Simulating German polling error: where the chances come from
Abstract. The German seat ranges, the chance of each party entering the Bundestag and the chance of each coalition holding a majority come from 2,000 simulated elections, each moving every party's national share by a random error the size of its past final-poll miss (at least one point), adding local noise in every Wahlkreis and running the full seat law. Unlike the British simulation, errors are drawn independently for each party, which keeps every party's own uncertainty but ignores how parties' errors move together. The most important output is the chance of clearing 5%, because that single event moves dozens of seats.
Summary
The seat ranges, the chance of each party entering the Bundestag and the chance of each coalition holding a majority all come from 2,000 simulated elections. In each one, every party's national share is moved by a random error whose size is that party's typical miss at the final polls of 2005 to 2025, never less than one point; every Wahlkreis gets a small random local change; and the full seat law is run. The outputs are counts across the 2,000 runs.
The design differs from the British one in one important way. In Britain each run replays a whole past polling miss, so that parties' errors move together as they have at real elections (Simulating UK polling error: where the chances come from). In Germany each party's error is drawn independently. That choice is simpler and easier to defend with six elections whose misses have no stable shape, but it ignores the fact that one party's gain is usually another's loss, and we say below what that does to the ranges. The simulation's most important output is not a seat range at all but the chance that a party near 5% gets in, because that single event moves dozens of seats and decides which coalitions add up.
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.
- Every outcome (3 October). The seats in every simulated election are shown as histograms at /de/outcomes.
From a best estimate to a chance
Our projection is a single set of national shares, the poll average (see Forecasting the Bundestag: an overview), and a single Bundestag worked out from it by the seat law (From votes to seats: the Bundestag seat law). A reader looking at that Bundestag will want to know how sure we are. Will the FDP get in? Does the governing coalition still have a majority if the polls are two points out? Which Wahlkreise could go either way?
To answer, we need a model of how wrong the polls might be, pushed through the seat law many times. The model is in GermanyForecast.Simulate. It uses the same seat engine as the projection, so every run is a complete Bundestag: qualifying parties, 630 seats, Land lists and every Wahlkreis winner, seated or not.
One run, step by step
A run has four steps.
- Move each party's national share. For each of the seven parties modelled (the Union, the AfD, the SPD, the Greens, the Linke, the BSW and the FDP), add a random error drawn from a normal distribution centred on zero, with a standard deviation equal to that party's root-mean-square final-poll miss, or one point if that is larger. Others take whatever is left.
- Spread the shares to the Länder and Wahlkreise. Scale each party's 2025 Land shares and Wahlkreis votes by its new national share, exactly as for the projection.
- Add local noise. Multiply each party's Erststimmen in each Wahlkreis by its own small random factor.
- Run the seat law. Record every party's seats, whether it qualified, every coalition's total, and every Wahlkreis winner and whether it was seated.
The random number generator has a fixed seed, so the same polls always produce the same chances. A change in a published chance therefore always reflects a change in the polls, never a different draw.
National error: one draw per party
The draw
Let \(S_p\) be today's national share for party \(p\) and \(e_{p,t}\) the error of the average final poll at past election \(t\) (poll minus result; see Measuring German polling error, 2005–2025). The typical miss is the root mean square over the \(T_p\) elections at which the party was measured,
and the standard deviation used in the simulation is
A party with no record at all would get 1.5 points; all seven parties now have at least one election. In run \(r\) the simulated share is
and others become \(\max(0.5,\ 100 - \sum_p S^{(r)}_p)\). The floors stop a party from falling to zero or below and keep a sliver of "others" in every run.
The values of \(\sigma_p\), from the six Bundestag elections since 2005, are:
| Party | Elections measured | RMS final-poll miss | \(\sigma_p\) used |
|---|---|---|---|
| CDU/CSU | 6 | 3.12 | 3.12 |
| FDP | 6 | 1.41 | 1.41 |
| SPD | 6 | 1.29 | 1.29 |
| Linke | 6 | 1.09 | 1.09 |
| Greens | 6 | 1.03 | 1.03 |
| AfD | 4 | 0.94 | 1.00 |
| BSW | 1 | 0.40 | 1.00 |
Table: Standard deviation of the simulated national error, points. RMS computed from the final-poll errors in Measuring German polling error, 2005–2025.
The Union's spread is three times that of most parties, and almost all of the difference is 2005. Without 2005 its RMS would be 1.91. We keep 2005 because it happened: it is the largest German polling miss since 1998 in our record, and a simulation that could not produce one like it would be too confident. But it means the Union's range is wide largely because of one election twenty years ago.
