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

Measuring German polling error, 2005–2025

Abstract. At the six Bundestag elections from 2005 to 2025, the average of the final polls missed each party by about one point, except in 2005, when every pollster had the CDU/CSU about six points too high. Since then no party has been missed the same way every time, and the AfD, understated in 2013 and 2017, has been polled almost exactly at the last two elections. We define the error, describe the 52 final polls and how they were sourced and graded, and report the gap between the two largest parties, each pollster's record, earlier elections back to 1953 and the state elections held since 2025.

Polling Forecast · updated 9 November 2026 · 15 min read · Germany, Bundestag, polling error, final polls, CDU/CSU, AfD, gap error, house effects, source grades, long view, state elections, Landtag

Summary

This paper sets out how we measure polling error in Germany: which polls count as final polls, how their sources are graded, how the error is defined, and what the record shows. At the six Bundestag elections from 2005 to 2025 the average of the final polls missed each party by about one point, except in 2005, when it missed by more than two. The worst miss was the CDU/CSU in 2005, overstated by 6.3 points on average, with every pollster between 5.8 and 6.8 points too high. Since then no party has been missed the same way every time. The Union has been overstated four times and understated twice; the SPD, the Greens and the Linke have each been both overstated and understated by more than a point; the AfD was understated in 2013 and 2017 and has been polled almost exactly since.

The record has three uses on the site. Its size sets the width of our ranges (Simulating German polling error: where the chances come from). Its lack of a steady direction is the reason we do not correct German polls (How we read German polls: the 5% line and the 2023 seat law). And its pollster-by-pollster detail feeds the ratings (Rating German pollsters: accuracy, the lead and openness). Two further bodies of evidence sit beside it: final polls at earlier Bundestag elections back to 1953, where we could source them, and the six state elections held since the 2025 federal election, which serve as a check between federal elections.

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.
  • When the polls missed (2 November): Six points lost in the last week: Merkel's vanishing lead and Germany's 2005 polling miss.
  • The track record across countries (28 September). Final polls at 133 elections in nine countries, measured the same way, are at /accuracy/elections.

What is measured, and against what

The quantity that decides the Bundestag is each party's national share of the valid Zweitstimmen, the list vote. That is what German voting-intention polls ask about (the "Sonntagsfrage": whom would you vote for if the Bundestag election were next Sunday?), and it is what we measure against. Results come from the Federal Returning Officer (the Bundeswahlleiterin) and are the final official figures. For 2021 we use the original result of 26 September 2021, which is what the polls were trying to predict, not the version revised after the partial repeat election in Berlin in February 2024.

Unlike in Britain, where error is measured in England to keep the SNP and Plaid Cymru out of the comparison (Measuring British polling error in England, 2015–2024), Germany needs no such step. The national figure is the one that decides the seats, and every party is polled on the same national basis. The one complication is the CDU and CSU. They are separate parties, but the CSU stands only in Bavaria, and almost every poll reports the two together as the Union. We measure the Union as one party.

The parties measured at each election are those that appear in the result and in at least half of that election's final polls: the Union, the SPD, the FDP, the Linke and the Greens throughout, the AfD from 2013 and the BSW in 2025. Everything else is "others" and is not scored.

Definitions

Let \(v_{p,t}\) be party \(p\)'s share of the valid Zweitstimmen at election \(t\), and let \(y_{j,p,t}\) be its share in final poll \(j\), one of the \(n_t\) final polls for that election. The error of a single poll is

\[ e_{j,p,t} = y_{j,p,t} - v_{p,t}. \tag{1} \]

The sign convention is the same everywhere on the site: error = poll minus result. A positive error means the polls overstated the party.

The industry error at election \(t\) is the error of the simple average of the final polls,

\[ e_{p,t} = \bar{y}_{p,t} - v_{p,t}, \qquad \bar{y}_{p,t} = \frac{1}{n_{p,t}}\sum_{j} y_{j,p,t}, \tag{2} \]

where \(n_{p,t}\) is the number of final polls reporting the party. Each pollster counts once, however large its sample.

To compare elections we use the average miss per party, the mean absolute industry error over the \(P_t\) parties measured at that election:

\[ \text{MAE}_t = \frac{1}{P_t}\sum_{p} \left| e_{p,t} \right|. \tag{3} \]

This is the "average miss per party" quoted throughout the site: the miss of the average poll, which is never larger than the average miss of the individual polls.

