Research · When the polls missed

Right winner, wrong size: how every poll overstated Labour's lead in 2024

Abstract. All 19 final polls listed by the British Polling Council overstated Labour's lead, averaging 17.2 points against an actual 10.3, the largest overestimate of Labour since 1979. Pollsters' reviews pointed to too few older and loyal Conservatives in their samples, high refusal rates, over-engaged online panels and some late switching away from Labour.

Polling Forecast · 2 November 2026 · 7 min read · United Kingdom, 2024 general election, polling error, final polls, inquiry, pollster methods

Summary

The polls called the Labour landslide of 4 July 2024 but overstated it. The 19 final polls listed by the British Polling Council averaged Labour 38.6%, Conservatives 21.4%, a lead of 17.2 points. Keir Starmer's Labour won Great Britain by 10.3 points, 34.7% to 24.4%, so the polls missed the lead by 6.9 points, and every one of them overstated it. The BPC called it the largest overestimate of Labour since 1979; the companies' own reviews found no single cause but pointed to late switching away from Labour, high refusal rates among older voters, unrepresentative minority subsamples and over-engaged online panels.

What the polls said

In mid-June Labour's lead was around 20 points: Focaldata's poll of 14–17 June had it at 22, which the firm says was in line with other pollsters at the time. The final round narrowed it only a little. The BPC's table of final polls, published on 8 July 2024:

Pollster	Fieldwork end	Con	Lab	LD	Ref	Grn
Savanta	3 July	20	39	10	17	5
We Think	3 July	23	41	11	15	7
JL Partners	3 July	23	38	13	17	5
Survation (phone)	3 July	20	38	12	17	7
Norstat	3 July	24	37	11	16	6
Opinium	3 July	21	41	11	17	7
Ipsos (phone)	3 July	19	37	11	15	9
Deltapoll	3 July	22	39	10	17	7
Prolific	2 July	18	36	13	15	6
People Polling	2 July	16	36	10	20	9
Whitestone Insight	2 July	21	38	10	18	7
BMG	2 July	22	39	11	16	7
Techne	2 July	21	40	11	16	6
Redfield & Wilton	2 July	22	41	10	16	6
YouGov	2 July	22	39	12	15	7
Verian	1 July	21	36	13	16	7
More in Common	1 July	24	39	12	15	5
Focaldata	1 July	23	40	12	16	5
Stonehaven	30 June	24	39	12	15	5
Average		21.4	38.6	11.3	16.3	6.5

Labour's lead ran from 13 points (Norstat) to 20 (Opinium and People Polling). The seven MRP seat projections from BPC members averaged 441 Labour seats and 101 Conservative.

What happened

In Great Britain Labour took 34.7%, the Conservatives 24.4%, Reform UK 14.7%, the Liberal Democrats 12.5% and the Greens 6.9%. Labour won 412 seats, the Conservatives 121, the Liberal Democrats 72, the SNP 9 and Reform UK 5.

Against the poll average, Labour was overstated by 3.9 points and the Conservatives understated by 3.0. Reform UK was 1.6 points too high, the Liberal Democrats 1.2 too low and the Greens 0.4 too low. The BPC noted that for every party other than the big two the average error was within 2 points. The MRP average overstated Labour by 29 seats and understated the Conservatives by 20. Because the winner was never in doubt, the miss drew less attention than 2015, but on the size of the error on the lead it was larger.

Why it happened: what the pollsters found

The BPC did not commission an independent inquiry. Instead, from September 2024 it collected its members' own reviews on a dedicated page, and with the Market Research Society it held a public event, "Polling Lessons from 2024", on 6 May 2025. Its summary of that event lists the open questions: weighting by education, age (especially the over-75s) and ethnicity; the handling of don't knows and refusals; and falling response rates causing differential non-response. Pollsters were waiting for the British Election Study's random-sample post-election survey to test the last of these.

