Research · When the polls missed

Six points lost in the last week: Merkel's vanishing lead and Germany's 2005 polling miss

Abstract. The final polls had the CDU/CSU on 41.5% and eight points clear of the SPD; it won 35.2% and finished less than a point ahead. Every institute overstated the Union by 5.8 to 6.8 points, and their own analyses found a late swing of Union-leaning voters to the FDP that the broadcasters' polls were not allowed to publish.

Polling Forecast · 2 November 2026 · 7 min read · Germany, 2005 Bundestag election, polling error, final polls, inquiry, pollster methods

Every election the polls got badly wrong →

Summary

At the early Bundestag election of 18 September 2005 the final polls put the CDU/CSU on 41.5% and the SPD on 33.4%, a lead of eight points. The Union won 35.2% and the SPD 34.2%; the lead was 0.9 points, a miss of 7.2. Every institute had the Union between 5.8 and 6.8 points too high, while the FDP was 2.7 points too low. The institutes' own analyses found a late swing they could not see or, under the broadcasters' rules, could not publish: Union-leaning voters moved their second vote to the FDP in the last days, and the record number of late deciders broke against the Union after the row over Paul Kirchhof's flat-tax plan.

What the polls said

Gerhard Schröder called the early election after the SPD lost North Rhine-Westphalia in May 2005. By September the Union's lead had shrunk, but every institute still had it above 40%. The last published poll of each:

Pollster (client)	Fieldwork	Published	CDU/CSU	SPD	FDP	Linke	Greens
Allensbach (FAZ)	10–15 Sept	16 Sept	41.5	32.5	8.0	8.5	7.0
Emnid	9–12 Sept	13 Sept	42.0	33.5	6.5	8.0	7.0
Forsa	12–16 Sept	16 Sept	41–43	32–34	7–8	7–8	6–7
Forschungsgruppe Wahlen (ZDF)	6–8 Sept	9 Sept	41.0	34.0	7.0	8.0	7.0
infratest dimap (ARD)	6–7 Sept	8 Sept	41.0	34.0	6.5	8.5	7.0
Average (Forsa at midpoint)			41.5	33.4	7.1	8.1	6.9

The two public broadcasters' institutes published nothing in the last nine days or more: by a standing agreement between the editors of ARD and ZDF, their polls from the final week stayed unpublished. Allensbach, in the FAZ just before the vote, and Forsa, with the midpoints of its ranges, both pointed to a CDU/CSU and FDP majority. The average put the Union and FDP together on 48.6% and the SPD, Greens and Linke on 48.4%.

What happened

The Federal Returning Officer's final result gave the CDU/CSU 35.2% of second votes (16,631,049; CDU 27.8%, CSU 7.4%), the SPD 34.2% (16,194,665), the FDP 9.8%, the Linke 8.7% and the Greens 8.1%. Turnout was 77.7%. The Union won 226 of 614 seats, the SPD 222, the FDP 61, the Linke 54 and the Greens 51. The Union and FDP together had 287 seats, short of the 308 needed, and the result was a grand coalition under Angela Merkel.

Against the average the Union was 6.3 points too high, the SPD 0.8 points too low, the FDP 2.7 too low, the Greens 1.2 too low and the Linke 0.6 too low. Almost the whole miss was the Union's, and much of it reappeared in the FDP. On first votes the Union held 40.9%, close to its 2002 level, which is the signature of ticket-splitting: Union voters kept their constituency vote at home and lent their second vote to the FDP.

Why it happened: what the pollsters found

Germany has no polling regulator and held no formal inquiry; the explanations came from the institutes and from academics.

infratest dimap: Hilmer and Müller-Hilmer, Zeitschrift für Parlamentsfragen 1/2006. The authors, both of infratest dimap, called it the most volatile federal election so far: the lowest turnout (77.7%), the most ticket-splitting (24.4% of polling-station voters), the most late deciders (29%) and the most party switchers (28%). Their last poll before the vote, which was not published, showed the Union's share stable but the proportion of Union voters considering a second vote for the FDP up from 12% to 17%. They tested the late-swing account against the Federal Returning Officer's split between postal and polling-station votes. Postal voters, who vote days or weeks before, gave the Union 39.4%; polling-station voters gave it 34.2%. The gap of 5.2 points compared with 2.3 points in 2002, about three points more than the usual structural difference, and the FDP did almost two points better at the polling station than the postal vote implied. The share of voters still undecided shortly before the vote was about 50% higher than in 2002; 87% of them voted, and the Union won only 23% of them. Their flow analysis had the Union losing a net 1.1 million voters to the FDP. They concluded that the Union lost ground in the last days and rejected talk of a "Waterloo for the pollsters". Their recommendation: institutes publishing on the Friday or Saturday before polling day should refine their questions on strategic voting and field as close to the decision as possible.

