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

What we don't know yet: limitations of the German forecast

Abstract. The German forecast rests on six Bundestag elections, with 2005 dominating the averages and the simulation's spread. Its uncertainty is concentrated at the 5% line, where independent party errors and an interim three-Wahlkreis rule are simplifications; it has one election of history for the BSW, an untested proportional swing with no candidates, too few regional polls, an uncertain election date and mostly second-hand polls. For each we say how much it matters and what evidence would change our mind.

Polling Forecast · updated 9 November 2026 · 14 min read · Germany, limitations, Bundestag, 2005, 5% threshold, BSW, new parties, proportional swing, regional polls, election date, publication practice

Summary

The German forecast makes fewer claims than the British one. It does not correct the polls, because a correction failed its test; it takes the poll average, runs it through a seat calculation that reproduces 2025 exactly, and puts its uncertainty where the seats are most sensitive, at the 5% line. That leaves less to go wrong in the method, but the evidence under it is thin. It rests on six Bundestag elections, one of which (2005) dominates the averages. Its seat projection moves every Land and Wahlkreis by proportional swing from 2025, with no candidates in it. It has one election of history for the BSW and none for any party that might appear next. The regional polls that could test its geography are too few. The election may come at any time before 25 March 2029. And most of its polls reach us second-hand.

This paper sets out each weakness, how large we think it is, and what evidence would change our mind. The British counterpart is What we don't know yet: limitations and open questions in our UK forecast; the method itself 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.
  • Pollster grades rescaled (13 October) (Rating German pollsters: accuracy, the lead and openness).

Six elections, dominated by 2005

Everything the German forecast knows about polling error comes from the final polls at six Bundestag elections, 52 polls in all.

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.

Six elections is little, and one of them is unlike the others. In 2005 every final poll overstated the CDU/CSU by more than five points, and the average overstated it by 6.3. No other party-election error in the record exceeds 3.0. The average miss per party in 2005 was 2.33 points; at no other election did it exceed 1.48.

The consequences run through the method:

Quantity With 2005 Without 2005
CDU/CSU mean error +1.29 +0.29
FDP mean error −0.79 −0.38
CDU/CSU spread in the simulation (root mean square error) 3.12 1.91
FDP spread in the simulation 1.41 1.00 (the one-point floor)

Table: How much of the German evidence is 2005. Points; computed from the final polls in our database.

The "pattern" that tempts a correction (the CDU/CSU overstated, the FDP understated) is mostly 2005. The leave-one-out test (How we read German polls: the 5% line and the 2023 seat law) is the guard against being misled by it, and it worked: the correction was closer at two elections, further out at three, and level at 2005 itself, where the other five elections had nothing like 2005's miss to teach. But 2005 still shapes the simulation. The CDU/CSU's national share is drawn with a standard deviation of 3.1 points, set by the root mean square of its six errors, and without 2005 it would be 1.9. The CDU/CSU's seat range on the home page is correspondingly wide. Whether that width is right depends on whether a 2005-sized miss can happen again, and six elections cannot say.

There is also a structural limit that the British record shares: the misses are not independent draws. All polls at an election share the same electorate, campaign and late movement, so 52 polls carry roughly six elections' worth of information, not 52. Earlier elections would help, but the long view from 1953 to 2002 (Measuring German polling error, 2005–2025) has no final polls at all for nine elections and one pollster each for two more; it is shown for context and not used.

What would change our mind. Each new Bundestag election is added to the record automatically and the backtest rerun. If, with more elections, a correction beat the plain average on the leave-one-out test, the code would switch it on without any change to the method: the rule is checked every time the forecast is built. We would also revisit the simulation's spread if a further election or two showed misses well inside or well outside it.

The 5% line

The seat forecast is most fragile at the threshold. A party at 4.9% gets no list seats; at 5.0% it would have had about 34 in 2025 (If we're wrong: replaying past Bundestag polling misses on today's polls). In 2025 the BSW finished 4.981%, about 9,500 votes short, and had it got in the CDU/CSU and SPD would not have had a majority. Small errors near the line decide governments.

