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

Rating German pollsters: accuracy, the lead and openness

Abstract. We score each German pollster out of 100 on its final Bundestag polls since 2005, its error on the gap between the top two parties, and what it publishes. Since 13 October 2026 accuracy and the lead are ranked against all German final polls and a short record is pulled towards the typical one: Forschungsgruppe Wahlen and YouGov grade B, most others C.

Polling Forecast · updated 9 November 2026 · 16 min read · Germany, pollster ratings, Forschungsgruppe Wahlen, Projektion, Politische Stimmung, raw figures, shrinkage, openness, ADM, state elections

Summary

We rate each German pollster out of 100 on the same scale as the British ratings (Rating the British pollsters: accuracy, the Labour–Conservative lead and openness): 50 points for the accuracy of its final Bundestag polls since 2005, 20 for its error on the gap between the two leading parties, and 30 for how much of its working it publishes. The record is pulled towards a deliberately poor pollster, so a pollster with one good election cannot score well by luck. The rating is shown for readers. It does not weight the polls: our German average gives every pollster's latest poll the same weight.

Three things differ from Britain. German final polls have been more accurate, so the "poor pollster" the ratings pull towards is a good one by British standards, and German grades run higher than British ones for the same evidence. The openness items are specific to Germany, where almost no pollster publishes full tables and only Forschungsgruppe Wahlen publishes what respondents said before its own adjustments. And that one raw series, FGW's "Politische Stimmung", was further from the result than FGW's own published "Projektion" at five of six elections, which is why, unlike in Britain, we do not start from raw figures (Why we start from British pollsters' raw answers rather than the published headline). Beside the rating we show each pollster's lean against today's average and its record at the state elections held since 2025; neither is scored.

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.

Pollster grades were rescaled on 13 October 2026. The first scale scored accuracy between fixed end points (a miss of 0.75 points or less earned full marks, 2.75 or more nothing) and pulled every short record towards a poor pollster, one that three in four final polls beat. Almost every pollster ended up with a D or an E, which told readers little. The rating now works like this:

  • Accuracy (50 points) is ranked against the country's own checked final polls. If a share \(r\) of them missed by less than the pollster's average miss, it scores \(50\,(1-r)^{0.75}\): the typical (median) final poll earns about 30 and one in the best tenth about 46. Ranking within the country means a pollster is not marked down because its country is harder to poll.
  • The lead (20 points) is scored the same way on the miss on the gap between the two leaders, \(20\,(1-r)^{0.75}\).
  • A short record is pulled towards the typical (median) final poll, as if the pollster had two more elections there, not towards a poor one.
  • Openness (30 points) is scored out of the items checked so far; an item not yet checked counts neither for nor against.
  • Every pollster with a final poll on record gets a grade. One whose publications we have not reviewed is graded on its record alone (accuracy and lead scaled to 100) and marked "record only".

The grade boundaries are unchanged: A from 75, B 60–74, C 45–59, D 30–44, E below 30. Across the active pollsters of all ten countries the grades went from mostly D and E to A 8, B 68, C 76, D 31 and E 5. Where the sections below give the old end points, the poor-pollster prior or grades from before 13 October, they describe the first version of the scale; the live tables show today's.

In Germany it gives B to Forschungsgruppe Wahlen (62) and YouGov (61); C to Verian, Forsa, Ipsos, INSA and Infratest dimap; D to Allensbach and pollytix; and E to GMS. These are the figures of 9 November 2026; the table on the pollsters page is current.

What the rating is for

The rating answers the same narrow question as in Britain: on the evidence we can check, how much should a reader trust this pollster's published figure, and how easily can they see how it was made? It does not say which pollster is right today, and it does not feed the forecast. The German projection is the plain average of each pollster's latest poll within 21 days of the newest, one poll per pollster, equally weighted (How we read German polls: the 5% line and the 2023 seat law). A pollster that publishes twice a week, as INSA does, counts no more than one that publishes monthly.

We do not weight by the rating, for the same reason we do not correct German polls: the evidence is six elections, most pollsters have fewer, and their records move from one election to the next. A weighting that could not be shown to help out of sample would be a hidden assumption.

The evidence: 52 final Bundestag polls

The accuracy parts of the rating use every final poll in our record for the Bundestag elections of 2005, 2009, 2013, 2017, 2021 and 2025. A final poll is a pollster's last published poll before polling day. There are 53 in the database and 52 are usable: Forsa published its 2005 final poll only as ranges, so it has no shares. Sixteen pollsters have at least one.

