 |
| 1095 | 3.06 | 5.14 | 4.32 | 4.51 | 5.58 |
Table: Polls' average miss per party at selected distances and elections, points. The election-day page shows all ten distances.
2021 stands out: from six months to two years out, the polls missed by between 4.3 and 5.9 points per party, because the SPD's late rise and the Union's fall came in the last year. 2017 shows the opposite shape, with the polls closer at one year (1.65) than at six months (3.80): six months out the SPD's short-lived surge was at its height, with the party polling 31.3% against the 20.5% it won. The curve smooths these out; the real paths are lumpier.
Polling-day ranges
The polling-day forecast runs the same simulation as the "today" forecast (Simulating German polling error: where the chances come from), with 1,000 runs, and adds the drift to each party's spread:
where \(\text{RMS}_p\) is the party's root-mean-square final-poll miss. Because the floor is 1, the centre is today's shares, and the whole difference between the two forecasts is the width.
| Party | Seats today | 80% range today | 80% range, polling day | Gets in, polling day |
|---|---|---|---|---|
| AfD | 193 | 176–204 | 134–226 | over 99% |
| CDU/CSU | 152 | 126–170 | 96–190 | over 99% |
| Greens | 111 | 99–120 | 62–148 | 99% |
| SPD | 99 | 86–109 | 50–135 | 98% |
| Linke | 74 | 63–82 | 18–111 | 95% |
| BSW | 0 | 0–0 | 0–65 | 35% |
| FDP | 0 | 0–40 | 0–67 | 40% |
The difference is large. At two to three years out the drift is about five points, against one to one and a half points of final-poll error for most parties. The seat ranges become several times wider, and the chances of getting in change character.
For a party near the 5% line, a wide spread flattens the chance of getting in towards a coin toss. With the drift at about 5.3 points and the FDP's final-poll error of 1.41, the combined spread is about 5.5 points, and equation (5) of Simulating German polling error: where the chances come from gives a party on 4% about a 43% chance of reaching 5%, and a party on 6% about 57%. A small party well below the line today has a much better chance of getting in on polling day than on today's polls, and a mid-sized party well above it has a real chance of falling out. Compare the "Gets in" columns of the two tables on the election-day page to see this for today's parties.
Coalition chances move the same way. A combination that holds a majority in almost every "today" run holds it in far fewer polling-day runs, because the parties it needs and the parties it hopes will miss the line both become much less certain.
The date, and the possibility of an early election
A polling-day forecast needs a date. Article 39 of the Basic Law sets the Bundestag's term at four years from its first meeting, and requires the next election no earlier than 46 and no later than 48 months after the term began. The current Bundestag first met on 25 March 2025, so a regular election falls between late January and 25 March 2029. The site uses the latest date, 25 March 2029, as its default. That is not a prediction that the election will be held on the last possible day. It is the one date fixed by law, and it is the cautious choice: the longest horizon gives the widest ranges.
An earlier election is possible. If the Chancellor loses a vote of confidence, the Federal President may, on the Chancellor's proposal, dissolve the Bundestag (Article 68), and an election follows within sixty days. That is how the elections of 2005 and 2025 came about. An early election shortens the horizon and narrows the drift: at one year it would be about 4.4 points, at six months about 3.8. The site notes this on the election-day page; it does not try to forecast whether the coalition will last.
The distance \(d\) is counted from the date of the latest poll in the average, not from the day the page is viewed.
Limitations
Five elections. The test rests on five elections, one of which (2009) has only two pollsters at most distances. The failure of the pull-back is consistent across distances, but five elections cannot rule out that it would help at the next one.
The drift includes the final-poll error. The RMSE at each distance measures the polls' miss against the result, which already contains the error that remains at the end of a campaign. Equation (4) then adds the final-poll error again. At long range the double count is small; at one or two months out it makes the ranges somewhat too wide. We prefer ranges that err on the wide side, but the choice is not calibrated.
Equal drift for all parties. A new or small party can move much further, in proportion, than a large one. The BSW, founded in 2024, came within a fraction of 5% at its first Bundestag election; the FDP fell from 14.6% to 4.8% in one parliament. A single drift for everyone understates the relative uncertainty of small parties and may overstate it for the large ones.
Independent drifts. As with the final-poll error, drift is drawn independently for each party. In practice one party's gain is usually a particular rival's loss.
No fundamentals. The model has no economic data, no leader ratings and no state election results. Each would add parameters that five elections cannot estimate.
New parties. A party founded after today's polls has no share to carry forward. The forecast cannot anticipate it, and German politics produced the AfD in 2013 and the BSW in 2024.
Conclusion
A poll two or three years before a Bundestag election is a rough guide to the result, missing each party by four to five points. But it has been a better guide than the last result at every distance we tested, and the British device of pulling parties back towards the last result made German forecasts worse. The polling-day forecast therefore keeps today's polls as its centre and widens the uncertainty with the time left, by a curve fitted to how far the polls have missed at each distance before the elections of 2009 to 2025. The result, on the election-day page, is a forecast of a different quantity from the "today" headline, built on a thinner body of evidence, and it is shown separately for that reason.
References
Erikson, R. S. and Wlezien, C. (2012). The Timeline of Presidential Elections: How Campaigns Do (and Do Not) Matter. Chicago: University of Chicago Press.
Gelman, A. and King, G. (1993). Why are American presidential election campaign polls so variable when votes are so predictable? British Journal of Political Science, 23(4), 409–451.
Grundgesetz für die Bundesrepublik Deutschland (Basic Law), Articles 39 and 68.
Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions. Journal of the Royal Statistical Society: Series B, 36(2), 111–147.
Wlezien, C., Jennings, W., Fisher, S., Ford, R. and Pickup, M. (2013). Polls and the vote in Britain. Political Studies, 61(S1), 66–91.
How to cite
Polling Forecast (2026). From today to polling day in Germany. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/germany-election-day. Live figures change as new polls arrive; the date above is when the text was last revised.



