Germany Bundestag electionby 25 Mar 2029in 855 days CDU/CSU + SPD + Greens 99% Calendar Track record Research
Research · Germany
EN DE

If we're wrong: replaying past Bundestag polling misses on today's polls

Abstract. We apply each Bundestag election's final-poll miss from 2002 on (the model uses no election before 2000) to today's German poll average and run every reading through the same seat engine. German misses are modest, but the 5% line turns them into jumps: a replay that lifts a small party over the threshold hands it thirty or more seats and changes which coalitions add up. The same replays change the Erststimme winner in some Wahlkreise, which the map and seat pages mark as at risk; under the 2023 law that changes who sits, not how many seats each party gets.

Polling Forecast · updated 9 November 2026 · 17 min read · Germany, Bundestag, lenses, polling error, scenarios, 5% threshold, Wahlkreise at risk, FDP, BSW, what-if

Summary

A lens is one stated assumption about how wrong today's polls are. For Germany the site shows nine readings of the same polls: the polls as they stand, our projection (which, because correcting German polls failed its test, is the same numbers), and seven replays that take the final-poll miss of each Bundestag election from 2002 (the first after the model's cutoff year, 2000) and apply it to today's average. Every reading goes through the same seat engine: proportional swing to the Länder, uniform swing to the 299 Wahlkreise, then the 2023 seat law. We show the readings side by side and do not average them.

The German replays behave differently from the British ones. The misses themselves are modest, about a point per party on average, and apart from the CDU/CSU's large miss in 2005 they move the big parties by a few dozen seats at most. The exception is any party close to 5% of the Zweitstimmen. A replay that adds a point and a half to such a party can lift it over the line and hand it thirty or forty seats that it would otherwise not have, taking them from everyone else. The spread between readings is therefore lumpy: most replays agree, and one or two jump. The same replays also change the winner in some Wahlkreise, 102 of them today, which the map and the seat pages mark as "at risk". Under the 2023 law a changed Wahlkreis winner does not change a party's seat total, so "at risk" means something narrower than in Britain.

The British counterpart of this paper is Lenses: reading the same British polls several ways. How the misses were measured is in Measuring German polling error, 2005–2025, and why we do not correct German polls is in How we read German polls: the 5% line and the 2023 seat law.

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 lenses on the map (November). The replays are also on the map as "The polls miss again", beside "You set the vote".
  • When the polls missed (2 November): Six points lost in the last week: Merkel's vanishing lead and Germany's 2005 polling miss.

Why Germany needs lenses when its polls are fairly accurate

Final Bundestag polls have been closer to the result than British ones. Across the seven elections from 2002 to 2025 the average of the final polls missed by between 0.8 and 2.3 points per party:

ElectionPolls' average missCorrected, leave one outVerdict
20020.951.15further out
20052.332.37further out
20091.100.97closer
20130.941.09further out
20171.481.40closer
20210.801.13further out
20250.891.01further out
Average1.211.30
Live Correcting German polls, each election left out in turn Computed from the database, polls to 17 November 2026.

The same table shows why we do not correct them. Taking off half the average past miss, worked out each time from the other six elections, was closer at two elections and further out at five. On average it was worse, 1.30 points per party against 1.21. The misses do not keep their direction: the CDU/CSU was overstated in 2005, 2009, 2017 and 2025 and understated in 2002, 2013 and 2021, and the same kind of reversal appears for every party.

Party19531972197619982002200520092013201720212025Average
SPD−0.2−0.1−1.6−0.5+0.5−0.8+2.3+1.0+1.2−0.3−1.2+0.4
CDU/CSU0.0+1.6+0.4+2.6−1.5+6.3+1.0−2.0+3.0−1.7+1.2+0.9
FDP–−2.3+0.1−0.7+0.9−2.8−1.3+0.6−1.4+0.20.0−0.5
Greens–––−0.3−1.3−0.9−0.4+1.0−1.3+0.8+1.4−0.1
Linke–––−0.6+0.5−0.7−0.5+0.2+0.4+1.5−2.0−0.1
AfD–––––––−0.8−1.6+0.40.0−0.5
BSW––––––––––−0.4−0.4
Live Bundestag: final-poll average minus result, by party (points) Computed from the database, polls to 17 November 2026.