Election Final polls Pollsters
2005 5 Allensbach, FGW, GMS, Infratest dimap, Verian (then Emnid)
2009 6 Allensbach, FGW, Forsa, GMS, INFO, Infratest dimap
2013 7 Allensbach, FGW, Forsa, GMS, Infratest dimap, INSA, Verian
2017 9 as 2013, plus Trend Research and YouGov
2021 12 as 2017, plus Civey, Ipsos and Wahlkreisprognose
2025 13 Allensbach, Cluster 17, Democracy Institute, FGW, Forsa, GMS, Infratest dimap, INSA, Ipsos, pollytix, Verian, Wahlkreisprognose, YouGov

Table: Final polls used in the ratings. Emnid's 2009 final poll could not be sourced.

The sourcing is uneven, and the rating inherits that. FGW's and Infratest dimap's final polls come from their own archives at every election, and Allensbach's from its own table from 2013; the rest come from Wikipedia's poll tables and are graded C (The German data: results, boundaries, polls and licences). A check against the wahlrecht.de final-poll tables (used as a check only, nothing copied) found 251 of 252 party figures in agreement. The one difference, Emnid's 2013 FDP figure, is left blank.

For pollster \(j\) at election \(t\), write \(e_{j,t,p} = \hat{v}_{j,t,p} - r_{t,p}\) for its share for party \(p\) minus the party's Zweitstimmen share. The parties are those measured at that election: CDU/CSU, SPD, Greens, FDP and Linke throughout, the AfD from 2013 and the BSW in 2025. The miss is the mean absolute error over the parties the poll reported,

\[ m_{j,t} = \frac{1}{|P_{j,t}|} \sum_{p \in P_{j,t}} \left| e_{j,t,p} \right|, \tag{1} \]

and the lead error is the error on the gap between the two parties that actually finished first and second,

\[ \ell_{j,t} = \left(\hat{v}_{j,t,a} - \hat{v}_{j,t,b}\right) - \left(r_{t,a} - r_{t,b}\right), \tag{2} \]

where \(a\) and \(b\) are the first and second parties in the result: the CDU/CSU and SPD in 2005, 2009, 2013 and 2017, the SPD and CDU/CSU in 2021, and the CDU/CSU and AfD in 2025. In Britain the lead is always Labour against the Conservatives. In Germany it follows the result, because the pair that decides who leads the government has changed.

Across the 52 polls the average miss is 1.30 points per party and the average absolute lead error 2.46 points. The closest single final poll was YouGov's in 2025 (0.60 per party); the furthest were GMS's and Emnid's in 2005 (2.55 each), in the election that dominates the German record (Measuring German polling error, 2005–2025).

Scoring accuracy and the lead

The scales are the British ones, unchanged. With \(\tilde{m}_j\) and \(\tilde{\ell}_j\) the rated miss and lead error defined in the next section,

\[ A_j = 50 \cdot \operatorname{clip}\!\left( \frac{2.75 - \tilde{m}_j}{2.00},\ 0,\ 1 \right), \qquad L_j = 20 \cdot \operatorname{clip}\!\left( \frac{9 - \tilde{\ell}_j}{8},\ 0,\ 1 \right). \tag{3} \]

A rated miss of 0.75 points or better earns the full 50, 2.75 or worse earns nothing; a rated lead error of 1 point or better earns the full 20, 9 or worse nothing. We kept the British end points on purpose, so that a German and a British pollster with the same record would get the same accuracy points. The next section shows that this does not make their grades comparable.

Shrinkage towards a poor pollster

Each pollster's record is pulled towards a prior as if it had two more elections at the prior's level (Efron and Morris, 1975):

\[ \tilde{m}_j = \frac{n_j \bar{m}_j + 2\, m^{*}}{n_j + 2}, \qquad \tilde{\ell}_j = \frac{n_j \bar{\ell}_j + 2\, \ell^{*}}{n_j + 2}, \tag{4} \]

where \(n_j\) is the number of the pollster's final polls with a usable miss and \(\bar{m}_j\), \(\bar{\ell}_j\) are its averages of \(m_{j,t}\) and \(|\ell_{j,t}|\). As in Britain, the prior is a poor pollster: \(m^{*}\) is the 75th percentile of the miss across all 52 German final polls and \(\ell^{*}\) the 75th percentile of the absolute lead error, by linear interpolation. The code is the same for every country (WorldPollsterRatings), with each country's own final polls.

