How we read German polls
How we turn Germany's polls into a forecast for the Bundestag (630 seats, 316 for a majority): the 16 polls of the last 30 days, corrected for how Germany's final polls have missed before. Today's average: AfD 26.7, CDU/CSU 21.0, Greens 15.4, SPD 13.7.
Our rule everywhere: measure how far final polls missed at past elections, take off half the average miss, and keep the correction only if it would have made past forecasts better when each election is predicted without knowing its result. In Germany the polls as published have done at least as well without one, so our projection takes them as they are and puts the effort into what decides the seats.
Why we don't correct
Averaged over the 7 elections with final polls from 2000 on (2002–2025; older ones are shown in the track record but not used by the model) the misses look like a pattern (the CDU/CSU overstated by 0.9 points, the FDP understated by 0.5), but most of that is 2005. Party by party the sign flips from one election to the next, and the polls as published were closer at 5 of 7 elections. The same holds on 2005–2025 alone (3 of 6). The test →
What we do instead
- The 5% line. Each simulation moves every party by its typical final-poll miss, then runs the full seat law. The share of runs where a party gets seats is its chance of being in the Bundestag.
- The seat law. Our calculation reproduces the official 2025 result exactly: every party, every Land list and all 299 Wahlkreise.
- Unseated winners. Since 2025 a Wahlkreis winner sits only if the party's list vote covers them, so we project every Wahlkreis too.
- If we're wrong. Today's polls replayed with each past election's miss, every seat worked out again.
- 57 final polls at the 7 Bundestag elections the model uses (2002, 2005, 2009, 2013, 2017, 2021, 2025), and 10 more at 1953, 1972, 1976, 1998, shown in the track record but not used by the model (it uses no election before 2000). We found no final polls at all for 1957–1969, 1980–1994.
- Results by Wahlkreis for every Bundestag election since 1949, each on its own boundaries, and the Federal Returning Officer's recalculations of 2002–2017 onto the next election's Wahlkreise, for the swing tests (pairs before 2000 are shown but not used by the model).
- 280 poll averages at set distances before polling day, for the election-day test.
- Forschungsgruppe Wahlen's raw figures beside its projection, for the raw-figures test.
- Final polls at every Landtag election since 2025.
- Average: each pollster's latest poll within 21 days of the newest, equal weight per pollster.
- Länder: each party's 2025 Zweitstimmen share in every Land, scaled by its national change (proportional swing). The CDU/CSU's share goes to the CSU in Bavaria and the CDU elsewhere.
- Wahlkreise: each candidate's 2025 Erststimmen share plus the party's national change in points (uniform swing; smaller parties and independents in proportion), with a small random local change in each simulation. Uniform swing named more Wahlkreis winners right at four of six past pairs of elections (tests tab).
- Seats: 630 seats by Sainte-Laguë/Schepers among parties with 5% of Zweitstimmen, three Wahlkreis wins or national-minority status (the SSW); then to each party's Land lists; then Wahlkreis winners seated in order of their Erststimmen share until the Land's seats run out (the 2023 law as amended by the Constitutional Court on 30 July 2024).
- Chances: 5000 simulations, each party moved by its typical final-poll miss (at least one point).
- Polling day: the same, with extra drift for the days left: rmse(days) = 2.08 + 0.93 × ln(days / 30).
How the local model did
Our seats come from moving each party's last Land Zweitstimmen by its national change in proportion (proportional swing), and each candidate's Erststimmen share by the same change in points (uniform swing), then the seat law. To test that step on its own, we took 2021, recalculated by the Federal Returning Officer onto the 2025 Wahlkreise, moved every party by its true national Zweitstimmen change to 2025, gave the BSW (new in 2025) its national share everywhere, and ran the 2025 seat law. No polls are involved: this is the error the local step adds even when the national vote is known exactly.
| 2025 | Seats | Real | Wahlkreise | Real |
|---|---|---|---|---|
| CDU/CSU | 208 (=) | 208 | 211 (+21) | 190 |
| AfD | 152 (=) | 152 | 44 (−2) | 46 |
| SPD | 120 (=) | 120 | 31 (−14) | 45 |
| Greens | 85 (=) | 85 | 10 (−2) | 12 |
| Linke | 64 (=) | 64 | 3 (−3) | 6 |
| SSW | 1 (=) | 1 | 0 (=) | 0 |
| Wahlkreise with the right winner | 269 / 299 |
| … and right on whether the winner is seated | 258 / 299 |
| Unseated winners: projected / actual | 26 / 23 |
| … the same Wahlkreis and party in both | 15 |
| Seats given to the wrong party | 0 / 630 |
| Land list seats in the wrong Land or party | 16 / 630 |
| Erststimme share error per party, points | 2.4 |
| Zweitstimme share error per party and Land, points | 1.4 |
Projected from 2021, with the change against the real result in brackets. Share error: the average gap between projected and real share for each party with 1% or more in a Wahlkreis (or Land), others left out.
The 30 Wahlkreise with the wrong winner
Largest Erststimme errors: Suhl – Schmalkalden-Meiningen – Hildburghausen – Sonneberg 6.5, Erfurt – Weimar – Weimarer Land II 6.2, Bad Kissingen 6.1, Berlin-Friedrichshain-Kreuzberg – Prenzlauer Berg Ost 6.1, Berlin-Neukölln 5.4
Where the swing still misses
- Candidates. The SPD's Erststimmen fell much less than its Zweitstimmen (from 26.4% to 20.1%, against 25.7% to 16.4%). Moving candidates by the party's Zweitstimmen change still costs the SPD: we give it 31 Wahlkreise against 45. Proportional swing, which we used until September 2026, cut SPD candidates much harder (11 Wahlkreise). Moved by each party's own Erststimmen change, 273 of 299 winners come out right; polls ask only about the Zweitstimme, so the forecast cannot do this.