So the headline figure is simply the poll average. That does not mean the polls will be right. It means that on the evidence of seven elections nobody can say in advance which way they will be wrong. A lens answers a narrower and more useful question: what happens to the Bundestag if the polls miss this time exactly as they did at a known past election. In Germany the answer depends less on the size of the miss than on where it falls, because a small miss on a party near 5% moves far more seats than a large miss on a party at 25%.

What a lens is

The replay

Let \(M\) be the seven parties the German forecast models: CDU/CSU, AfD, SPD, Greens, Linke, BSW and FDP. Let \(v_p\) be today's average for party \(p\), which is each pollster's latest poll within 21 days of the newest, equally weighted, and let \(S = \sum_{p \in M} v_p\) be their combined share. Everything else is "others", \(100 - S\).

For a past election \(t\), the error of the final polls on party \(p\) is the average of the final polls minus the official Zweitstimmen share,

\[ e_{p,t} = \bar{y}_{p,t} - r_{p,t}, \tag{1} \]

so a positive error means the polls overstated the party. The replay of election \(t\) first subtracts that error from every party it measured, with a floor of 0.1 points:

\[ \tilde{v}^{t}_{p} = \begin{cases} \max\left(0.1,\; v_p - e_{p,t}\right) & p \in P_t,\\[2pt] v_p & \text{otherwise,} \end{cases} \tag{2} \]

and then rescales the seven parties back to their original combined share:

\[ v^{t}_{p} = \tilde{v}^{t}_{p} \times \frac{S}{\sum_{q \in M} \tilde{v}^{t}_{q}}. \tag{3} \]

Equations (2) and (3) are what LensMath.Replay does. Two details matter.

The first is the set \(P_t\). A party is measured at election \(t\) only if it had a result and appeared in at least half of that election's final polls. The CDU/CSU, SPD, Greens, FDP and Linke are measured at all seven elections. The AfD is measured from 2013, when it first stood, and the BSW only in 2025. So the replays of 2002, 2005 and 2009 leave the AfD unadjusted, and only the 2025 replay touches the BSW. Six of the seven replays therefore say nothing about the BSW except through the rescale, and three say nothing about the largest party in today's polls.

The second is that the rescale in (3) keeps "others" where they are. In the British lenses the rescale is to 100 and every party, including others, moves when the adjustments do not cancel. Here only the seven modelled parties are rescaled among themselves. If a replay's errors do not sum to zero, the rescale spreads the difference over all seven in proportion to their adjusted shares, including parties the replay did not touch. A consequence is that no replay can express the polls having overstated or understated the modelled parties as a group against the small parties: that part of each past miss is removed by the rescale.

Everything after the replay is the same for every reading

The adjusted shares go into the same engine as the projection (GermanyForecast.Run), described in How we read German polls: the 5% line and the 2023 seat law and From votes to seats: the Bundestag seat law:

  1. Länder. Each party's 2025 Zweitstimmen share in each Land is multiplied by the ratio of its national share to its 2025 national share (proportional swing), the Land's shares are rescaled to 100, and each party's Land shares are then scaled so that together they give back its national share (the CDU and CSU together). The CDU/CSU's share goes to the CSU in Bavaria and to the CDU elsewhere.
  2. Wahlkreise. Each main party's 2025 Erststimmen share in each Wahlkreis moves by the party's national change in points (uniform swing); smaller parties and independents are scaled by the ratio instead. The winner is the party with the most. On six pairs of elections, 2002→2005 to 2021→2025, uniform swing named 1,602 of 1,794 Wahlkreis winners and proportional swing 1,521, so uniform swing is used.
  3. Seats. The 2023 seat law: 630 seats by Sainte-Laguë/Schepers among parties with 5% of the Zweitstimmen, three Wahlkreis winners or national-minority status; each party's seats divided among its Land lists; Wahlkreis winners seated in order of their Erststimmen share until the party's seats in that Land run out.