Centred on zero
The error is centred on zero. A run is as likely to raise a party as to lower it. This follows from the test in How we read German polls: the 5% line and the 2023 seat law: we found no direction in the past misses that would have helped predict the next one, so the simulation borrows only their size. In Britain, too, the simulation is centred on zero, because the correction has already been applied to the central estimate; in Germany there is no correction, and the centre is simply the polls.
Why independent draws
The British simulation replays whole past misses: in each run it picks one of the four elections from 2015 to 2024, takes the pattern of errors all parties had then, and scales it by one random number. That keeps the correlations between parties. When the Conservatives were understated, Labour was usually overstated, and the simulation reproduces it.
We did not do that for Germany, for two reasons. First, the German patterns are not stable enough to be worth replaying as a structure. The Union was overstated with the SPD understated in 2005, both overstated in 2009 and 2017, and the Union understated with the SPD overstated in 2013; the pairs change from one election to the next. Second, a pattern-based draw can only produce errors in the directions spanned by past patterns. For a party with one past election, like the BSW, or none, the pattern gives it almost nothing. Independent draws give every party its own uncertainty.
The price is that the draws ignore a real constraint. Vote shares add up to 100, so an error that raises one party must lower others, and at real elections the two largest parties' errors are often opposite. In the simulation, others absorb the difference instead. Two consequences follow. The spread of the gap between the two largest parties is somewhat understated when their real errors are negatively correlated, because the draws do not push them apart as often as history would. And the seat ranges of the parties that make up a coalition are treated as if they moved independently, which, if their errors are in fact negatively correlated, overstates the spread of the coalition's total. The past-miss lenses, described below, partly make up for this: each replays one real pattern in full.
From national shares to seats in a run
Länder and Wahlkreise
The simulated national shares go through exactly the same projection as the projection: proportional swing from 2025 to every Land's Zweitstimmen and every Wahlkreis's Erststimmen, with the Union's Land share given to the CSU in Bavaria and the CDU elsewhere (From votes to seats: the Bundestag seat law). There is no separate regional error. A run that raises the AfD by two points nationally raises it by the same proportion in every Land.
Local noise
Wahlkreis contests have their own uncertainty: candidates, local campaigns, tactical voting for the first vote. In each run every party's projected Erststimmen in every Wahlkreis is multiplied by an independent random factor,
A candidate on 30% of the Erststimmen therefore has a local standard deviation of about 2.4 points before the other candidates' noise, one on 10% about 0.8. The noise is proportional to the share, so large parties' candidates are the most exposed to it. It affects which candidate wins each Wahlkreis and the ranking of winners by share, which decides who is seated; it does not affect the Land Zweitstimmen and so, except through the three-Wahlkreis rule, does not change any party's total seats. The 8% figure is a setting, not an estimate: 2025 is the only election under the current law, and we have not fitted the local error to data.
The seat law
Each run then applies the full law: the 5% threshold with the three-Wahlkreis and minority exceptions, Sainte-Laguë/Schepers nationally and to the Land lists, and Zweitstimmendeckung. This is what makes the simulation worth running. The seat law is sharply non-linear at the threshold, and the only reliable way to see what a normal error in votes does to seats is to run the law.
What comes out
After 2,000 runs, every output is a count.
- Seat range
- Each party's seats across the runs are sorted and read off at the 10th and 90th percentiles. That is the 80% range shown on the site.
- Chance of getting in
- The share of runs in which the party wins at least one seat, by any of the three routes.
- Coalition majority
- The share of runs in which the coalition's parties together hold more than half the seats, that is, 316 or more of 630.
- Wahlkreis chances
- For each Wahlkreis, the share of runs in which each party's candidate wins it, and the share in which the projection's projected winner both wins and is seated.
Today's figures, as of 17 November 2026:
| Party | Polls | Seats | 80% range | Gets in |
|---|---|---|---|---|
| AfD | 26.7 | 193 | 176–204 | over 99% |
| CDU/CSU | 21.0 | 152 | 126–170 | over 99% |
| Greens | 15.4 | 111 | 99–120 | over 99% |
| SPD | 13.7 | 99 | 86–109 | over 99% |
| Linke | 10.2 | 74 | 63–82 | over 99% |
| FDP | 4.1 | 0 | 0–40 | 25% |
| BSW | 3.1 | 0 | 0–0 | 3% |
With 2,000 runs, a reported chance has a Monte Carlo standard error of at most \(\sqrt{0.25/2000} \approx 1.1\) points. Differences of a point or two between chances mean nothing.