- Where a party's new voters live. Die Linke nearly doubled, but its gains came in the big cities, not its old eastern base, so doubling its 2021 vote puts it ahead in Rostock and Marzahn-Hellersdorf, which the AfD won. The Greens held their city seats (Berlin, Köln, Münster, Freiburg) better than their national fall.
- Unseated winners follow from the winners. With too many CDU/CSU Wahlkreise we still leave 26 winners without a seat against 23; 15 of the real ones are named correctly.
- Seats per party follow the national Zweitstimmen alone under the 2025 law, so with the true shares they are right by construction. But summed over the Länder, proportional swing does not give back the national shares exactly: left alone, the BSW comes out at 5.01% instead of 4.98% and wins 35 seats. The test scales each party's Land votes to its true national total, and the forecast does the same with the poll shares, so the 5% line is judged on the share that went in.
Proportional or uniform? Every pair of elections since 1949
The same replay on every pair of Bundestag elections we hold: the earlier result moved by each party's true national Zweitstimmen change, once in proportion and once in points, against what happened. Wahlkreise can be compared from 2002→2005, where the Federal Returning Officer recalculated each earlier result onto the next election's Wahlkreise; Länder from 1949→1953. Only pairs with both elections from 2000 on count (14 older pairs are shown but not used by the model). In the Wahlkreise uniform swing named 1602 of 1794 winners right, proportional swing 1521, better at 4 of 6 pairs. In the Länder proportional swing was closer at 4 of 6 pairs. So the forecast moves Wahlkreise uniformly and Länder in proportion.
| Pair | Winners right, uniform | … proportional | Erststimme error, uniform | … proportional | Land error, uniform | … proportional |
|---|---|---|---|---|---|---|
| 2021 → 2025 | 269 / 299 | 256 | 2.5 | 2.3 | 1.8 | 1.3 |
| 2017 → 2021 | 244 / 299 | 215 | 2.4 | 2.5 | 1.8 | 1.6 |
| 2013 → 2017 | 276 / 299 | 276 | 3.1 | 3.0 | 2.0 | 1.7 |
| 2009 → 2013 | 269 / 299 | 276 | 1.9 | 1.6 | 1.4 | 1.0 |
| 2005 → 2009 | 270 / 299 | 250 | 2.0 | 2.6 | 1.6 | 1.8 |
| 2002 → 2005 | 274 / 299 | 248 | 2.7 | 2.6 | 1.9 | 2.0 |
| 1998 → 2002 not used by the model |
– | – | – | – | 2.1 | 1.7 |
| 1994 → 1998 not used by the model |
– | – | – | – | 1.2 | 1.2 |
| 1990 → 1994 not used by the model |
– | – | – | – | 2.4 | 2.1 |
| 1987 → 1990 not used by the model |
– | – | – | – | 1.7 | 1.6 |
| 1983 → 1987 not used by the model |
– | – | – | – | 0.9 | 0.8 |
| 1980 → 1983 not used by the model |
– | – | – | – | 1.2 | 1.0 |
| 1976 → 1980 not used by the model |
– | – | – | – | 0.7 | 0.6 |
| 1972 → 1976 not used by the model |
– | – | – | – | 0.8 | 0.7 |
| 1969 → 1972 not used by the model |
– | – | – | – | 1.5 | 1.5 |
| 1965 → 1969 not used by the model |
– | – | – | – | 1.2 | 1.0 |
| 1961 → 1965 not used by the model |
– | – | – | – | 1.9 | 2.1 |
| 1957 → 1961 not used by the model |
– | – | – | – | 2.2 | 2.9 |
| 1953 → 1957 not used by the model |
– | – | – | – | 1.5 | 1.5 |
| 1949 → 1953 not used by the model |
– | – | – | – | 2.1 | 2.2 |
Error: average gap in points per party with 1% or more, others left out. Uniform swing applies to the main parties; smaller parties and independents always move in proportion. Before 1990 West Berlin did not elect members directly, so the Länder are the ten western ones (nine before the Saarland voted in 1957). Source: the Federal Returning Officer's results by Wahlkreis for every election since 1949 (btw_kerg).
Correcting final polls
Each election is predicted with the correction worked out from the other elections only, so it never sees the result it is tested on. Correcting German polls was closer at 2 of 7 and further out at 5 (average miss per party 1.30 against 1.21 for the polls as published). So our projection leaves these polls as they are.
Elections before 2000 (1953, 1972, 1976, 1998) are not in this test: they are shown in the track record but not used by the model.
| Election | Polls' average miss | Corrected | Verdict |
|---|---|---|---|
| 2002 | 0.95 | 1.15 | Correction further out |
| 2005 | 2.33 | 2.37 | Correction further out |
| 2009 | 1.10 | 0.97 | Correction closer |
| 2013 | 0.94 | 1.09 | Correction further out |
| 2017 | 1.48 | 1.40 | Correction closer |
| 2021 | 0.80 | 1.13 | Correction further out |
| 2025 | 0.89 | 1.01 | Correction further out |
Pulling back towards the last result
Not closer than the polls as they stand (0 of 10 distances better), so not used. Details →
The seat law
Given the official 2025 Zweitstimmen and Erststimmen, our calculation gives every party's seats, every Land list and every seated and unseated Wahlkreis winner exactly as the Federal Returning Officer did.
- Proportional swing by Land ignores local candidates, which matter for Wahlkreise.
- Parties near 5% dominate the uncertainty; one pollster's methods can move them a point.
- Regional polls are too few to use.
- Eleven elections with final polls, but only six with more than three pollsters (2005–2025), and 2005 still dominates the averages.
Figures as of . Published by Polling Forecast.



