From today to polling day: how far British polls move before a general election
Abstract. Months or years before a British general election, the Conservatives, Labour and the Liberal Democrats have finished closer to their previous result than their polls suggested, while insurgent parties have not. Taking a little under two thirds of the big three's change since the last election at face value, and widening the uncertainty with the time left, beat the polls as they stood on average at every distance across the seven elections from 2001 to 2024 that the model uses.
What has changed since this paper was written
Updated 9 November 2026. The rest of the paper is as first published, except where a sentence below says it was corrected.
- How the forecast has moved (3 October). /uk/history shows the projection over time, with earlier weeks rebuilt on the polls that were public then.
- A time stamp (13 October). Every forecast page says when the forecast was last rebuilt and the date of the latest poll, in UK time.
The question: today is not polling day
Everything else on this site answers a conditional question: if the election were held today, what would happen? That is the question opinion polls themselves answer, and it is the one our measurements of polling error address, since those errors are measured on final polls taken in the last days of a campaign (see Measuring British polling error in England, 2015–2024). But the next general election is not today. At the time of writing it is almost three years away, and a reader who wants to know who will win it is asking a different question.
The distinction matters because voting intention moves. A poll taken in the middle of a parliament records how people say they would vote at that moment, with no campaign under way, no choice of prime minister in front of them and, for many, little thought given to the matter. The literature on the "timeline" of elections finds that polls become steadily more informative about the result as polling day approaches, and that early polls carry a large component of short-lived movement that has disappeared by the time votes are cast (Wlezien et al., 2013; Erikson and Wlezien, 2012). Gelman and King (1993) framed the same puzzle for American presidential elections: campaign polls swing about a great deal, yet the outcome is predictable from fundamentals known well in advance, because the campaign tends to bring voters back to positions that their circumstances would have predicted.
The original brief for this project drew the line sharply. Errors measured on final polls apply only to the "election today" estimate; a forecast for polling day needs its own model of how votes move before then, and the two must not be mixed. This paper describes that second model: the evidence it rests on, its form, how it was fitted and tested, how its uncertainty is set, and how it enters the seat simulation. The result is shown on the election-day forecast, which is kept separate from the "election today" headline on the home page.
The evidence: forty years of polls at fixed distances from polling day
Source
The evidence is every Great Britain voting-intention poll since 1983 in Mark Pack's PollBase (the Q2 2026 edition), a long-running compilation of published British polls that is free to use with credit and a link, both of which are given on the election-day page.1 For each poll we take the fieldwork date and the shares for the Conservatives, Labour, the Liberal Democrats (the Liberal/SDP Alliance before 1988), the UKIP–Brexit Party–Reform family and the Greens. In the 2017–19 and 2019–24 parliaments, when PollBase carries separate UKIP and Brexit Party columns, the two are added together to give the family's share.
Results come from two places. For 1983 to 2005 we use the general-election rows in PollBase; for 2010 to 2024 we use our own constituency results, aggregated to Great Britain. The 1983 result enters only as the previous result for 1987, so the elections in the data are the ten from 1987 to 2024. The model uses seven of them, 2001 to 2024. Elections before 2000 stay in the data and on the pages but count for nothing that shapes a forecast (the cutoff year is a setting on the admin page, 2000 by default), so 1987, 1992 and 1997 play no part in the curve, the test or the uncertainty below.
Horizon points
For each of those ten elections we average the polls at twelve fixed distances before polling day: 30, 61, 91, 182, 273, 365, 456, 548, 639, 730, 913 and 1,095 days, running from about one month to three years. Call the distance \(d\). The average at distance \(d\) uses every poll whose fieldwork date falls within 30 days either side of the date \(d\) days before the election (a 60-day window), with three restrictions:
- only polls taken more than 14 days after the previous general election count, so that a window reaching back into the last parliament is simply truncated;
- a window needs at least three polls, or that distance is skipped for that election;
- a party is averaged only if it appears in at least three polls and in at least a third of the polls in the window.
Each resulting record, a horizon point, holds the election year, the distance, the party, the poll average \(\bar{y}_{p,t}(d)\), the number of polls behind it, the party's share at the previous general election \(v_{p,t-1}\), and its share at this election \(v_{p,t}\). There are 439 such points (331 of them from 2001 on), stored in model.HorizonPoint by database script 12 and read by the site through web.vw_HorizonPoint. The number of polls in a window ranges from 6 to 174, reflecting how much more often Britain was polled after 2010.
Two gaps are structural. The short parliaments of 2015–17 and 2017–19 do not reach back three years, so of the seven elections the model uses there are six at 913 days and five at 1,095. And the insurgent parties enter only from 2015, the first election for which both the UKIP family and the Greens have a previous result and enough polling to average. That gives eight insurgent cases at each distance (UKIP/Brexit/Reform and the Greens at 2015, 2017, 2019 and 2024), fewer at the longest distances.