For each party we summarise its errors across elections by their mean, which measures direction, and their root mean square (RMS), which measures size:

\[ \bar{e}_p = \frac{1}{T_p}\sum_{t} e_{p,t}, \qquad \text{RMS}_p = \sqrt{\frac{1}{T_p}\sum_{t} e_{p,t}^{2}}, \qquad \text{RMS}_p^2 = \bar{e}_p^{\,2} + s_p^2, \tag{4} \]

where \(s_p^2\) is the variance of the errors around their mean. The identity on the right splits the size of a party's misses into a steady part (bias) and a part that changes from one election to the next.

Finally, the gap between the two largest parties decides who leads the next government, so we also score it. If \(a\) and \(b\) are the two parties that finished first and second, the gap error is

\[ e^{\text{gap}}_t = (\bar{y}_{a,t} - \bar{y}_{b,t}) - (v_{a,t} - v_{b,t}) = e_{a,t} - e_{b,t}. \tag{5} \]

A positive gap error means the polls put the winner too far ahead.

The evidence: final polls and how they were chosen

What counts as a final poll

A final poll is the last poll by a given pollster whose fieldwork ended before polling day, provided it was published before voting began and fell within the 31 days before the election. One poll per pollster per election is kept. Election-day forecasts and exit polls are excluded, and so are MRP models (YouGov published one in 2025) and releases marked on Wikipedia as intermediate results.

That gives 53 pollster-elections from 16 pollsters. One of them cannot be scored: Forsa published only ranges before the 2005 election, so its 2005 final poll has no shares and Forsa is scored from 2009. The benchmark is therefore 52 final polls.

The number of final polls grew from 5 in 2005 to 6 in 2009, 7 in 2013, 9 in 2017, 12 in 2021 and 13 in 2025. Allensbach, Forschungsgruppe Wahlen, GMS and infratest dimap have a final poll at all six elections, Verian (published as Emnid until 2017 and as Kantar in 2021) at five, Forsa at five that can be scored, and INSA at four; the rest have one to three.

Two gaps are known: Emnid's 2009 final poll could not be found in any usable source, and a 2005 row labelled "TNS Forschung", with figures identical to infratest dimap's poll of that week, is excluded because its identity could not be verified.

Grading the source

Every figure carries a source grade, on the same scale as in Britain. Grade A is the pollster's own published release. Grade C is a reputable secondary transcription, here Wikipedia's tables. Three pollsters have archives we could read directly: Forschungsgruppe Wahlen's Politbarometer archive (from 2004), infratest dimap's Sonntagsfrage series (from 2000) and Allensbach's own table (from July 2010). Their final polls from those dates are grade A, which gives 16 of the 53 pollster-elections. The rest, including every INSA, Forsa, Verian, GMS and YouGov final poll, are grade C.

Grade C is a real limitation, and we do not hide it. Almost every Wikipedia poll row cites wahlrecht.de, a long-running compilation of German polls. We did not copy from that site, but we used it as a check once our final-poll table was finished: 251 of 252 party figures agreed. The one exception, Emnid's FDP figure in 2013 (5 in one source, 6 in the other), is left blank because no primary source settles it. The spot-checks against pollsters' own releases that could be made, for example YouGov's published deviations from the 2025 result, matched every figure. The German data: results, boundaries, polls and licences describes the sources and checks in full.

Unlike the British benchmark, the German one keeps grade C figures: we need only the national headline, not an England figure extracted from a poll's tables, and a transcribed headline that agrees with an independent compilation is sound enough to use.

A worked example: 2025

The 2025 election, held early on 23 February, had the most final polls. Their figures and the result were:

Pollster Fieldwork end CDU/CSU AfD SPD Greens Linke BSW FDP
YouGov 20 Feb 29 20 16 13 8 5 4
Ipsos 21 Feb 30 21 16 12 7 4.5 4.5
INSA 22 Feb 29.5 21 15 12.5 7.5 5 4.5
Forschungsgruppe Wahlen 20 Feb 28 21 16 14 8 4.5 4.5
Wahlkreisprognose 21 Feb 29.5 20.5 14.5 12 8 4.5 4.5
Forsa 20 Feb 29 21 15 13 8 3 5
Cluster 17 18 Feb 30 21 15 13 7 4 4
pollytix 12 Feb 29 19 17 14 6 5 4
Allensbach 20 Feb 32 20 14.5 12 7.5 4.5 4.5
GMS 19 Feb 31 20 15 13 6 4 4
Verian 28 Jan 30 20 15 14 4 5 4
infratest dimap 12 Feb 32 21 14 14 6 4.5 4
Democracy Institute 30 Jan 27 25 15 13 5 6 5
Average 29.69 20.81 15.23 13.04 6.77 4.58 4.35
Result 28.52 20.80 16.41 11.61 8.78 4.98 4.33
Error +1.17 0.00 −1.18 +1.43 −2.01 −0.40 +0.02

Table: Final polls and result, Bundestag 2025, Zweitstimmen, percentage points. Rows ordered by each poll's own average miss.

