Research · When the polls missed

The hung parliament that never came: unrepresentative samples and the 2015 polling miss

Abstract. Every BPC member's final poll had the Conservatives and Labour within a point or two of each other, and the Conservatives won Great Britain by 6.6 points and a majority. The Sturgis inquiry found the samples held too many Labour supporters and too many politically engaged people, and ruled out shy Tories and late swing as major causes.

Polling Forecast · 2 November 2026 · 7 min read · United Kingdom, 2015 general election, polling error, final polls, inquiry, pollster methods

Summary

The final polls of the nine British Polling Council members had the Conservatives and Labour level: on average 33.7% to 33.6%. David Cameron's Conservatives won Great Britain by 6.6 points, 37.8% to 31.2%, and an overall majority, so the polls missed the lead by 6.5 points. The independent inquiry chaired by Patrick Sturgis concluded in March 2016 that the main cause was unrepresentative samples: the polls had too many Labour supporters and too few Conservatives, and weighting did not fix it. Late swing, shy voters and turnout misreporting made at most a small contribution.

What the polls said

All campaign long the polls showed a tie. The inquiry counts 91 polls in the short campaign; only 19 put the gap between the two big parties above 2 points, and zero was the most common lead. The final round, from the inquiry's Table 1:

Pollster	Fieldwork end	Con	Lab	LD	UKIP	Grn
Populus	6 May	34	34	9	13	5
Ipsos MORI	6 May	36	35	8	11	5
YouGov	6 May	34	34	10	12	4
ComRes	6 May	35	34	9	12	4
Survation	6 May	31	31	10	16	5
ICM	6 May	34	35	9	11	4
Panelbase	6 May	31	33	8	16	5
Opinium	5 May	35	34	8	12	6
TNS UK	4 May	33	32	8	14	6
Lord Ashcroft*	6 May	33	33	10	11	6
BMG*	5 May	34	34	10	12	4
SurveyMonkey*	6 May	34	28	7	13	8
BPC members' average		33.7	33.6	8.8	13.0	4.9

* Not BPC members in May 2015.

No published poll put the Conservatives more than a point ahead, except SurveyMonkey's poll for the Washington Post, which had the lead right (6 points) but both big parties too low. Survation ran a telephone poll on 6 May showing Conservatives 37, Labour 31, and did not publish it until after the count.

What happened

In Great Britain the Conservatives took 37.8%, Labour 31.2%, UKIP 12.9%, the Liberal Democrats 8.1% and the Greens 3.8%. The Conservatives won 330 seats, Labour 232, the SNP 56 and the Liberal Democrats 8. The exit poll at 10pm, projecting 316 Conservative seats and 239 Labour, was the first sign that the campaign polls had been wrong.

Against the BPC members' average, the Conservatives were 4.1 points too low and Labour 2.4 too high. The smaller parties were close: the inquiry puts the mean absolute error at 1.0 points for the Liberal Democrats and 1.4 for UKIP and the Greens. In Scotland the three final-week polls overstated Labour by 2.4 points and understated the SNP by 2.7. The miss was concentrated in the gap between the two parties that decided who would govern.

Why it happened: what the pollsters found

The BPC and the Market Research Society announced an independent inquiry about twelve hours after the polls closed. Its nine members, chaired by Professor Patrick Sturgis of the University of Southampton, had no role in the 2015 polls. They received micro-data from all nine BPC members and the re-contact surveys that six of them ran after the election. The report was published in March 2016. It tested eight explanations:

Analysis errors. The panel reproduced every published final estimate from the raw data and confirmed the weighting targets were correct. Unlike 1992, there were no wrong population totals.
Postal, overseas and unregistered voters. No detectable effect. Postal voters were 21% of the weighted samples against 20.9% of ballots.
Late swing. At most about 1 point on the lead. In re-contact surveys the raw average shift to the Conservatives was 1.8 points, but only 0.6 weighted by sample size, and below zero using final-poll respondents alone. The panel put Survation's unpublished 37–31 poll down to sampling variation and a fieldwork window of 3pm to 9pm on one day.
Deliberate misreporting ("shy Tories"). Rejected: it could not be reconciled with the re-contact surveys or the two random-sample surveys after the election.
Turnout weighting and differential turnout misreporting. At most a very small contribution.
Mode. Online and phone final polls were equally wrong, though phone polls had shown the Conservatives 1 to 2 points higher during the parliament.
Herding. The spread of estimates shrank in the final week more than sampling would explain, but the panel found no evidence of collusion.
Unrepresentative samples. The primary cause.