No reading changes the baseline, the swing model or the law. Any difference between two rows of the table below comes only from the assumed polling error.

The readings on the German pages

The polls as they stand

The poll average with no adjustment. This is the figure most coverage reports, and it is the one other German forecasts start from.

Our projection

Our rule in every country is the same: measure the past misses, take off half, and keep the correction only if it would have made past forecasts better with each election predicted from the others. In Britain the rule passes for the raw answers (Correcting the raw answers: the method behind our UK projection). In Germany it fails, so the projection lens has an empty adjustment and is labelled "Our projection (the polls, uncorrected)". Its row in the table is identical to the polls row. We keep both rows so that the German table has the same shape as the British and French ones, and so that it would show a difference at once if a future backtest passed and the correction were switched on. The code checks this every time the forecast is built (ErrorHistory.CorrectionHelps), so no one has to remember to do it.

Seven past misses

One replay for each of 2002, 2005, 2009, 2013, 2017, 2021 and 2025, built from the 57 final polls in our record, between five (2002 and 2005) and thirteen (2025) per election. Each is labelled "If the polls miss like 2002" and so on. Elections before 2000 (1953, 1972, 1976 and 1998 have final polls in the long view) are shown on the track-record page but not replayed: the model does not use them.

The reader's own

The what-if tool lets a reader set an error for each of the seven parties, from −6 to +6 points in half-point steps. It applies equations (2) and (3) to the poll average, runs the same seat engine, and shows the seats, the coalitions that reach 316, and the Wahlkreise whose winner differs from the projection. Its starting presets are the seven replays, the projection and the 2025 result. A preset converts each lens's shares into slider errors rounded to the nearest half point, so a preset reproduces its replay only to within a quarter of a point per party. The tool shows the central projection only, without the simulation, so that it can recalculate on every slider movement. The prediction tool does the same from shares the reader types in.

What Germany does not have

The British site has three further lenses built from pollsters' records and from subsets of pollsters. The German site does not. With a 21-day window there are only a handful of pollsters in the average (today 7), and most have one to six checked final polls with a lead error that is small by British standards. A "friendliest pollsters" lens would mostly reshuffle the Greens and the AfD by a point, and a "pollster's own record" lens would rest largely on figures transcribed from secondary sources (Rating German pollsters: accuracy, the lead and openness). We may add them; they are not there now.

Every reading through the same seat engine

ReadingCDU/CSUAfDSPDGreensLinkeBSWFDP
The polls as they stand152193991117400
Our projection (the polls, uncorrected)152193991117400
If the polls miss like 2002159189941186900
If the polls miss like 200510218710211476048
If the polls miss like 20091391878011075038
If the polls miss like 2013165197911047200
If the polls miss like 20171241968711668038
If the polls miss like 20211651911021076400
If the polls miss like 20251421921071008800
Live Seats under each reading of today's polls Computed from the database, polls to 17 November 2026.

The table is recomputed whenever a poll is added, so the numbers move. Three features follow from the structure of the misses and should survive most changes in the polls.

Most replays move the big parties by a few dozen seats, not hundreds. The largest single error in the record is the CDU/CSU's +6.3 points in 2005; every other party-election error is 3 points or less, and most are under 2. A replay of 2005 therefore takes the most off the CDU/CSU. Replays of 2002, 2013 and 2021, the three elections at which the CDU/CSU was understated, add to it.

The AfD barely moves. Its errors run from −1.6 (2017) to +0.4 (2021), and three replays do not adjust it at all. Whatever the replay, the AfD's seats change only by the rescale and by the knock-on effect of other parties entering or leaving the Bundestag. That is a statement about the record, not about the AfD's polls today: four elections of history do not rule out a bigger miss.