The firms' reviews do not agree on the main cause:

Ipsos (March 2025) understated the Conservatives by 5 points. Refusals were the key factor: 11% of likely voters would not say how they would vote, against 4% in 2019, rising to 18% among over-65s and among those with no qualifications. Re-weighting its over-65s to match its own large post-election survey moved the Conservatives from 19% to 22% and Labour from 37% to 35%. Weighting by past vote would have helped Labour's figure only a little.
Survation (24 March 2025) found no fault in interviewer conduct, weighting or treatment of undecideds. When it re-contacted respondents from its final phone poll, their Labour lead fell to 12.1 points, close to the actual UK margin of 10.1. It concluded that voters, especially those who had said Labour, changed their minds after fieldwork ended.
Verian, which polled a random-probability panel recruited from Royal Mail addresses, re-contacted 1,482 of its final-poll respondents and found no late swing specifically away from Labour. Its own error came from fixing turnout at 31.5 million, which was too high; correcting that and adding a leadership question to its imputation of don't knows would have given Labour 34% and the Conservatives 22%.
More in Common (31 August 2024) put about a point of its 4-point Labour overstatement down to late decisions and about a point to unrepresentative Muslim respondents, whose swing away from Labour in some seats its polls did not pick up. It suspected the rest lay in the kind of people who join online panels.
Focaldata (14 April 2025) concluded that not weighting by past vote had been a mistake, and that highly engaged respondents, more likely to back Labour, were over-represented.
Electoral Calculus and Find Out Now ran a recall poll of 18,616 people a week after the election. It still showed a 19-point Labour lead. Only 42% of 2019 Conservatives in their sample stayed loyal, against about 48% in reality, which pointed to loyal Conservatives missing from the panel rather than to late swing.

The common thread is that the polls found too few Conservatives, especially older ones, and too many engaged Labour voters. Late swing explains part of it for some firms and none of it for others.

What our data shows

The error sat in England and Wales, not Scotland. Our England figures from eight pollsters' final polls had Labour at 39.2% against 34.5% on the day, 4.7 points too high, and the Conservatives at 22.7% against 25.9%. The six Scotland-only polls in the final four weeks had Labour at 34.7% against 35.3%: no overstatement at all. The SNP was 2.4 points too high there.

As in 2015, and unlike 2017, the problem was in the raw samples. For six pollsters we hold the weighted table of everyone asked, before turnout filters and don't-know handling, as well as the published England figure. England produced 133 Labour voters for every 100 Conservatives. The raw tables averaged 186; the published figures, 173. The adjustments helped, most of all at JL Partners (199 down to 161) and Number Cruncher Politics (191 to 162), but they started from samples that had far too few Conservatives.

In our long-view table, all 20 final polls overstated Labour's lead, and the average error of 7.5 points is the second largest since 1945, behind only 1992. Two elections in the last four, 2015 and 2024, now show the same thing: the samples, not the adjustments, short of Conservative voters.

What the pollsters changed

Changes so far are firm by firm:

Ipsos has introduced a new voting intention method and moved from telephone interviewing to an online panel recruited offline; Scot Goes Pop reported the switch in Ipsos's Scottish poll with fieldwork of 12–18 June 2025.
Focaldata now weights by recalled past vote and by political interest, using targets from the British Election Study adjusted for that survey's own non-response. It has also aligned its MRP and standard questions, with explicit options for independent candidates.
More in Common published the micro-data from its final poll, said it would release micro-data every two months, and said it was likely to add a weight for country of birth interlocked with ethnicity, or for political engagement, once the BES post-election data are out.
Verian has built a new turnout estimate for its random-probability panel.
Survation found no methodological fault, put its error down to real late movement, and said it was cautious about changes that might not carry over to other elections.

The BPC event added recommendations on standard guidelines for reporting MRP, better handling of late deciders, and a mix of methods.

What it means for our forecast

2024 is the largest of the four misses our UK correction is fitted on. The correction starts from raw answers in England and takes off half the average past miss, party by party, kept only because it would have improved forecasts of each past election made without knowing the result. The raw answers understated the Conservatives in England at all four elections, by between about 2 and 7 points, so the correction adds to their share; 2024 adds to the case for it. Scotland and Wales are left as polled; 2024's Scottish polls show why.

The caveat is the one 2017 taught: the Conservatives were in government at all four elections. If the people who refuse to answer polls change, the lean can change with them, so the ranges on our forecast keep the full spread of past misses. The method is in Measuring polling error, and each firm's record is in our pollster ratings.

Sources

How to cite

Polling Forecast (2026). Right winner, wrong size: how every poll overstated Labour's lead in 2024. Retrieved 2 November 2026 from https://www.pollingforecast.com/research/missed-uk-2024. Live figures change as new polls arrive; the date above is when the text was last revised.