The Union's average is \(386/13 = 29.69\) against a result of 28.52, an error of \(+1.17\) by equation (2). The average miss per party is \((1.17 + 0.00 + 1.18 + 1.43 + 2.01 + 0.40 + 0.02)/7 = 0.89\) points by equation (3). The gap between the Union and the AfD was 7.7 points; the polls had it at 8.9, a gap error of \(+1.2\).

Three features of 2025 recur in the record. The average was closer than most of its members: it missed by 0.89 points per party, the individual polls by between 0.60 (YouGov) and 2.00 (Democracy Institute). The largest miss was shared: every poll understated the Linke, by between 0.8 and 4.8 points, the signature of an error common to the industry. The AfD, by contrast, was right on average, with the polls scattered from 1.8 points too low to 4.2 too high. And two polls were stale, ending on 28 and 30 January; they are kept because they were those pollsters' last word, and excluding them would change the rule after seeing the result.

The BSW row shows how a small error decides seats. The average put the party at 4.58%, five of the thirteen polls had it at 5% or above, and it finished about 9,500 votes short of the line.

The record, 2005–2025

The table below applies the same calculation to all six elections. It is computed from the database each time the page is built.

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.

Four points stand out.

2005 was different. The average miss per party was 2.33 points, more than twice that of any later election, and almost all of it was one party. Every final poll had the Union between 5.8 and 6.8 points too high and the FDP between 1.8 and 3.3 points too low. Since 2005 the average miss per party has stayed between 0.80 (2021) and 1.48 (2017).

The Union's misses are the largest, and they change sign. It was overstated in 2005, 2009, 2017 and 2025 and understated in 2013 and 2021. In 2013 every final poll had it too low, by between 1.5 and 3.5 points; in 2017 every one had it too high, by between 1.1 and 4.1. Both were industry misses in opposite directions, four years apart.

No party has a steady lean. The FDP was understated three times (2005, 2009 and 2017) and within 0.6 points otherwise; its mean is negative, but at the last two elections it was polled almost exactly. The Greens were overstated in 2013, 2021 and 2025 and understated in 2005, 2009 and 2017. The Linke was overstated by 1.5 points in 2021 and understated by 2.0 in 2025.

The AfD has not been understated since 2017. A common view is that German polls understate the AfD because some of its voters do not say so. That was true in 2013 and 2017, by 0.8 and 1.6 points. In 2021 the polls overstated it by 0.4, and in 2025 they had it exactly. The raw figures tell part of the story: Forschungsgruppe Wahlen's unadjusted "Politische Stimmung" had the AfD at 14% in its release ten days before the 2025 election, against 20.8% in the result. The pollsters' adjustments have been correcting for exactly this, and doing so well.

Using equation (4), the summary for each party is:

Party Elections Mean error RMS error Share of RMS² that is bias
CDU/CSU 6 +1.29 3.12 17%
FDP 6 −0.79 1.41 31%
SPD 6 +0.36 1.29 8%
Linke 6 −0.19 1.09 3%
Greens 6 +0.10 1.03 1%
AfD 4 −0.52 0.94 30%
BSW 1 −0.40 0.40 –

Table: Summary of industry errors, 2005–2025, points. The last column is \(\bar{e}_p^2 / \text{RMS}_p^2\).

For the SPD, the Greens and the Linke almost all of the error is variation from one election to the next, with little or no steady direction. For the Union, the FDP and the AfD about a sixth to a third is bias, but for the Union and the FDP that bias comes mostly from 2005. Without 2005, the Union's mean error is +0.29 and its RMS 1.91; the FDP's mean is −0.38. This is the pattern that makes correction fail (see below): the steady part of the error is small, and much of it belongs to one election twenty years ago.