The FDP column jumps. The FDP has been understated by more than a point three times (2005, 2009 and 2017) and overstated or measured almost exactly four times (2002, 2013, 2021 and 2025). The next section explains why that turns into a column that is either zero or forty-odd.

How the 5% line makes some readings jump

The discontinuity

Under the seat law a party below 5% of the Zweitstimmen with fewer than three Wahlkreis winners gets nothing, and its votes are ignored when the 630 seats are shared. A party just over the line gets roughly its share of the votes that count:

\[ \text{seats}_p \approx 630 \times \frac{v_p}{\sum_{q \in Q} v_q}, \tag{4} \]

where \(Q\) is the set of qualifying parties. In 2025, 13.7% of Zweitstimmen went to parties that won no seats (BSW 4.98%, FDP 4.33%, Freie Wähler 1.55% and 2.87% for the rest). A party with exactly 5% would have had about \(630 \times 5 / 91.3 \approx 34\) seats. The seat count is a step function of the vote share with a step of thirty-odd seats at 5%, and nothing like it anywhere else in the range where German parties poll.

Why the replays hit the step

A replay adds \(-e_{p,t}\) to each party. For the FDP those amounts are −0.9 (2002), +2.8 (2005), +1.3 (2009), −0.6 (2013), +1.4 (2017), −0.2 (2021) and 0.0 (2025), before the small rescale. So, roughly:

  • if the FDP is polling between about 3.7% and 5%, the replays of 2005, 2009 and 2017 put it over the line and the other four leave it out;
  • if it is polling between about 2.2% and 3.7%, only the 2005 replay puts it in;
  • if it is polling at 5.9% or more, no replay takes it out (between 5% and 5.9%, the 2002 replay, and below 5.6% also the 2013 one, can).

In the first case the FDP column has three non-zero entries of similar size and four zeros, rather than a spread around a central value. The same happens, mirrored, to every other party. When the FDP enters, its thirty to fifty seats come from the others in proportion to their size, so the CDU/CSU, AfD, SPD, Greens and Linke all lose seats in exactly the replays where the FDP gains. A reader who looks only at the CDU/CSU column will see its lowest totals in the 2005 and 2017 replays, and will be seeing two effects at once: the replay takes points off the CDU/CSU directly, and it lets the FDP in.

The BSW sits in the same zone but has a record of one election, in which the polls understated it by 0.4 points. Only the 2025 replay moves it, and only by that much, so no replay puts it over 5% unless it is already polling within about 0.4 points of the line. The replays cannot express the most important uncertainty about the BSW, which is whether a new party's support is measured well at all. That uncertainty belongs to the simulation (Simulating German polling error: where the chances come from), which gives every party a random national error of at least one point, and to the reader's own lens.

The three-Wahlkreis route

A party below 5% still qualifies with three Wahlkreis winners. The rule was abolished in 2023 and restored by the Federal Constitutional Court on 30 July 2024 as an interim rule. It matters for Die Linke, whose candidates won six Wahlkreise in 2025, and would matter for the CSU if its national share ever fell below 5%. Because every replay recomputes the Wahlkreis winners, the engine applies this route automatically: a replay that pushed Die Linke under 5% would still give it seats if its projected Wahlkreis wins stayed at three or more. Neither the FDP nor the BSW won a Wahlkreis in 2025, so for them the 5% line is in practice the only route.

The coalitions jump with it

A party entering the Bundestag changes the majority line for everyone else, because it takes seats that would otherwise have gone to them. A coalition that clears 316 seats with five parties in the Bundestag can fall short with six. The coalition table below is built on the projection and the simulation; the lens table above shows the arithmetic under each replay.

CoalitionSeatsMajority
CDU/CSU + SPD2510%
CDU/CSU + Greens2630%
CDU/CSU + SPD + Greens36299%
SPD + Greens + Linke2841%
CDU/CSU + AfD34591%
Live Coalitions: seats and chance of a majority Computed from the database, polls to 17 November 2026.