The 5% line
For a party well above 5%, the simulation adds a range of a few seats either side of the central projection. For a party near the line it does something quite different, and this is where it matters most.
Ignore for a moment the three-Wahlkreis route, which parties near 5% do not take in our projections. A party polled at share \(S\) with spread \(\sigma\) gets in, approximately, when its simulated share reaches 5%, so its chance is
where \(\Phi\) is the standard normal distribution function. The approximation is close rather than exact because 5% is measured against the total of the simulated shares, others included, and because of the floors in equation (3). The table shows what it gives at a range of shares.
| Polled share | \(\sigma\) = 1.00 (AfD, BSW) | \(\sigma\) = 1.41 (FDP) |
|---|---|---|
| 3.0 | 2% | 8% |
| 3.5 | 7% | 14% |
| 4.0 | 16% | 24% |
| 4.5 | 31% | 36% |
| 5.0 | 50% | 50% |
| 5.5 | 69% | 64% |
| 6.0 | 84% | 76% |
Table: Chance of reaching 5% by equation (5), with the spreads used today.
Two features of the table matter. First, at the line itself the chance is a coin toss, whatever the spread. Second, a wider spread helps a party below the line and hurts one above it. The FDP's larger past misses make it more likely to scrape in from 4% than the BSW, and less likely to be safe at 6%.
Because entry is all or nothing, the seat range of a party near 5% is not an interval in the usual sense. Its seats in a run are either zero or, at the threshold, about thirty-five: in 2025 the BSW would have won 34 had it got in. A range of "0 to 40" means that the party gets in often enough to be counted at the 90th percentile, not that twenty seats is a plausible result. The site shows the chance of getting in beside the range for this reason, and it is the better guide.
Coalitions
The site reports five combinations.
| Coalition | Seats | Majority |
|---|---|---|
| CDU/CSU + SPD | 251 | 0% |
| CDU/CSU + Greens | 263 | 0% |
| CDU/CSU + SPD + Greens | 362 | 99% |
| SPD + Greens + Linke | 284 | 1% |
| CDU/CSU + AfD | 345 | 91% |
The seats column is the central projection; the majority column is the share of runs in which the combination reaches 316. A coalition's chance depends above all on who else gets in. When a small party misses the line, its votes are wasted and every qualifying party's share of seats rises; when it clears the line, it takes some thirty-odd seats from the others. That single event can move a two-party coalition across the majority line, as it would have in 2025, when a BSW that cleared 5% would have left the CDU, CSU and SPD with 310 seats instead of 328. The chance of a coalition majority is therefore driven as much by the small parties' chances of getting in as by the large parties' shares.
The CDU/CSU and the AfD are shown together for the arithmetic only. The CDU/CSU has ruled out working with the AfD, a commitment known as the Brandmauer (firewall), and we make no forecast that such a coalition will form. We show it because the rest of the table cannot be read without it: in a Bundestag where the AfD holds a large block of seats, the number of seats left to everyone else, and so which other combinations can reach a majority, depends on it.
Coalition chances inherit the limitation of independent draws discussed above. At real elections the errors of coalition partners have sometimes run together (the Union and the SPD were both overstated in 2009 and 2017) and sometimes against each other (2005 and 2013). Independent draws average over both. Whether that makes the coalition chances too confident or too cautious depends on the next election's pattern, and we have not quantified it.
Unseated Wahlkreis winners
Under the 2023 law a Wahlkreis winner takes a seat only if the party's Land seats cover them. In 2025 that left 23 Wahlkreise without a member. On today's projection the number is 22.
The simulation gives each Wahlkreis two chances: that the projection's projected winner wins, and that the winner also takes a seat. The two differ most where a party is projected to win more Wahlkreise in a Land than its list vote there covers, and where its winners' shares are close to the cut. A winner projected at the bottom of its party's ranking in a Land can have a high chance of winning and a much lower chance of sitting. The Wahlkreise page shows both for every Wahlkreis, and the map shades them.
The local noise of equation (4) is what drives the difference. It moves winners up and down their party's ranking within a Land, and the ranking decides who is seated. Because that noise is a setting rather than an estimate, the seated chances are less well founded than the winning chances.
