How the forecast works
In short: we average the polls for England, Scotland and Wales separately, measure how far polls have missed before, carry those misses into ranges and "what if" lenses, and move every seat's 2024 result by the change in its nation.
This page is the short version. How we correct the polls, with today's figures, is on its own page; the full method, with the maths, the tests and what we rejected, is in the research papers.
1. Polling averages by nation
Great Britain polls mix three different party systems. So we average each nation on its own terms:
- England: the England figure inside each GB poll, taken from the pollster's own tables. Where there is no England column we rebuild it from the English regions, and record how. Only polls graded A on the headline basis count.
- Scotland and Wales: Scottish and Welsh Westminster polls. Where there are none recent, the 2024 result moved by the GB change.
- Great Britain: the three nations combined, weighted by votes cast in 2024, then checked against the GB headline average.
Each average uses the latest poll from each pollster in the last 28 days (120 days for Scotland and Wales, where polls are rarer), with equal weights, so frequent pollsters do not dominate.
2. How wrong polls have been
England is where we measure polling error: the SNP and Plaid Cymru do not stand there, and it holds 543 of the 632 seats in Great Britain. For 2015, 2017, 2019 and 2024 we compared each pollster's final poll with the England result. See poll accuracy.
Scotland and Wales have their own records, from Scotland-only and Wales-only final polls at the same four elections (see Scotland and Wales). Their polls miss by about as much as England's, but not the same way twice, so correcting them for their average past miss didn't help when tested.
We tested whether adjusting the pollsters' published figures for their past error would have helped, predicting each election only from the ones before it. It did not help reliably. Correcting what people actually told the pollsters (their raw answers) did, so that is our headline: our projection. The full evidence is in Correcting the raw answers and Why we start from the raw answers. The size of past errors sets the ranges:
| Party | 80% range (±, points) |
|---|---|
| Conservative | 4.4 |
| Green | 0.8 |
| Labour | 3.9 |
| Lib Dem | 1.4 |
| Others | 1.0 |
| Reform UK | 2.0 |
3. Lenses: what if the polls are wrong?
A lens is a way of reading today's polls. The headline is our projection: the pollsters' raw answers, corrected by half their average past miss (our method), because that beat the published figures in 2015–2024 tests. It falls back to "polls as they stand" (the published figures, averaged) only when fewer than two pollsters' raw tables from the last four weeks are available. The others apply a past error to today's numbers: the average miss since 2015, or exactly what happened in 2015, 2017, 2019 or 2024. In Scotland and Wales the "average miss" and "as wrong as in…" lenses use each nation's own past error, including for the SNP and Plaid Cymru. Lenses built from England-only evidence (our reading, pollster records) leave Scotland and Wales as polled.
Three more lenses look at pollsters rather than elections:
- Each pollster corrected by its own record. Every pollster in today's England average has its own average final-poll miss (primary benchmark polls only) taken off. Pollsters with no checked record get the industry average miss instead. Records are short, one to four elections each, so this is an indication, not a correction we trust.
- Only the pollsters best for Labour / the Conservatives / Reform UK. The average is rebuilt from the half of today's pollsters showing that party highest. This shows how much the forecast depends on which pollsters you believe. A pollster being good for a party is not evidence that it is right.
Because every pollster is weighted equally in the average, each of these is exactly an adjustment to the England average, so it runs through the same seat model as every other lens.
Seats at risk from polling error
For every seat we record the winner under each past-miss lens (the average miss, and the 2015, 2017, 2019 and 2024 misses). A seat is "at risk" if any of them gives a different winner from our projection. The map's "Risk from polling error" colouring and the list on the forecast page are built from this.
Your own polling error
The what-if tool lets you set the error for each party yourself (up to 8 points either way) and runs it through the same seat model. It shows the central projection only, not the simulation.
4. From votes to seats
Each seat starts from its 2024 result on today's boundaries. Its parties move by their change in the seat's nation. We tested three simple ways of doing this on the last election, starting from the 2019 result and giving each method the true 2024 vote shares:
| Seat model | Seats called correctly |
|---|---|
| Uniform swing (current default) | 479 of 631 (75.9%) |
| Blend of uniform and proportional | 461 (73.1%) |
| Proportional swing | 445 (70.5%) |
All three under-called Labour in 2024 (329 or fewer against 411 won), because they cannot see tactical voting or where each party's vote rose most. A model using each seat's demographics and tactical voting is being built and will replace uniform swing only if it beats it on this test.
By-elections. Where a seat has had a by-election since 2024, its starting point moves part of the way towards that result. We first take out the national change on polling day (from our own poll average that day), so only what was special to the seat is left, and carry half of that into the projection. A party that gained the seat gets 3 more points for its new MP. Half is what the evidence shows: at the 33 by-elections in Great Britain from 2015 to 2024, parties kept about half of their by-election lead over the polls at the next general election, and winners who had gained the seat ran a few points further ahead. Most by-election gains were held (North Shropshire, Tiverton and Honiton, Wakefield), some went back (Hartlepool, Mid Bedfordshire). A by-election where the Conservatives or Labour did not stand (Clacton, 2026) is not used for vote shares. In every case the seat is counted as held by the party of its current MP.
Restore Britain had no vote in 2024, so we spread its national share using Reform UK's 2024 vote in each seat: stronger where Reform was strong, weaker where it was weak, with the national total unchanged. How steep to make this is set from the two Westminster contests Restore has fought: Makerfield (6.8%) and Holborn and St Pancras (2.5%). Northern Ireland is not modelled and is shown at its 2024 result; the Speaker's seat stays with the Speaker.
5. Ranges and chances
The chances come from 1,000 simulations. Each picks one real past election's error pattern (2015, 2017, 2019 or 2024: England's in England, and Scotland's and Wales's own from the same election), scales it up or down at random around zero, adds a little national and seat-level noise, and re-runs every seat. Using real error patterns means parties move together the way they have in practice: when the Conservatives are understated, Labour has usually been overstated.
6. Election day, not just today
Everything above is "if the election were held today". The polling-day outlook carries today's polls forward to a polling day you choose (by default the latest possible, 15 August 2029). From Mark Pack's PollBase we found that, months or years out, the Conservatives, Labour and the Lib Dems finished much closer to their previous result than their polls suggested at the general elections from 2001 to 2024 (the model uses no election before 2000; 1987 to 1997 are in the data but not used): the best predictor takes only about three-fifths of their change since the last election at face value, rising towards all of it in the final weeks. The challengers (UKIP, the Brexit Party, Reform, the Greens) showed no such pull back and are carried forward as polled. Predicting each election from the others, this beat the polls as they stood at most distances. The simulation then adds the extra uncertainty of the time left.
Model versions and runs
| Run | Date | Model | Status | Polls |
|---|---|---|---|---|
| 1 | 24 Sep 2026 | PollModel 1.1 | Live | 8 |
Figures as of . Published by Polling Forecast.