Shared and pollster-specific error

As in Measuring British polling error in England, 2015–2024, the average squared poll error at an election splits exactly into the square of the industry error and the spread of the polls around their average:

\[ \frac{1}{n_t}\sum_{j} e_{j,p,t}^{2} = e_{p,t}^{2} + \frac{1}{n_t}\sum_j \left(y_{j,p,t} - \bar{y}_{p,t}\right)^{2}. \tag{6} \]

The share of the first term is the share of the miss common to the industry. For the largest misses it is high: 100% for the Union in 2005 (every poll between +5.8 and +6.8), 89% for the Union in 2013, 92% for the Union in 2017, 86% for the SPD in 2009 and 73% for the Linke in 2025. No choice of pollster would have avoided these, and averaging more polls does nothing to reduce them. Elsewhere the shared part can be small: 0% for the AfD and the FDP in 2025 and 17% for the SPD in 2021, where the industry was right on average and the pollsters scattered around it.

The gap between the two largest parties

The party that finishes first normally leads the government, so the gap between the first two matters more than any single share.

Election Top two Real gap Polls' gap error Worst poll
2005 CDU/CSU, SPD 0.9 +7.2 +8.1
2009 CDU/CSU, SPD 10.8 −1.3 −3.8
2013 CDU/CSU, SPD 15.8 −3.0 −5.8
2017 CDU/CSU, SPD 12.4 +1.7 +4.6
2021 SPD, CDU/CSU 1.7 +1.4 +3.3
2025 CDU/CSU, AfD 7.7 +1.2 −5.7

Table: Gap between the two largest parties, points. The polls' gap error is equation (5) on the average; the worst poll is the single final poll with the largest absolute gap error.

2005 is the one election at which the polls called the wrong kind of result. They had the Union about eight points clear of the SPD; it finished less than a point ahead, and a grand coalition followed. At every later election the average got the order of the top two right. In 2021 the SPD's lead was 1.7 points and the polls had it at about 3.1: right in direction, but within a normal miss. Single polls could mislead more: in 2025 Democracy Institute had the Union only two points ahead of the AfD.

Pollster by pollster

Each pollster's record is the average miss per party of its own final polls, over the elections it covers.

Pollster Elections Average miss per party
Civey 1 0.84
Wahlkreisprognose 2 0.85
Ipsos 2 0.92
INSA 4 1.03
Forsa 5 1.08
Cluster 17 1 1.08
pollytix 1 1.20
Allensbach 6 1.20
Forschungsgruppe Wahlen 6 1.23
YouGov 3 1.39
infratest dimap 6 1.45
Trend Research 2 1.46
INFO GmbH 1 1.51
Verian 5 1.57
GMS 6 1.58
Democracy Institute 1 2.00

Table: Average miss per party of each pollster's final polls, 2005–2025, points.

Two cautions apply. The first is that the records cover different elections. The five pollsters with a 2005 final poll (Allensbach, Forschungsgruppe Wahlen, GMS, infratest dimap and Verian) carry it in their records, when every pollster missed by more than two points per party; a pollster that started in 2021 never faced it. The second is that the differences are small next to the differences between elections. The best long record, INSA's 1.03 over four elections, is about a fifth of a point better than Forschungsgruppe Wahlen's 1.23 over six, while the industry's own miss moved between 0.80 and 1.48 from one election to the next after 2005.

The pollster ratings (Rating German pollsters: accuracy, the lead and openness and the ratings page) use these records, pulled towards the record of a poor pollster so that one good election cannot earn a high score, alongside the gap error and how much each pollster publishes.

Why the record is not used to correct

The obvious use of the table is to subtract each party's average miss from today's polls. Our rule, the same in every country, takes off half the average miss and keeps the correction only if it would have improved past forecasts, each election predicted without knowing its own result. In Germany it would not: correcting was further out at 5 of six elections and worse on average, because an error that changes sign cannot be predicted from its own mean, and the part that does not change sign is largely 2005. How we read German polls: the 5% line and the 2023 seat law reports the test. The record is used instead to set the size of the uncertainty: in each simulated election, every party's share is moved by a random error with a standard deviation equal to its RMS error from the summary table, but never less than one point (Simulating German polling error: where the chances come from).

The long view

Six elections are few. To see whether they are typical, we collected every earlier Bundestag final poll we could source.

ElectionFinal pollsAverage miss per partyBiggest miss
195332.09Others +6.1
197211.21FDP −2.3
197610.78SPD −1.6
199850.82CDU/CSU +2.6
200250.92CDU/CSU −1.5
200552.33CDU/CSU +6.3
200961.10SPD +2.3
201370.94CDU/CSU −2.0
201791.48CDU/CSU +3.0
2021120.80CDU/CSU −1.7
2025130.89Linke −2.0
Live Final polls against the result, every Bundestag election we could source Computed from the database, polls to 17 November 2026.