With the current record the German prior is

\[ m^{*} \approx 1.48, \qquad \ell^{*} \approx 3.29 . \]

The British values are about 2.50 and 8.3. A poor German final poll misses by less than an average British one, and its lead error is less than half a British poor poll's. Plugged into the same scales, the German prior is worth far more points:

Checked elections Rated miss Accuracy points Rated lead error Lead points Best possible total (full openness) With no openness points
0 1.48 31.9 3.29 14.3 76 (A) 46 (C)
1 0.98 44.2 2.19 17.0 91 (A) 61 (B)
2 0.74 50.0 1.64 18.4 98 (A) 68 (B)
3 0.59 50.0 1.32 19.2 99 (A) 69 (B)

Table: The ceiling for a hypothetical pollster with no error at any checked election, using the current German prior.

This is the most important caveat about the German ratings, and we would rather state it than hide it. In Britain an unproven pollster cannot score above a D however open it is. In Germany the same rule puts an unproven pollster at 46, a C, before any openness points. A pollster that matched the German average at all six elections (a miss of 1.30 and a lead error of 2.46) would earn about 51 record points and reach a B with ten openness points. The British scale was set to be punitive against British errors; against German errors it is generous.

We have kept it for now, for two reasons. The German polls really have been more accurate, and a scale that gave the same grades in both countries would have to pretend otherwise. And the relative order, which is what a reader comparing German pollsters needs, does not depend on the end points. But German and British grades should not be compared: a German B means "better than most German final polls", not "as good as a British B". A German scale with its own end points is one of the open items in What we don't know yet: limitations of the German forecast.

A worked example

Forschungsgruppe Wahlen has a final poll at all six elections. Its misses were 2.15 (2005), 1.46 (2009), 0.81 (2013), 1.35 (2017), 0.91 (2021) and 0.71 (2025), an average of 1.23. Its lead errors were +6.1, +0.2, −2.8, +2.1, +0.3 and −0.7, an average absolute error of 2.04. Then

\[ \tilde{m} = \frac{6 \times 1.23 + 2 \times 1.48}{8} \approx 1.29, \qquad \tilde{\ell} = \frac{6 \times 2.04 + 2 \times 3.29}{8} \approx 2.35, \]

so \(A = 50 \times (2.75 - 1.29)/2 \approx 36\) and \(L = 20 \times (9 - 2.35)/8 \approx 17\). With six elections the prior still carries a quarter of the weight, but because FGW's own record is close to the prior here, shrinkage moves it little. Its openness points (below) are about 21, for a total of 74, a B.

Openness

The 30 openness points reward publishing enough that a figure can be checked. The British items (raw stage, turnout step, past-vote table, likelihood to vote, England figure) mostly do not exist in German publications, so the German list is different:

Item Points Why it counts
Raw or unprojected figures 8 Shows what respondents said before the pollster's own adjustments
Full tables published 8 Sub-groups and weighting can be checked
Method note 4 How the sample was drawn and weighted
Sample size 3 The basics for judging a poll
Undecided share 3 Shows how firm the figures are
Fieldwork dates 2 When the poll was taken
Member of the industry body (ADM) 2 Signs up to its standards

Table: German openness items.

Each item is Yes, Partial or No, scoring the full points, a third, or nothing, exactly as in Britain:

\[ O_j = \sum_{i} w_i\, s_{j,i}, \qquad s_{j,i} \in \left\{1, \tfrac{1}{3}, 0\right\}. \tag{5} \]

The review was done on 26 September 2026 from each pollster's own Sonntagsfrage pages and releases, with membership taken from the member list of the ADM (Arbeitskreis Deutscher Markt- und Sozialforschungsinstitute). Other memberships were not checked. The notes against each judgement are shown on the pollster's page. What we found, in brief:

  • Forschungsgruppe Wahlen is the only pollster that publishes a raw figure: its "Politische Stimmung", weighted for demographics but before its political projection, next to the headline "Projektion" in most regular releases since 2011. Each release gives the sample size, fieldwork dates and mode. There are no cross-tabulations, and the undecided share appears only in some releases. FGW is not on the ADM list.
  • YouGov publishes the fullest tables: a monthly PDF with cross-tabulations by 2025 vote, age, sex, education and East/West, with unweighted bases per column. Its shares are weighted only, so its raw item is Partial.
  • Infratest dimap's ARD-DeutschlandTREND pages carry a method box (population, mode, sample split by telephone and online, fieldwork, margin of error) but no full tables.
  • Forsa publishes no releases on its own site; its figures reach the public through RTL/ntv, with sample size, fieldwork and a combined share of non-voters and undecided.
  • INSA's site shows the latest "Meinungstrend" with sample size, fieldwork and margin of error; the figures are published mainly by BILD. No tables or undecided share were found.
  • Allensbach keeps its own Sonntagsfrage table, with fieldwork dates and shares to the half point but no sample size; its detailed reports appear in the FAZ, behind a paywall, and were not checked.
  • Ipsos publishes a monthly series in a public chart and describes its method in general terms.

