Research · When the polls missed

Fixing 2015 too hard: turnout models and the majority that vanished in 2017

Abstract. The final polls had the Conservatives 8 points ahead and they won by 2.5, losing their majority; the Conservative share was right and Labour's 5 points too low. Raw poll data were close to the result, and the turnout models pollsters had added after 2015 pushed the headlines towards the Conservatives, while Survation and YouGov's MRP model came close.

Polling Forecast · 2 November 2026 · 7 min read · United Kingdom, 2017 general election, polling error, final polls, inquiry, pollster methods

Summary

The nine British Polling Council members' final polls averaged Conservatives 43.7%, Labour 35.8%: a lead of 7.9 points, enough for a comfortable majority. Theresa May's Conservatives won Great Britain by 2.5 points, 43.5% to 41.0%, and lost their majority, so the polls missed the lead by 5.4 points. They had the Conservative share almost exactly right and Labour's 5.2 points too low. The cause was the turnout adjustments pollsters had added after 2015: their raw data were close to the result, and their models of who would vote pushed the figures towards the Conservatives.

What the polls said

The final round ranged from a 1-point to a 13-point Conservative lead. The final BPC member polls, as published by the BPC the day after the election:

Pollster	Fieldwork end	Con	Lab	LD	UKIP	Grn
Opinium	4 June	43	36	8	5	2
Survation (phone)	7 June	41	40	8	2	2
Ipsos MORI (phone)	7 June	44	36	7	4	2
ICM	7 June	46	34	7	5	2
ComRes	7 June	44	34	9	5	2
YouGov	7 June	42	35	10	5	2
Panelbase	7 June	44	36	7	5	2
Kantar Public	7 June	43	38	7	4	2
BMG	7 June	46	33	8	5	3
Average		43.7	35.8	7.9	4.4	2.1

Outside the BPC, SurveyMonkey's final poll had Conservatives 42, Labour 38, and Qriously, which polled through mobile apps, had Labour ahead 41 to 39. YouGov's MRP model, built with the LSE's Ben Lauderdale and Jack Blumenau and first published in The Times on 31 May, projected a hung parliament with an expected 302 Conservative seats and 269 Labour. The pound fell half a per cent on the day it appeared.

What happened

In Great Britain the Conservatives took 43.5%, Labour 41.0%, the Liberal Democrats 7.6%, UKIP 1.9% and the Greens 1.6% (BPC figures, based on all seats except Kensington). The Conservatives won 317 seats, 13 fewer than in 2015 and nine short of a majority. Labour won 262 (up 30), the SNP 35 (down 21) and the Liberal Democrats 12.

Against the BPC average, the Conservatives were overstated by 0.2 points, Labour understated by 5.2, the Liberal Democrats overstated by 0.3 and UKIP by 2.5. The BPC noted this was only the second election since 1987 at which the polls had underestimated Labour. Survation, at a 1-point lead, missed the margin by 1.5 points and the YouGov MRP by 0.9. BMG and ICM, at 13 and 12 points, missed by about 10.

Why it happened: what the pollsters found

The BPC said on 9 June 2017 that "the final polls were not ideal" but that it saw no need for another formal inquiry, pointing to the Sturgis report as the reference for the issues involved. Instead each member that published a final poll was asked for a "lessons learned" report. They were presented at a BPC conference at the Royal Statistical Society on 20 February 2018, with talks from ComRes, Survation, YouGov, Kantar and Ipsos MORI and a session from Ben Lauderdale on MRP.

John Curtice, the BPC's president, set out the main finding in The Conversation on 15 June 2017. In 2015 the raw, unweighted numbers in the final polls had pointed to the same result as the headlines: the samples themselves had too many Labour voters. In 2017 the raw tables had Labour on average just one point behind the Conservatives, close to the actual result. The headline leads were produced by the adjustments, which, Curtice wrote, "proved to be a step too far".