Forschungsgruppe Wahlen. Its post-election newsletter put the FDP's gains down to the grand-coalition debate: in the final week 53% of Union supporters opposed a coalition with the SPD, and 68% of voters thought Kirchhof had hurt the Union. In a 2009 review Matthias Jung, of the institute's board, compared the last published Politbarometer (9 September: Union 41, SPD 34) with the result. The left-wing camp was close; the error was inside the centre-right camp, the Union too high and the FDP too low. In the institute's survey in the week before the election (1,603 respondents), 41% of those intending to vote FDP named the Union as the party they liked best, and a week before the vote 42% of Union voters could imagine voting FDP. Ten days out, 22% were still unsure whether and how they would vote. The final-week poll that showed the FDP rising could not be published under the ARD and ZDF agreement.

Konrad-Adenauer-Stiftung: Viola Neu, March 2006. The CDU's foundation noted that all institutes had the Union above 40% and the FDP at 6.5–7% except Allensbach (8.0%), and questioned the trend to one- or two-day fieldwork and falling response rates.

Everhard Holtmann, GWP 1/2006, drawing on a WZB survey by Bernhard Weßels (December 2005). Between mid-August and mid-September 30% of undecided voters moved toward the SPD against 23% toward the Union. About 15% of declared Union voters moved to the FDP, and 93% of those still gave their first vote to the CDU or CSU.

What our data shows

Our database holds five scored final polls for 2005 (Allensbach, Forschungsgruppe Wahlen, GMS, infratest dimap and Emnid, now Verian; Forsa published only ranges). Their average is the same 41.5% for the Union and 33.4% for the SPD. Every one missed by more than two points per party on average, and the Union error ran from 5.8 to 6.8 points, so this was a shared error, not one house.

The timing matters. The polls that finished latest (Allensbach on 15 September, Forsa on 16 September) were no closer on the Union than the ones that finished on 7 and 8 September. Only the FDP moved: Allensbach, the latest, had it highest at 8.0%. That fits a swing in the last two or three days, after the final fieldwork. But the postal vote gave the Union 39.4%, still two points below the poll average, so the polls were probably somewhat high even before the swing.

2005 stands alone in the German record. On the lead between the two largest parties, our final-poll averages missed by 3.1 points in 1998, 2.0 in 2002, 7.2 in 2005, 1.3 in 2009, 3.0 in 2013, 1.7 in 2017, 1.4 in 2021 and 2.3 in 2025. On the Union's own share, the direction flips: overstated in 2005, 2009, 2017 and 2025, understated in 2013 and 2021. Across the six elections from 2005 to 2025 the Union was overstated by about 1.3 points on average, but without 2005 that falls to about 0.3.

What the pollsters changed

The final week is no longer silent. In 2005 the last published ZDF Politbarometer was nine days before the vote. In 2013 ZDF published one on 19 September (fieldwork 18–19 September) for the election of 22 September, and in 2017 on 21 September (fieldwork 20–21 September) for the election of 24 September: three days before polling day, inside the window where the 2005 swing happened.
Strategic voting on the questionnaire. infratest dimap's own recommendation was that institutes publishing at the end of the week should build second-vote and coalition options into their final questions; its unpublished 2005 data had already caught the movement from Union to FDP. Forschungsgruppe Wahlen pointed to the party-potential questions it had asked regularly since 2002, which showed how many voters were open to more than one party in their camp.
"Mood" and "projection" kept apart. Forschungsgruppe Wahlen's answer to the criticism was to insist on the difference between its raw "political mood" and its weighted projection, and that neither is a forecast of a vote weeks away.

What it means for our forecast

2005 is the reason our German forecast does not correct the polls. Our rule takes off half of each party's average past miss and keeps the correction only if it would have improved past forecasts, each election predicted without knowing its result. In Germany it fails that test: the Union's apparent overstatement is mostly 2005, and for every party the sign of the miss changes from one election to the next. So the centre of our forecast is the poll average. 2005 does its work in the size of the simulated error instead: it is the largest miss in the record, and it widens the range our simulations allow for the Union and for the gap between the two largest parties. The method is in our paper on German polling error, and each institute's record, 2005 included, is in our German pollster ratings.

Sources

How to cite

Polling Forecast (2026). Six points lost in the last week: Merkel's vanishing lead and Germany's 2005 polling miss. Retrieved 2 November 2026 from https://www.pollingforecast.com/research/missed-de-2005. Live figures change as new polls arrive; the date above is when the text was last revised.