In 2025 itself the BSW finished about 9,500 votes short of 5%. Had it got in, the CDU/CSU and SPD would not have had a majority (How we read German polls: the 5% line and the 2023 seat law). Replays that move a party across the line are the ones to read for coalition arithmetic.

Wahlkreise at risk

The measure

For each Wahlkreis \(i\), let \(w^{L}_{i}\) be the party that wins the most Erststimmen under reading \(L\) and \(w^{B}_{i}\) the winner under the projection. The site counts how many of the seven replays give a different winner:

\[ F_i = \sum_{t \in \{2002, 2005, 2009, 2013, 2017, 2021, 2025\}} \mathbf{1}\!\left[\, w^{t}_{i} \neq w^{B}_{i} \,\right]. \tag{5} \]

A Wahlkreis is "at risk" if \(F_i \geq 1\). Today 102 of the 299 are. The code is LensMath.Flips, run over the lenses in the "past" group only; the polls-as-they-stand lens and the reader's lens do not count.

As in Britain, this is a count of historical scenarios, not a probability. A Wahlkreis that changes under two of seven replays is not "two sevenths likely" to change hands. The simulation gives each Wahlkreis a chance for each party, and the seat pages show that chance next to the count. The replays are also not independent. The 2005 and 2017 misses had the same shape (the CDU/CSU overstated by three points or more, the FDP understated), and 2002, 2013 and 2021 had roughly the opposite one, so a Wahlkreis that flips under one of a pair will often flip under the other.

What kind of Wahlkreise flip

The pattern follows from the errors. The replays that take most off the CDU/CSU (2005 and 2017) flip the most Wahlkreise, most of them from the CDU/CSU to the AfD where the two are close on Erststimmen. The replays that add to the CDU/CSU (2002, 2013 and 2021) flip some the other way, many from the AfD to the CDU/CSU. Smaller numbers of changes involve the SPD, the Greens and Die Linke. The "At risk" tab described below shows, for every Wahlkreis, which replay changes it and to whom.

Why "at risk" means less than in Britain

In Britain a constituency that changes hands is a seat that changes party. In Germany it is not. The party totals are set entirely by the Zweitstimmen; the Wahlkreise decide only which of a party's candidates sit, and whether a winner is covered by the party's list share in that Land. A Wahlkreis flip from the CDU to the AfD does not, by itself, give the AfD a seat: the AfD's total is already fixed by its national Zweitstimmen, and the extra Wahlkreis winner takes one of those seats instead of a list candidate. The flip changes who represents the Wahlkreis. Under the 2023 law it can also leave the Wahlkreis with no directly elected member at all, if the new winner's party has more Wahlkreis winners in the Land than seats. On the projection, the number of projected winners who would not be seated is 22; in 2025 it was 23.

So the Wahlkreise-at-risk count answers the local question, whether this area's member could be from a different party, and not the national one. The national question, how many seats each party could lose, is answered by the lens table and by the simulation's ranges. The exception is the three-Wahlkreis rule, where a flip can change whether a party is in the Bundestag at all.

How the map and the seat pages use the readings

The home page has a chart headed "If we're wrong": for each party, a dot for each replay, a ring for the polls as they stand and a bar for the projection. Below it, a link gives the number of Wahlkreise that change hands under at least one past miss and leads to the list.

The map (/de/map) has a reading selector. In "winner" mode it colours every Wahlkreis by its projected winner under any of the nine readings; under the projection the colour is paler where the simulation is less sure. In "risk" mode it colours each Wahlkreis grey if no replay changes the winner, amber if one does and red if two or more do, and the tooltip names the years whose misses would change it. The other modes (seated or not, one party's chance, one party's projected Erststimmen share, the 2025 result) do not depend on the replays. The same modes work on the hex map, where every Wahlkreis has the same size, and on the real boundaries.