The prior
The obvious alternative to a poll taken years out is to assume nothing has changed since the last election. We call the previous result the prior for this reason: it is what one would predict with no polls at all. A forecast for polling day can then be framed as a choice of how far to move from the prior towards today's polls, and the data can tell us how far that has been worth moving at each distance.
What the data show
The live table in section 5 sets out the results at every distance, but three findings carry the model.
The big three drift back. For the Conservatives, Labour and the Liberal Democrats, the result has typically landed somewhere between the previous result and the mid-term polls. Parties that surged in mid-term polls gave back much of the surge; parties that slumped recovered much of the slump. The examples are familiar, some from the older part of the record: Labour in the 1987 parliament polled in the high thirties a year out and finished at 31.5; the Conservatives in 2014 polled around 32 a year before an election in which they took 37.7. Most striking is that from about six months out, the polls as they stood have missed the big three's result by about as much as simply repeating the previous result would have, and at several distances by more. This is not a claim that polls are useless at that range, only that they overstate how much has changed.
The insurgents do not. For the UKIP–Brexit–Reform family and the Greens, pulling the polled share back towards the previous result made predictions worse, not better. In the exploratory analysis at three, twelve and twenty-four months out, their polls as they stood missed by 4.1, 4.1 and 3.8 points on average, and the pulled-back version by 4.3, 4.4 and 4.6; at six months the two were level. With eight cases per distance this is thin evidence, but it points in one direction. It also makes sense: an insurgent's "previous result" is often a poor guide because the party was barely competing last time (UKIP took 3.2% in 2010 and 12.9% in 2015; Reform took 2.1% in 2019, as the Brexit Party, and 14.7% in 2024). Where it points the other way, as with UKIP's collapse from 12.9% to 1.9% in 2017, the collapse had causes the prior could not know about.
No separate governing-party effect. A common view is that governments recover in the run-up to an election beyond the general drift back. We tested for it by comparing the residuals of the governing party with those of the others after the drift-back model. The governing party's average residual was −0.4 points at six months, +1.0 at one year and −0.8 at two years (ten cases each, on the full 1987–2024 record). The sign changes with the distance, so we treat this as noise and add no governing-party term.
The model
Form
For the Conservatives, Labour and the Liberal Democrats, the projected polling-day share is a weighted average of today's polled share and the party's share at the previous election:
Here \(y_p\) is today's share for party \(p\), after whatever lens the reader has chosen (by default our projection; see Lenses: reading the same British polls several ways), \(v^{\,\mathrm{prior}}_{p}\) is the party's share at the 2024 election, \(d\) is the number of days from the latest poll to polling day, and \(w(d)\) is the weight given to today's polls. The second form of the equation is the more intuitive one: start from the last result and count only a fraction \(w(d)\) of the change since then.
The weight falls from 1 on polling day towards a floor as the distance grows:
where \(f\) is the floor, the share of the change since the last election that still counts however far out we are, and \(\tau\) is a time constant in days that sets how quickly the weight falls to the floor. The fitted values are \(f =\) 0.62 and \(\tau =\) 30 days. With these, the weight is about 0.76 at one month, 0.64 at three months and 0.62 at six months, and from about four months onwards it is indistinguishable from the floor. In words: a little under two thirds of the big three's change since the last election is taken at face value from roughly three months out, and a little more than that in the final weeks. On the full 1987–2024 record the fit was \(f = 0.52\) and \(\tau = 45\) days, a stronger pull back: the three elections before 2000 had larger mid-term swings that later reversed.
We chose this form for three reasons. It has the right limits (\(w = 1\) on polling day, since today's polls are then the best information there is; \(w \to f\) at long range). It has only two parameters, which matters with seven elections. And it matches the shape of the weights fitted at each distance separately, which (see the table in section 5) fall quickly over the first few months and then wander between about 0.5 and 0.7 with no clear trend.
Everyone else
For Reform UK and the Greens the projection is simply today's share: \(\hat{v}_p(d) = y_p\). The same applies to parties with no history to learn from (Restore Britain, Your Party) and to the SNP and Plaid Cymru, which the horizon data do not cover. After equation (1) has been applied to the big three, all shares in the nation, including Others, are rescaled to sum to 100. The rescaling matters more than it might seem, as the worked example in section 7 shows: when the big three are pulled back towards a 2024 total that is higher than their combined polling today, every other party's share is scaled down in proportion. "Taken as polled" means that no pull-back is applied to them, not that their final share is untouched.
Applied nation by nation
The curve is fitted on Great Britain polls and results, because that is what PollBase holds over forty years. The site applies it in each of England, Scotland and Wales separately, using that nation's own 2024 result as the prior. So Labour in Scotland is pulled back towards its 2024 Scottish share, not its GB share. This is an assumption: we have no long series of nation polls with which to test whether the drift back has the same strength in each nation. It is the natural assumption, but it is one.