The Landtag elections since 2025 show how often threshold parties are misjudged. In Baden-Württemberg the final polls had the FDP and Die Linke at 5.5 to 6% and both finished on 4.4%, out. In Sachsen-Anhalt the BSW was polled at 4% by every pollster and got 5.3%, in. In Mecklenburg-Vorpommern the CDU was polled at 6 to 10% and got 4.9%, out. These are state elections with different electorates, and the average miss per party was not unusual. But in each case a miss of a point or so crossed the line.

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.

We handle the threshold in the simulation, not the central forecast. Each of 2,000 runs moves every party's national share by a random normal error with the standard deviation of that party's past final-poll misses, at least one point, and runs the full seat law. The share of runs in which a party wins seats is its chance of entering the Bundestag (Simulating German polling error: where the chances come from).

PartyPollsSeats80% rangeGets in
AfD26.7193176–204over 99%
CDU/CSU21.0152126–170over 99%
Greens15.411199–120over 99%
SPD13.79986–109over 99%
Linke10.27463–82over 99%
FDP4.100–4025%
BSW3.100–03%
Live Our projection: Zweitstimmen and seats Computed from the database, polls to 17 November 2026.

Three simplifications in that simulation matter most at the threshold.

Errors are drawn independently for each party. In reality polling errors are linked: if the polls overstate one party, they understate others, since shares add up to 100. In the simulation the parties' errors are unrelated and the seven shares are not rescaled; "others" absorb the difference. A run in which the FDP is understated is not more likely to have the CDU/CSU overstated, although in 2005 and 2017 that is exactly what happened. The effect on each party's own chance of entering is probably small; the effect on coalition chances, which depend on several parties at once, could be larger.

The spread is a party's average record, not its record near 5%. The FDP's spread is set by its misses at shares from 4.3% to 14.6%. Whether polls err differently for a party hovering at the line (tactical Zweitstimmen from supporters of a larger partner, the "Leihstimmen" often discussed in German coverage) is not modelled. The state elections above hint that threshold parties can be misjudged in either direction.

The three-Wahlkreis rule is uncertain in law. It was abolished in 2023 and restored by the Federal Constitutional Court on 30 July 2024 as an interim rule until the Bundestag legislates. The seat calculation applies the rule as it stood for the 2025 election. A new electoral law before the next election could change it, or change the threshold, the number of seats or the treatment of unseated winners. We would have to change the calculation, and the 2025 check that it reproduces the official result would no longer apply.

What would change our mind. A study of polling error for parties within a point or two of a threshold, across Bundestag and Landtag elections, would tell us whether their spread should be wider or narrower than their average. The state-election record is growing by several elections a year; once it is long enough we will test whether it improves threshold chances.

New parties: the BSW and whoever comes next

The BSW stood for the first time in 2025. Its record is one election, at which the final polls understated it by 0.4 points. That is too little to estimate its spread, so the simulation uses the one-point floor, and only one of the six replays in the lenses moves it at all.

A single election says little about how well the BSW's support is measured. The AfD's record is the nearest guide to how a new party's polling can go: understated at its first two elections (by 0.8 in 2013 and 1.6 in 2017), then roughly right. FGW's raw figures understated the AfD by far more, and FGW's projection added most of that back (Rating German pollsters: accuracy, the lead and openness). Whether pollsters' projections handle the BSW as well is unknown. Several pollsters did not show the BSW separately at all in some releases after 2025, which also means today's BSW average rests on fewer pollsters than the other parties'.

A party that is not one of the seven modelled parties is folded into "others", and in the seat calculation "others" never qualify. If a new party rose above 5% before the next election, the forecast would give it no seats until we added it to the model. Adding a party is a small change in code, but the party would have no 2025 base in any Land or Wahlkreis to swing from, and proportional swing cannot work from zero. We would need another way to place it.

What would change our mind. More elections with the BSW on the ballot, federal and state; and, for a new party, a rule for placing it geographically, such as its results in Land or European elections, decided before it is needed rather than after.