The Wahlkreis list (/de/seats) has an "At risk from polling error" tab. It lists every Wahlkreis with \(F_i \geq 1\), sorted by \(F_i\) and then by the projection's chance, with one column per replay showing "=" where the winner is unchanged and the alternative winner where it is not.

Each Wahlkreis page has a table "Under each reading of the polls" giving the winner under all nine readings, with "(different)" against those that differ from the projection, and a line such as "A different winner under 2 of 7 past misses".

The what-if and prediction tools show, for the reader's own reading, the seat totals, the coalitions, the number of winners not seated and a list of Wahlkreise whose winner differs from the projection, with the option to highlight them on the hex map.

Why the readings are shown side by side

The reasons are those set out for Britain in Lenses: reading the same British polls several ways, and two are stronger in Germany.

Averaging the replays is the correction that failed. The mean of the seven replay vectors, party by party, is the mean past error. Taking off half of it is the correction whose leave-one-out test is shown above, and it scored 1.30 points per party against 1.21 for the polls as they stand. Taking off all of it, which is what averaging the replays amounts to, did worse still: run the same way on our database, it missed by 1.39 points per party against 1.21, and was further out at five of the seven elections. Averaging the scenarios would bring back, at full strength, a correction that does not work.

The threshold makes averages of seat totals meaningless. With a step of about 34 seats at 5%, the average of the FDP's seats across the seven replays is a number of seats the FDP cannot win: either it is in with thirty or more, or it is out with none. The same holds, less visibly, for every other party, since their seats depend on whether the FDP or the BSW is in. The seat total under the average error is not the average of the seat totals, and near a threshold the two are far apart.

The spread is the information. A reader who sees four replays with the FDP out and three with it in has learned where the forecast is fragile: not in how the big parties divide the seats, but in whether a small party crosses a line. That is what the simulation's "gets in" column puts as a probability (Simulating German polling error: where the chances come from), and what the replays show as named, reproducible cases.

What a replay cannot show

Misses that have not happened. The replays can only repeat the seven errors in the record. The largest AfD error in them is 1.6 points and the only BSW error is 0.4. A miss of a new shape, for example the AfD overstated by three points or the BSW understated by two, is outside the catalogue. The reader's own lens is the only way to see one.

Misses that differ by region. A replay moves each party by the same number of points nationally, and proportional swing then spreads that change across the Länder in proportion to the party's 2025 strength there. A miss that fell mainly in the East, or mainly in the cities, cannot be expressed. Regional polls, which would be the evidence for such a miss, are too few to use (What we don't know yet: limitations of the German forecast).

Local effects. The Wahlkreis projection has no candidates in it. A popular incumbent, a retirement or a tactical Erststimme campaign can decide a close Wahlkreis whatever the national polls do. Many at-risk Wahlkreise are close precisely because the 2025 result there was close, and local factors are most likely to decide those.

The date. The replays apply past final-poll misses to today's polls, as if the election were held now. Before a real election the polls will move by much more than any of these misses: polls this far out have missed by about 5.2 points per party by polling day (From today to polling day in Germany). The replays describe the error that remains when the polls are measured in the last week, not the drift before then.

Seven elections. Every replay rests on one election's final polls, five to thirteen of them. The replays start in 2002 because the model uses no election before 2000 (the cutoff is an admin setting). The 2005–2025 finals come from pollsters' and outlets' releases; the 2002 finals, like the older ones in the long view back to 1953 (Measuring German polling error, 2005–2025), come only from secondary sources (grade C).

References

Bundesverfassungsgericht, judgment of 30 July 2024, 2 BvF 1/23.

Bundeswahlgesetz (BWG), as amended by the Act of 8 June 2023 (BGBl. 2023 I Nr. 147).

Jennings, W. and Wlezien, C. (2018). Election polling errors across time and space. Nature Human Behaviour, 2(4), 276–283.

How to cite

Polling Forecast (2026). If we're wrong: replaying past Bundestag polling misses on today's polls. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/germany-lenses. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like