One rule needs stating plainly. An item we could not check scores the same as No. Verian's site returned an error on the day of the review, Civey's method page was not found, Wahlkreisprognose's pages are rendered by script, GMS and pollytix had nothing we could find, and Trend Research was not checked beyond its home page. Their low openness scores reflect what we could verify, not necessarily what they publish. A pollster whose publications become checkable will move up at the next review.

Grades and the current table

The score is \(S_j = \operatorname{round}(A_j + L_j + O_j)\), graded A (75 and above), B (60–74), C (45–59), D (30–44) and E (below 30). A rating is provisional when the pollster has fewer than two checked final polls. The site's table lists every pollster with a final poll or a poll since 2025, and marks those with no poll in the last year as not polling now.

PollsterGradeScoreAccuracyLeadOpennessElections
CiveyA773314301
INFO GmbHB66279301
Forschungsgruppe WahlenB623011217
YouGov DeutschlandB612713213
VerianC59219306
ForsaC573213126
Ipsos DeutschlandC53331282
INSAC52321284
WahlkreisprognoseC50371302
Infratest dimapC46247157
AllensbachD4428977
Cluster 17D43301301
Trend ResearchD39281102
pollytixD37251201
GMSE2819906
Democracy InstituteE2316701
Live German pollster ratings (out of 100) Computed from the database, polls to 17 November 2026.

Given the caveat above, the spread matters more than the letters. The pollsters with six checked elections (FGW, Infratest dimap, Allensbach and GMS) are separated mainly by their records and by openness; Forsa, INSA and Ipsos rank high on accuracy but publish less. The provisional pollsters' record points are pulled strongly towards the prior, because with one election the prior carries two-thirds of the weight.

Each pollster's lean now

Beside the rating the site shows how each pollster differs from the others today. A pollster's lean on party \(p\) is its latest poll within the 21-day window minus the average:

\[ h_{j,p} = s_{j,p} - \bar{s}_{p}. \tag{6} \]

Because the average is built from the same polls, the leans sum to roughly zero across pollsters. A lean is not an error. The pollster furthest from the rest may be the one closest to the result. Each pollster's page therefore shows its lean now next to its lean at past elections, the mean of \(e_{j,t,p}\) over its final polls. For FGW, for example, the past lean is CDU/CSU +1.3 and FDP −0.8, much of it from 2005. A pollster that leans the same way today as it erred before gives a reader something to weigh; one whose past errors went both ways gives much less.

Two German details affect the leans. Several pollsters do not report every party: FGW, Infratest dimap and Allensbach have at times not shown the BSW separately, and FGW's releases between October 2025 and January 2026 put the BSW and FDP inside a single "others" figure. A missing figure is left out of that party's average, so a lean on the BSW or FDP compares a pollster with a smaller group. And INSA publishes the CDU and CSU separately as well as together; we use the combined figure.

State-election records

Germany has no by-elections in the British sense, but it has frequent Landtag elections polled by many of the same firms. Six have been held since the 2025 Bundestag election. For each, we take the last poll by each pollster that finished within 31 days of polling day and compare it with the official 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.

On the pollster pages each pollster's state record is shown beside its Bundestag record. At the time of writing, averaged over the elections each covered, Infratest dimap (six elections) missed by 1.68 points per party, FGW (six) by 1.72, INSA (six) by 1.92, pollytix (two) by 2.29 and Forsa (one) by 3.06.

These records are not part of the score, for three reasons. The electorate, the party system and often the parties' positions differ by Land, so a state miss is weak evidence about Bundestag polls. Fewer pollsters poll most state elections, so the comparison between them is thin. And two of the six results are not final: Berlin's is the provisional result, and Mecklenburg-Vorpommern's is taken from Wikipedia because the official sites could not be reached. The records are a check between Bundestag elections: a pollster that missed badly in every Land would be a warning, and so far none has done so consistently.