The evidence for the last point came from the British Election Study and British Social Attitudes surveys, which used random samples and face-to-face interviews after the election. They showed Conservative leads of 7 and 6 points, close to the result. The polls' samples differed from them in specific ways. Within the 65-plus band, the online polls had too many people under 70 and too few over 75, and no respondents at all aged 90 or over. They had too many politically engaged people: almost every poll overstated how many respondents had voted in 2010, most of all among 18 to 24-year-olds. They also understated the Conservative lead among 45 to 64-year-olds. YouGov's own review, summarised by the BPC in 2017, reached the same conclusion: its sample had too many people who were very interested in politics, especially among the young.

The report made twelve recommendations. Pollsters should ask whether people had already voted by post, review turnout probabilities that relied on a 1-to-10 self-rating, review the handling of don't knows, get more representative samples within weighting cells, and look for new quota and weighting variables tied both to vote choice and to the chance of taking part in a poll. BPC rules should require members to publish their weighting variables and targets, flag changes in method, release micro-data on request and give confidence intervals.

What our data shows

Our England figures inside the GB polls tell the same story. Across seven pollsters' final polls, the Conservatives averaged 36.4% in England against 40.9% on the day, an error of 4.6 points, and Labour 34.5% against 31.7%, an error of 2.9.

The key question is where in the process the error entered. For two pollsters, YouGov and Survation, we hold both the weighted table of everyone asked (before turnout filters and don't-know handling) and the published headline. On the day England had 129 Conservative voters for every 100 Labour voters. The two raw tables had 103 on average, and the headlines 106. The turnout and don't-know steps moved the numbers the right way, but by almost nothing: the error was already in the sample. That matches the inquiry's finding and is the opposite of what happened two years later.

In our long-view table, every one of the twelve final polls in 2015 understated the Conservative lead. The inquiry ranks 2015 alongside 1970 as the second-worst performance on the lead since 1945, behind 1992. What made it worse in public perception is that no poll was close: the best final poll still missed the Conservative and Labour shares by an average of 2.8 points, something that had otherwise happened only in 1992.

What the pollsters changed

Before the 2017 snap election the BPC published each member's changes since 2015. The common themes were turnout and political balance:

ComRes replaced self-rated likelihood to vote with a turnout model based on age and social grade, built by regression on ward-level turnout. Applied to its 2015 polls it would have produced a 5-point Conservative lead. It moved all polling online.
ICM moved from telephone to an online panel, doubled samples to 2,000, added past vote to its quotas, weighted turnout by age and social grade, weighted interest in politics to the 2015 BES face-to-face survey and weighted fully to the 2015 result.
Ipsos MORI weighted by newspaper readership, added education to quotas and weights in 2016, and tightened its turnout filter to people who rated themselves 9 or 10 out of 10 and did not say they never, rarely or only sometimes vote.
Kantar Public (formerly TNS) built a new likely-voter model from age, stated intention and past voting, weighted correctly for the over-70s, added education, and asked whether people had already voted by post.
YouGov recruited more low-interest, young, less-qualified panellists and non-voters, moved its top age band to 65-plus, added education by age, swapped party ID for past vote by region, and weighted to attention to politics using BES targets.
Opinium counted only people certain to vote and weighted strictly to the 2015 result.

What it means for our forecast

2015 is one of the four elections our UK correction is fitted on. Because the error was in the samples rather than the adjustments, we start from each pollster's raw answers in England and take off half the average past miss, party by party. We keep that rule only because, tested on each election without knowing its result, it beat the published polls. Scotland and Wales are left as polled, since their records show no steady lean.

The 2015 lesson behind that design is that a consensus among pollsters is not evidence. Every firm can share the same blind spot, which is why the ranges on our forecast use the full spread of past misses rather than the agreement between this year's polls. The details are in Measuring polling error, and each firm's record is in our pollster ratings.

Sources

How to cite

Polling Forecast (2026). The hung parliament that never came: unrepresentative samples and the 2015 polling miss. Retrieved 2 November 2026 from https://www.pollingforecast.com/research/missed-uk-2015. Live figures change as new polls arrive; the date above is when the text was last revised.