MRPs and mega-polls: what they measure, how they have done, and how we treat them

Summary

An MRP turns one very large poll into a call for every constituency. Since the 2024 election these projections, often sold as mega-polls, have become the most quoted numbers in British political coverage, because they give the one figure a reader wants: seats. This paper sets out what an MRP measures, how the published ones have done at the last three general elections, how the 29 published since July 2024 compare with our own projection, and how we use them.

The short version is this. An MRP's seat count is two things stacked together: a poll, and a model that spreads that poll across 632 seats. The poll carries the same house lean as any other poll from the same firm; a sample of tens of thousands shrinks random error, not that lean. At the 2024 election every final MRP gave Labour more seats than it won, from 425 to 516 against 411, because the polls underneath overstated Labour by about four points and the seat count magnified it. We therefore keep stand-alone MRP releases out of our polling average, count the MRP-based headline figures that some firms now publish as ordinary polls, and use the published seat projections as a check on our own seat model rather than an input to it. When an MRP differs sharply from our projection, the vote shares underneath usually explain most of the gap. The October 2026 mega-poll is a good example: its vote shares, run through our model, give Labour 358 seats, close to its own 353.

What an MRP is

MRP stands for multilevel regression and post-stratification. It works in two steps.

The first is the regression. A pollster interviews a large sample, typically 5,000 to 20,000 people, and fits a model of how each respondent says they will vote, using their age, sex, education, past vote and other characteristics, together with facts about their constituency: its 2024 result, its region, the share of graduates or retired people living there. The "multilevel" part lets the model borrow strength across seats, so that a constituency with only a handful of respondents is estimated mostly from similar seats elsewhere.

The second is the post-stratification. The pollster takes census counts of how many people of each type live in each constituency, applies the fitted model to each type, and adds them up. The result is an estimated vote share for every party in every seat, and so a winner in every seat.

The method is sound and well tested. Lauderdale, Bailey, Blumenau and Rivers (2020) describe it for the United States and the United Kingdom, including YouGov's constituency estimates at the 2017 election. Its strength is the shape of an election: which kinds of seat move, where a party's vote is efficient, where a three-way contest makes the result fragile. A uniform-swing model like ours adds one national change to every seat and can see none of that.

A mega-poll is not always an MRP

"Mega-poll" is a newspaper word. It usually means a large online survey run through an MRP, but not always. Some releases are a single very large poll with the seats worked out by a simpler swing model; some combine several weeks of a firm's regular polling; some are commissioned by a party or a campaign group and published selectively. The size of the sample is what makes the headline, but size does different things in each case.

Sample size matters less than it seems. A sample of 6,699, the size of the October 2026 mega-poll, averages about 11 respondents per constituency in Great Britain. Even 16,758, YouGov's sample in September 2026, is about 27 per seat. No seat is measured directly. Every seat estimate leans on the model's view of what voters like those in that seat do across the country, which is the point of the method and also the reason a seat-level call should be read as an estimate from the national poll, not a local poll.

How they have done

2017

YouGov's MRP was the first to be published widely in Britain. Its final estimate called a hung parliament, at a time when most forecasts had a clear Conservative majority, and put the Conservatives about 15 seats below the 317 they won. YouGov reported that it called about 93 per cent of seats correctly. This was the result that made MRP's name.

2019

YouGov's final MRP gave the Conservatives 339 seats, a majority of 28, and Labour 231. The Conservatives won 365 and Labour 202. The direction was right and the size of the majority was well short, because the poll underneath had the Conservative lead too narrow.

2024

2024 was the first election with many MRPs, and the first test of them as a group. Every one of them overstated Labour.

Final MRP	Labour	Conservative	Liberal Democrat	Reform UK
YouGov	425	98	72	3
More in Common	429	125	51	2
Focaldata	444	108	57	2
Electoral Calculus	449	60	71	18
Ipsos	453	115	38	3
We Think	465	76	52	3
Survation	470	68	59	15
Savanta	516	53	50	0
Result	411	121	72	5

Figures are each firm's final release where we hold it, otherwise the seat tallies compiled by Peter Inglesby, which count the seats where each party leads and can differ by a few from a firm's headline.¹ The British Polling Council put the average of the final MRPs at 441 Labour seats against 412 (including the Speaker's seat) and 101 Conservative against 121.

The seat-by-seat record was weaker than the totals suggest. Martin Rosenbaum compared each final model with the result in every seat. The best, YouGov's, called 53 seats wrongly; the worst called 134 wrongly; and in 245 seats at least one of the eight models picked the wrong winner.²

The cause was mostly the polls, not the models. The final polls overstated Labour by about four points in Great Britain, 38.6 per cent on the British Polling Council's average against 34.7 per cent in the result, and understated the Conservatives by three. Under first past the post a few points move scores of seats, and with Labour ahead in hundreds of seats by margins of five to fifteen points, the error in the vote shares came out larger in the seat count. The models were working as designed on inputs that were wrong in one direction.

What the record says

Three elections are a small record, but they point the same way. MRPs have been good at the shape of an election: they called the hung parliament of 2017, and they picked up Labour's efficient vote in 2024. Their seat totals have carried the error of the polls underneath, and at a landslide that error grows. A large sample did not protect them from it, because a large sample shrinks random error, not the bias that every poll from one firm shares.

This parliament

We hold 29 seat projections published for the next general election between December 2024 and October 2026, from nine firms or partnerships, collected from each firm's release where linked and from Wikipedia's table of seat projections. Their range is very wide. Labour's figure runs from 30 seats (Electoral Calculus and Find Out Now for PLMR, December 2025) to 353 (the same partnership, October 2026). Reform UK's runs from 71 to 398.