FGW's raw "Stimmung" against its "Projektion"

What the two series are

Every German headline figure is a projection. FGW is explicit about it: its "Projektion" adjusts the weighted answers for long-term party attachment and tactical considerations to approximate an election result, and it publishes the "Politische Stimmung" (the weighted answers before that step) alongside. Since 2013 it has stopped publishing the Stimmung in the final pre-election weeks, so for each election we compare the last release that had both, using the Stimmung and the Projektion from the same wave:

ElectionPartiesRawProjection
200553.052.15
200951.861.46
201361.301.15
201762.601.69
202162.302.69
202572.411.17
Live Forschungsgruppe Wahlen: raw "Stimmung" against the "Projektion" in the same release, average miss per party Computed from the database, polls to 17 November 2026.

The projection was closer at five of the six elections. The exception, 2021, is the weakest comparison: the last release with a Stimmung was in mid-August, six weeks before polling day, when the Greens were far higher than they finished, and FGW's final projection that year missed by only 0.91 points per party.

What the projection corrects

The raw figures have a steady lean. They overstated the CDU/CSU at all six elections, by between 0.5 and 4.8 points, and understated the AfD at all four it contested:

Election AfD raw AfD projection (same wave) AfD result
2013 4 4 4.7
2017 7 10 12.6
2021 7 11 10.3
2025 14 20 20.8

Table: FGW's AfD figures, raw and projected, against the Zweitstimmen result.

In 2025 the raw answers had the AfD at 14%; FGW's projection had 20% and the result was 20.8%. Some AfD voters do not say so when asked directly, and FGW's projection adds most of them back. This is the adjustment the raw figures lack, and it is the largest single difference between the two series.

Why we don't start from raw figures, unlike in Britain

In Britain our projection starts from the raw stage of every pollster whose full tables we can read, because the raw answers' misses have been steadier than the headlines' and can be partly corrected, and because that correction beat the published figures out of sample (Why we start from British pollsters' raw answers rather than the published headline, Correcting the raw answers: the method behind our UK projection). Three things stop us doing the same in Germany.

There is one raw series. Only FGW publishes a raw figure. YouGov's tables are weighted, and no other pollster publishes anything before its projection. A method that starts from raw figures would rest on one pollster's figures, with every house effect of that pollster in the forecast.

The published figure is better. At five of six elections FGW's projection was closer than its raw figure, by between 0.15 and 1.24 points per party. In Britain the published headlines were closer than the raw answers too; the case for the raw stage there is that its error is steadier, not smaller. In Germany the steadiness is also there (the CDU/CSU and AfD leans above), but correcting a single pollster's raw figures by their average past lean would amount to rebuilding FGW's projection from six observations, and FGW has decades of its own data to build it from.

The final weeks are missing. The Stimmung is withheld at exactly the time a forecast is tested. Any raw-based correction would be estimated from waves one to six weeks before polling day and applied to final polls it has never seen.

We have not tested a raw-based correction for Germany. We show FGW's Stimmung on the polls page ("Raw figures" tab) beside every Projektion since the 2025 election, so readers can see the gap, and we use the projection.

What the rating leaves out

As in Britain, the rating leaves out how often a pollster polls, who commissions it, and which way it leans today. It also leaves out mode. German pollsters use very different methods, from Allensbach's face-to-face quota interviews to online panels, but mode is a means, and the record already shows how well each pollster's means have worked. Civey's river sampling, which some aggregators exclude, is treated like any other method: it has one final poll, a good one, and is rated provisionally on it.

Limitations

Six elections, one of them unusual. No pollster has more than six final polls, and in 2005 every pollster with a usable final poll overstated the CDU/CSU by more than five points. Pollsters whose records start later are spared that election and look better for it: Forsa's 2005 poll has no shares, INSA's record starts in 2013 and Ipsos's in 2021. The shrinkage limits how much a short record can say, but it does not correct for which elections a pollster happened to cover.

The sources. Most final polls are grade C, transcribed from Wikipedia, whose rows mostly cite wahlrecht.de. They agree with an independent check, but a transcription error would pass straight into a rating.

The scale. The end points and the prior were set for Britain. Against German errors the result is generous, and German and British grades are not comparable.

Methods change. Emnid became Kantar Public and then Verian; FGW moved from telephone to telephone and online in 2025; Ipsos has polled for SPIEGEL since September 2026. A record built on 2005–2025 may describe a pollster that no longer works that way.

Openness is what we could verify. Unknown counts as No. Several low openness scores say as much about the review as about the pollster.

The first real test of whether higher-rated German pollsters are more accurate will be the next Bundestag election.

References

Efron, B. and Morris, C. (1975). Data analysis using Stein's estimator and its generalizations. Journal of the American Statistical Association, 70(350), 311–319.

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). Rating German pollsters: accuracy, the lead and openness. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/germany-pollsters. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like