The earlier record is patchy. No pollster's own archive reaches back before 2000, so every figure before 2005 is grade C, keyed by hand from Wikipedia's tables, which cite the original studies for 1953 and 1972. We found no usable final polls at all for 1957, 1961, 1965, 1969, 1980, 1983, 1987, 1990 or 1994, and 1972 and 1976 have one pollster each. The 1953 row needs a further caution: two of its three sources combined the FDP with two smaller parties, so their "others" figure is a derived sum, and the large miss on "others" that year is partly an artefact of that.

With those limits, the long view supports two conclusions. The elections of 1998 and 2002, with five pollsters each, missed by less than a point per party, much like the elections since 2009, which makes 2005 look like the exception rather than the norm. And the largest miss at each election has fallen on different parties in different directions: the Union too high in 1998, too low in 2002, far too high in 2005. The long view adds no evidence of a steady lean.

State elections since 2025

Between federal elections the nearest thing to a test is a Landtag (state parliament) election: the same pollsters, mostly the same parties, and a real result. Six have been held since the 2025 Bundestag election. For each we take the last poll by every pollster ending within the 31 days before polling day and compare it with the list-vote result.

ElectionPollstersAverage miss per partyAfD: polls minus result
Hamburg 202531.74+2.5
Baden-Württemberg 202632.31−0.1
Rheinland-Pfalz 202631.31−0.1
Sachsen-Anhalt 202652.22−1.4
Mecklenburg-Vorpommern 202651.84−1.4
Berlin 202651.57+2.1
Live State elections since 2025: final polls against the result Computed from the database, polls to 17 November 2026.

The average miss per party, 1.85 points across 177 poll figures, is larger than at recent Bundestag elections, as one would expect from smaller samples of smaller electorates. The direction varies from state to state. The Linke was understated by about 4.4 points in Berlin but overstated by about 3.5 in Sachsen-Anhalt and 3.7 in Mecklenburg-Vorpommern. The CDU was overstated by 4.8 points in Sachsen-Anhalt and understated by 2 to 2.5 in Hamburg, Baden-Württemberg and Rheinland-Pfalz. The AfD was polled within 2.5 points everywhere: too high in Hamburg and Berlin, too low in Sachsen-Anhalt and Mecklenburg-Vorpommern, and almost exact in the two western states that voted in March 2026. Averaged over all 24 pollster-elections, its error is +0.1.

We use the state elections as a check, not as an input. A consistent miss for one party across several states, especially an understated AfD, would be an early warning that the Bundestag polls might share it. So far there is none. Two of the six results are still preliminary: Berlin's is the returning officer's provisional result, and Mecklenburg-Vorpommern's is taken from a secondary source because the official site could not be reached over a verified connection. The German data: results, boundaries, polls and licences records both, and the state elections page shows every poll.

Limits

Six elections. Every summary statistic here rests on six numbers per party, four for the AfD and one for the BSW. A seventh election like 2005 would raise the Union's RMS sharply; one like 2021 would lower it. The absence of a steady direction is itself a finding from a small sample.

Secondary sources. Two thirds of the final polls are transcriptions. They agree with an independent compilation in 251 of 252 figures, but agreement between two secondary sources is not the same as a pollster's own table.

Headlines only. German pollsters publish their adjusted figures, and only one publishes its raw answers. We cannot look inside the other pollsters' polls as we do in Britain, so we cannot tell whether their samples lean one way and their adjustments another.

Industry change. Pollsters change methods, owners and names (Emnid is now Verian). A pollster's record describes what it did then, not necessarily what it does now.

Final polls only. These errors say nothing directly about a poll taken years before an election, when real opinion has time to move; that is the subject of From today to polling day in Germany.

References

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

Shirani-Mehr, H., Rothschild, D., Goel, S. and Gelman, A. (2018). Disentangling bias and variance in election polls. Journal of the American Statistical Association, 113(522), 607–614.

Die Bundeswahlleiterin, Wiesbaden. Final results of the Bundestag elections, 2005–2025 (open data). Datenlizenz Deutschland – Namensnennung 2.0.

How to cite

Polling Forecast (2026). Measuring German polling error, 2005–2025. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/germany-polling-error. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like