Proportional swing and local candidates

The forecast moves each party's 2025 Zweitstimmen share in every Land by the ratio of its national poll share to its 2025 national share, and moves each party's 2025 Erststimmen in every Wahlkreis the same way. This is proportional swing. It is the simplest model that respects the scale of German regional differences (the AfD's 2025 share was 32.0% in the East and 18.0% in the West), and it keeps every share above zero. It has three known problems.

It amplifies change where a party is already strong. A party that gains a third of its vote nationally gains a third everywhere, so where it had 32% it gains more than ten points and where it had 18% it gains six. If a party's growth is actually concentrated where it was weak, or capped where it was strong, the Land and Wahlkreis projections will be wrong in a systematic way. The regional polls below are the only evidence we have on this, and they are too few to test it.

It has no candidates. The Erststimme is cast for a person. Incumbency, a well-known candidate, a retirement, a local controversy or a tactical campaign can decide a close Wahlkreis whatever the national polls say, and none of it is in the model. The simulation adds a random local change in every run (each party's Erststimmen in each Wahlkreis multiplied by a factor with a standard deviation of 8%), but that change is independent from Wahlkreis to Wahlkreis and from party to party, and its size was chosen, not estimated. Under the 2023 law this matters less than in Britain, because the Wahlkreise decide who sits, not how many seats each party gets. It matters more for the list of unseated winners, which depends on the ranking of a party's winners by Erststimmen share within each Land.

It assumes Bavaria moves with the rest. Polls report the CDU and CSU together. The forecast gives the Union's national change to the CSU in Bavaria and to the CDU everywhere else in the same proportion. If the CSU's Bavarian vote moved differently from the CDU's elsewhere, the seat split between them, and the seated status of Bavarian Wahlkreis winners, would be wrong even with a correct national figure. The single Bavarian poll we hold is not enough to say.

We have not tested proportional swing against a uniform or mixed alternative on a German backtest. The obvious test, predicting 2025 from the 2021 result recalculated onto the 2025 boundaries with the true 2025 national shares, is possible with the data we hold and has not been run. The Strukturdaten for every Wahlkreis are loaded but not used in the model.

What would change our mind. That backtest, run for proportional, uniform and mixed swing, would tell us which to use. A model using the Strukturdaten or regional polls would have to beat the winner of that test before replacing it.

Regional polls: too few

A handful of pollsters publish Bundestag voting intention for the East and West or for a single Land. Since the 2025 election we hold 68 such polls: 30 for the East and 29 for the West, nearly all from YouGov and INSA, and nine for single Länder, all from Forsa and infratest dimap. YouGov's come from its own result tables and three others from the commissioning outlet's chart; the other 27, mostly INSA's, come from Wikipedia's tables.

That is too few to use. Two pollsters dominate the East and West series, so any house effect of theirs would pass straight into the regional forecast. Single-Land polls cover three of sixteen Länder. And a regional poll's sample is a fraction of a national poll's, so its sampling error is larger. The forecast therefore uses the national average moved by each Land's 2025 result, and shows the regional polls on the Länder page against the 2025 result in the same area, so a reader can see where they disagree with proportional swing.

What would change our mind. More regional polls from more pollsters. If the East and West polls consistently disagreed with what proportional swing implies for the East and West, by more than their sampling error and across pollsters, we would test blending them in, and keep the blend only if it improved a backtest.

The election date

The next Bundestag election must be held by 25 March 2029, 48 months after the Bundestag first met. The forecast assumes that date for the election-day projection, 855 days after the latest poll. But German elections can come early: two of the six in our record, 2005 and 2025, followed a lost confidence vote.

The date matters because polls move. Polls this far out have missed by about 5.2 points per party by polling day (From today to polling day in Germany), several times the final-poll miss. The election-day forecast widens its ranges accordingly, from a curve fitted to five elections (2009 to 2025):

Days before polling dayElectionsPolls' missPrevious result's missPull-back's missRMSE
3051.435.111.481.84
6152.165.112.242.93
9152.455.112.543.39
18252.785.112.884.01
27352.935.113.134.08
36552.945.113.184.07
54853.205.113.574.01
73054.535.114.605.40
91354.425.114.525.48
109554.525.114.625.59
Live Bundestag: polls against the result by distance from polling day, each election predicted from the others Computed from the database, polls to 17 November 2026.