Most of that range is real movement in the polls. Reform UK led most polls through 2025 and into 2026, and on those vote shares every model gave it a large plurality. As Labour recovered in the summer of 2026 the projections followed. The projections from different firms in the same weeks also differ, sometimes by 50 seats or more for one party, and when they do the vote shares underneath usually differ too.

Our page MRPs and mega-polls plots every one of them against our own projection on the same day, and lists them in a table beside our figures for that date. It is updated when a new one is published.

The October 2026 mega-poll, taken apart

On 16 October 2026 the Daily Mail reported an Electoral Calculus and Find Out Now MRP for the public affairs firm PLMR, with fieldwork from 25 September to 2 October and 6,699 respondents. It gave Labour 353 seats, a majority of 56, with the Conservatives on 90, Reform UK on 81 and the Liberal Democrats on 70. Our projection on the same day had Labour on about 307 seats, short of a majority. The release also reported that a third of Labour voters would vote Conservative to keep Reform out, which invites the reading that the difference is tactical voting.

We tested that by running the poll's own vote shares (Labour 28, Reform UK 21, Conservative 18, Liberal Democrat 11, Green 8, Restore Britain 8) through our seat model, with no tactical voting at all. It gives Labour 358 seats, the Conservatives 97, Reform UK 46 and the Liberal Democrats 78. On Labour our plain uniform swing lands within five seats of the MRP.

So the difference from our projection is in the vote shares. Labour's share in the mega-poll, 28 per cent, is almost the same as in our average of every poll, 27.4. The Conservative share is not: 18 against 21.5. And Restore Britain is on 8 per cent, against about 3 in the average. Moving three and a half points back from Restore Britain to the Conservatives, and changing nothing else, takes Labour from 358 to 330 seats and the Conservatives from 97 to 145. With Labour on 28 per cent, Labour's majority depends on how the vote on the right is split, and this poll had it split much more evenly than other polls do.

Tactical voting matters at the margin, and in the opposite place from the headline. On the mega-poll's shares our model gives Reform UK 46 seats; the MRP gives it 81. If we add the anti-Reform switching the release describes, a third of the other parties' voters moving to whoever is best placed to beat Reform, our model gives Reform 10 seats and Labour 377. The MRP is therefore kinder to Reform than a uniform swing on the same vote, not harsher. Whatever tactical voting its model assumes, it is not what produces the Labour majority.

The test has limits. The vote shares we used are the rounded figures in the newspaper report, our model treats Restore Britain as a new party with no seat-level history, and a uniform swing cannot reproduce an MRP's local detail. But the result is large enough to survive those limits: the mega-poll's Labour majority comes from its vote shares, and the question to ask of it is whether the right is really that evenly split.

How we treat MRPs

As polls. An MRP's vote shares are a poll from that firm, with that firm's methods and house lean. Our pollster records are built from each firm's final polls; we treat its MRP vote shares as no better and no worse than its other polls.

Not in the average, with one exception. We keep stand-alone MRP releases out of our polling average. In our poll records they are marked as model output rather than standard voting intention, with the reason given. One release of 16,000 interviews, weighted by sample size, would outweigh every other poll for weeks, and its seat model plays no part in its vote shares in any case. The exception is a firm whose regular headline figure is itself produced by MRP: YouGov's voting intention has been produced through its MRP model since 2025, and we count that headline like any other firm's.

As a check on our seats. Our seat model is uniform swing, chosen because it beat the alternatives on a test of 631 seats from 2019 to 2024 (From votes to seats: swing models and their test). It is the part of our forecast we trust least, because it cannot see tactical voting or uneven swing. Published MRPs are the best outside check on it. When an MRP and our model disagree on the same vote shares, as they do on Reform UK's seats in October 2026, it tells us where our model is most likely to be wrong.

As something to try. The tactical voting section of our prediction page lets a reader add the kind of switching some MRPs assume, against Reform UK, against the Conservatives, or a pact on the right, on any vote shares. We tested it on the 2024 election: with two thirds of anti-Conservative voters switching where it mattered, uniform swing called 544 of 631 seats correctly, against 479 with no tactical voting, and Labour's total moved from 329 to 402, against 411 won. That is a large improvement on one election. We have not yet tested it on 2017 and 2019, and until we have, it stays a reader's tool and not part of our headline projection.

What would change our approach

Two results would. If we built our own MRP from respondent-level data, and it beat uniform swing on the same seat test at more than one election, it would replace uniform swing, under the same rule that applies to any seat model. We do not have the data for that today: we work from published tables, not individual responses. And if the published MRPs, given the same vote shares as the ordinary polls, called more seats correctly than our seat model at the next election, we would weight their seat patterns in our own projection. The comparison on MRPs and mega-polls is the start of that record.

References

British Polling Council (2024). MRP and seat projection overview, 2024 general election. britishpollingcouncil.org.

British Polling Council (2025). Polling lessons from 2024. britishpollingcouncil.org.

Inglesby, P. (2024). Seat tallies of the final MRP projections, 2024 general election (online).

Lauderdale, B. E., Bailey, D., Blumenau, J. and Rivers, D. (2020). Model-based pre-election polling for national and sub-national outcomes in the US and UK. International Journal of Forecasting, 36(2), 399–413.

Rosenbaum, M. (2024). Seat-by-seat comparison of the final MRP projections with the 2024 result. rosenbaum.org.uk.

Wikipedia. Opinion polling for the next United Kingdom general election: seat projections (accessed 16 October 2026).

How to cite

Polling Forecast (2026). MRPs and mega-polls: what they measure, how they have done, and how we treat them. Retrieved 17 October 2026 from https://www.pollingforecast.com/research/mrp-mega-polls. Live figures change as new polls arrive; the date above is when the text was last revised.