Lenses beside the simulation
The simulation says how wrong the polls might be in general. The lenses say what happens if they are wrong in one specific way that has happened before. Each lens takes today's poll average, subtracts the errors of one past election's final polls, rescales, and runs the seat law once (If we're wrong: replaying past Bundestag polling misses on today's polls).
| Reading | CDU/CSU | AfD | SPD | Greens | Linke | BSW | FDP |
|---|---|---|---|---|---|---|---|
| The polls as they stand | 152 | 193 | 99 | 111 | 74 | 0 | 0 |
| Our projection (the polls, uncorrected) | 152 | 193 | 99 | 111 | 74 | 0 | 0 |
| If the polls miss like 2002 | 159 | 189 | 94 | 118 | 69 | 0 | 0 |
| If the polls miss like 2005 | 102 | 187 | 102 | 114 | 76 | 0 | 48 |
| If the polls miss like 2009 | 139 | 187 | 80 | 110 | 75 | 0 | 38 |
| If the polls miss like 2013 | 165 | 197 | 91 | 104 | 72 | 0 | 0 |
| If the polls miss like 2017 | 124 | 196 | 87 | 116 | 68 | 0 | 38 |
| If the polls miss like 2021 | 165 | 191 | 102 | 107 | 64 | 0 | 0 |
| If the polls miss like 2025 | 142 | 192 | 107 | 100 | 88 | 0 | 0 |
The lenses are not averaged into the simulation, and they are not probabilities. They are useful for two things. They show, in the plainest terms, how much a repeat of 2005, when the Union was overstated by six points and the FDP understated by nearly three, would change the Bundestag. And they carry what the independent draws leave out: each lens moves the parties together, in a pattern that really occurred. 102 Wahlkreise change hands under at least one of the six.
Before polling day
Everything above describes an election held today. The election-day forecast runs the same simulation with a second source of uncertainty: how far opinion may move before polling day. Each party's standard deviation becomes
where \(\delta(d)\) is the drift for \(d\) days to go, estimated from German polls at fixed distances before past elections. At 855 days it is 5.2 points, several times most parties' final-poll error, and it applies equally to every party. The polling-day simulation uses 1,000 runs. From today to polling day in Germany gives the evidence and the formula.
What the simulation does not capture
Correlation between parties. The national errors are independent. At real elections they are not, and the effect on coalition chances in particular is unmeasured.
Regional error. A miss that falls more heavily in the East than the West, or in one Land, is not represented. The state elections since 2025 show misses that differ a good deal from state to state (Measuring German polling error, 2005–2025), and a regional component would widen the Land and Wahlkreis ranges.
Seat-model error. Proportional swing from 2025 has not been tested against a past election (From votes to seats: the Bundestag seat law). The local noise stands in for its errors but is independent from Wahlkreis to Wahlkreis, while real errors of a swing model tend to run in the same direction across similar places.
The Union's record. The Union's spread rests mostly on 2005. If that election was a one-off, the Union's range is too wide; if the next miss is like it, the other parties' ranges are too narrow.
Small records. The AfD's spread rests on four elections and the BSW's on one. Both are set by the one-point floor, not by evidence.
Everything outside polling error. Events, changes of leader, a new party and a change to the electoral law before 2029 are outside the model. The last would matter most: the three-Wahlkreis rule is an interim rule, and the threshold itself may be changed.
Calibration with six elections
A probability forecast is well calibrated if events given 20% happen about 20% of the time. Six elections cannot test that. What we can say is that the design reproduces, by construction, the size of the misses seen since 2005: a party's simulated national error has a standard deviation equal to its past root-mean-square miss, or more. It does not reproduce their correlations, and it does not know about misses larger than any in the record. The long view of German polling back to 1998 (Measuring German polling error, 2005–2025) found no election, other than 2005, with an average miss above 1.5 points per party, which is some comfort about the size of the spreads. It says nothing about their shape.
We think the ranges for the large parties are about right, possibly too wide for the Union, and that the chances of the small parties getting in are the best-founded numbers the simulation produces, because they depend mainly on the size of past misses, which is what the simulation is built from. Coalition chances are less secure, because they depend on correlations the simulation leaves out. All of them should be read as a guide to how the forecast could go wrong in familiar ways, not as a complete account of the uncertainty.
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.
How to cite
Polling Forecast (2026). Simulating German polling error: where the chances come from. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/germany-simulation. Live figures change as new polls arrive; the date above is when the text was last revised.