Two limits follow. The curve is fitted to five elections, and the evidence at long distances, where fewer polls exist, is thinnest. And unlike in Britain, pulling parties part of the way back towards their last result did not beat the polls as they stood at any of the ten distances tested, so the German election-day forecast keeps today's shares and only widens the ranges. If the election came early, the ranges would narrow at once; the forecast has no way to anticipate that.

What would change our mind. Each election adds one more set of distances to the curve. If a pull-back towards the last result began to beat the polls in the leave-one-out test, it would be used, as it is in Britain.

Publication practice

The German forecast depends on polls reaching us, and German pollsters publish less than British ones.

Many INSA polls are second-hand. Of 508 German polls in the database, 223 are grade C, taken from Wikipedia's tables, and almost every one of those rows cites wahlrecht.de rather than the pollster. Since the 2025 election most national polls come from the pollster's own release (117) or its client's own article or chart (111); 159 are still grade C, 138 of them INSA's, because INSA keeps its past Meinungstrends in a paid archive and BILD's reports are mostly behind BILDplus. Where we could compare, the Wikipedia rows agree with the releases apart from a handful of figures, dates and samples (The German data: results, boundaries, polls and licences). Forsa's polls are grade B, because forsa publishes only through its clients.

Some new polls are missed. Every morning a reader for each pollster reads its newest poll from its own page, or its client's, and stores it after checks. INSA's weekend poll is not on INSA's page, GMS does not post every release, and a Verian poll FOCUS prints only in the magazine is missed; those are entered by hand if at all. A gap does not distort the average much, because each pollster counts once whatever its frequency, but a pollster whose polls stop arriving drops out of the 21-day window and the average is built from fewer.

Pollsters report different things. Some do not show the BSW or the FDP separately in every release; FGW put both in "others" for several months. Undecided shares are defined differently and published irregularly. Allensbach publishes to the half point. Ipsos's fieldwork dates come from Wikipedia. Each of these is small, and each makes the average a little less like-for-like.

Raw figures are rare and withheld when they matter. Only FGW publishes a figure before its projection, and since 2013 it has stopped doing so in the final weeks before an election. This is the main reason the British approach of starting from raw figures cannot be used (Rating German pollsters: accuracy, the lead and openness).

The pollster ratings are calibrated on Britain. The ratings use British end points and a prior set at the 75th percentile of German final polls. Because German polls have been more accurate, an unproven German pollster starts at a C where a British one starts at a D or below. The order of German pollsters is informative; their grades are not comparable with British ones. A German scale is an open item.

What would change our mind. Pollsters publishing their own series in a readable form, which would let us grade more polls A and automate their collection; FGW continuing to publish its Stimmung through the campaign.

Summary

Limitation How much it matters What would address it
Six elections, 2005 dominant The simulation's spread for the CDU/CSU and FDP rests heavily on 2005 More elections; the backtest reruns itself with each
Threshold parties The largest single source of seat uncertainty Study of polling error near thresholds, including Landtag elections
Independent errors in the simulation Coalition chances may be too confident or too cautious Draw correlated errors from past elections' error patterns
Three-Wahlkreis rule is interim A new law could change the calculation Implement any new law when passed
New parties BSW has one election of history; a new party would get no seats in the model More elections; a rule for placing new parties
Proportional swing untested Land and Wahlkreis projections may be biased Backtest 2021 to 2025 against uniform and mixed swing
No candidates Close Wahlkreise and unseated winners uncertain Candidate data where the official source publishes it
Regional polls too few East–West differences taken on trust from 2025 More regional polls from more pollsters
Election date Early election would narrow ranges suddenly None possible; ranges shown for the latest date
Second-hand polls Transcription errors would pass into the average Pollsters' own series; more checks against releases
Ratings calibrated on Britain German grades read high A German scale

Table: Known limitations of the German forecast.

The forecast's claim is narrow: the German polls, read as they are, carried through a seat calculation that reproduces the last election exactly, with their uncertainty where it matters. The next Bundestag election will be the first real test, and we will publish how it went.

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). What we don't know yet: limitations of the German forecast. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/germany-limitations. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like