The House of Lords Select Committee on Political Polling and Digital Media reached the same conclusion in its report The politics of polling (April 2018). Most companies used turnout models that assumed the pattern of 2015 would repeat. In 2017, voters under 50 were more likely to turn out than that, and much more likely to vote Labour. The committee gave two examples from Peter Kellner's analysis: Ipsos MORI's unadjusted data showed the two parties level and its turnout adjustment produced an 8-point Conservative lead; ICM's figures went from a 6-point lead before turnout weighting to 12 after it. Will Jennings, in evidence to the committee, put the mean absolute error on the final Conservative–Labour lead at 5.3 points.

The committee also recorded a caution from the British Election Study: there was no surge in youth turnout. The error was not that young people voted in record numbers but that the models assumed turnout gaps by age and class that were wider than the real ones.

The polls that did well had not leaned on these models. Survation's phone polls used the same approach as its unpublished final poll of 2015, and its published figures sat almost on top of its raw data (see below). YouGov's MRP model, according to Lauderdale and colleagues' 2019 account, estimated vote by constituency type from a large daily sample and missed the national margin by 0.9 points; its largest error on any party was Labour, underestimated by 2.8.

What our data shows

Our stage-by-stage records show exactly where the error came in. For four pollsters we hold both the weighted table of everyone asked and the published figures, broken down for England. On the day England had 108 Conservative voters for every 100 Labour voters. The four raw tables averaged 107. The four published England figures averaged 122.

Pollster	Raw table (Con per 100 Lab)	Published	Result
Ipsos MORI	106	133	108
ComRes	101	129	108
Opinium	116	120	108
Survation	107	106	108

Ipsos MORI and ComRes, the two firms whose adjustments moved the numbers most, turned near-correct raw data into the two widest misses in this group. Survation, which made almost no turnout adjustment, published close to its raw figures and close to the result. In England the four final polls averaged Labour 3.9 points too low and the Conservatives 0.6 too high.

The spread was also unusual. Across the 13 final polls in our long-view table, the error on the lead ran from BMG overstating the Conservative lead by 10.6 points to Qriously understating it by 4.4. In 2015 every firm missed the same way; in 2017 the firms disagreed, and the ones that adjusted most were furthest out. Scotland had its own miss: the Scotland-only polls overstated the SNP by 4.1 points and understated the Conservatives by 2.9.

What the pollsters changed

The BPC's November 2019 summary of changes since 2017 reads largely as a list of turnout models being taken out:

ComRes (by then Savanta ComRes) stopped weighting by the turnout model it used in 2017.
ICM stopped weighting to a pre-set level of turnout by demographic group, moved to a 0-to-10 likelihood scale, stopped imputing don't knows from past vote, and added 2016 referendum vote to 2017 vote as a weighting target.
Ipsos MORI stopped weighting to the 2015 BES turnout ratios by age and tenure, dropped children in household as a weight and added ethnicity.
BMG stopped weighting by stated probability of voting and by whether people voted last time, and moved to online only.
Kantar rebuilt its likelihood-to-vote model on data collected after 2017 and adjusted its past-vote weighting for people, especially 2017 Labour voters, who forget how they voted.
YouGov slightly reduced the downweighting of people who did not vote last time.
Survation polled by phone only, as in 2017.

Most firms also began asking people to choose only from the parties standing in their own constituency. MRP spread quickly: at the 2024 election seven BPC members published MRP seat projections.

What it means for our forecast

2017 shaped our UK correction more than any other election. We start from each pollster's raw answers in England, before their turnout and don't-know steps, because in 2017 those steps were the problem. And we take off half the average past miss rather than all of it: taking off the whole miss did better on average across 2015, 2017, 2019 and 2024, but did worse in 2017, when the usual direction reversed. A correction has to help at every election when each is predicted without knowing its result, not just on average.

2017 is also the reason the ranges around our forecast are wide in both directions. The polls' long lean towards Labour is real, but it can flip. The tests are set out in Measuring polling error, and each firm's record is in our pollster ratings.

Sources

How to cite

Polling Forecast (2026). Fixing 2015 too hard: turnout models and the majority that vanished in 2017. Retrieved 2 November 2026 from https://www.pollingforecast.com/research/missed-uk-2017. Live figures change as new polls arrive; the date above is when the text was last revised.