Fitting and testing
Least squares on a grid
The two parameters are estimated by least squares over every horizon point for the big three, pooled across elections and distances. For each candidate pair \((f, \tau)\) the predicted share is computed from equation (1) and the squared error against the actual result summed:
The search runs over \(f\) from 0 to 1 in steps of 0.01 and \(\tau\) from 5 to 600 days in steps of 5. A grid is crude but transparent, and with two parameters it is cheap. The fit is recomputed each time the site starts, from the stored horizon points, so the numbers quoted here and on the site come from the same code (Elections.Core/Method/ElectionDayModel.cs).
Leave one election out
A curve fitted to all seven elections and then scored on the same seven would flatter itself. The test therefore follows the cross-validation logic of Stone (1974): the curve is refitted seven times, each time leaving out one election, and each left-out election is predicted using only the curve fitted to the other six. The comparison at each distance is between three predictions of the big three's shares: the polls as they stood, the previous result alone, and our method. The table reports the mean absolute error per party for each, the number of elections at which ours had the smaller average miss, and the root-mean-square error of our method, which is used in the next section as the measure of uncertainty.
| Days out | Elections | Weight fitted here | Weight on curve | Polls MAE | Last result MAE | Ours MAE | Ours better | Ours RMSE |
|---|---|---|---|---|---|---|---|---|
| 30 | 7 | 0.81 | 0.76 | 2.72 | 4.61 | 2.58 | 4 of 7 | 3.30 |
| 61 | 7 | 0.63 | 0.67 | 4.15 | 4.61 | 3.71 | 4 of 7 | 4.69 |
| 91 | 7 | 0.58 | 0.64 | 4.48 | 4.61 | 4.00 | 4 of 7 | 5.04 |
| 182 | 7 | 0.50 | 0.62 | 4.71 | 4.61 | 4.03 | 4 of 7 | 5.58 |
| 273 | 7 | 0.71 | 0.62 | 3.59 | 4.61 | 3.26 | 3 of 7 | 4.53 |
| 365 | 7 | 0.71 | 0.62 | 4.09 | 4.61 | 3.73 | 5 of 7 | 4.82 |
| 456 | 7 | 0.72 | 0.62 | 4.32 | 4.61 | 4.17 | 3 of 7 | 5.01 |
| 548 | 7 | 0.64 | 0.62 | 4.65 | 4.61 | 4.41 | 4 of 7 | 5.24 |
| 639 | 7 | 0.61 | 0.62 | 4.88 | 4.61 | 4.30 | 4 of 7 | 5.22 |
| 730 | 7 | 0.71 | 0.62 | 4.84 | 4.61 | 4.48 | 4 of 7 | 5.60 |
| 913 | 6 | 0.56 | 0.62 | 5.77 | 4.48 | 5.10 | 5 of 6 | 6.02 |
| 1095 | 5 | 0.54 | 0.62 | 5.40 | 4.47 | 5.00 | 3 of 5 | 6.56 |
Two further columns need explanation. "Weight fitted here" is the single weight that best fits that distance alone, estimated by regressing the result's change since the last election on the polls' change, through the origin, using all elections. It is descriptive: it shows the pattern the curve is smoothing. "Weight on curve" is equation (2) at that distance.
The method beat the polls as they stood on average at every distance, and in at least half the elections at ten of the twelve distances: at one year the average miss fell from 4.1 to 3.8 points, and at two years from 4.9 to 4.5 (the table shows the figures live). It also beat the previous result alone at every distance up to two years; at 30 and 36 months, with six and five elections, the previous result alone did better (4.5 against 5.1 and 5.0). Up to two years out the combination is better than either ingredient. The gains are smaller than on the full 1987–2024 record, where at one year the miss fell from 4.4 to 3.2 and at two years from 6.1 to 4.2.
This is not a large or a clean sample. Seven elections is seven elections, and some of the improvement at long range comes from a few elections with large mid-term swings (2010, 2017, 2024). The honest summary is that the drift back has been a regular feature of British elections for forty years, that since 2001 it has been weaker than before, that a two-parameter description of it has held up out of sample in three to five elections of seven (or six, or five) at each distance, and that there is no guarantee the next parliament behaves like the last seven.
Uncertainty that grows with the time left
Drift standard deviation
The projection in equation (1) is a central estimate. The uncertainty around it has two sources: the ordinary polling error that remains on polling day, and the movement of opinion between now and then. We call the second drift. Its size is taken from the leave-one-out test itself: the root-mean-square miss of our method at each distance (the last column of the table) is the typical error a forecast made that far out has had. It grows with distance, but more slowly than linearly, and a logarithmic fit describes it well:
for the Conservatives, Labour and the Liberal Democrats, where \(\sigma\) is one standard deviation in percentage points. For Reform UK and the Greens the same form is fitted to the root-mean-square miss of their polls as they stood (since those are the predictions we make for them):
Both are fitted by ordinary least squares on the twelve distances. The coefficients are estimated from the stored horizon points each time the site starts, so they change only if the horizon data are rebuilt; the values above are those at the time of writing.
The code adds three details to these formulae. First, distances under 30 days are evaluated at 30 and then scaled down linearly, \(\sigma(d) = (d/30)\,\sigma(30)\), so that the drift shrinks to nothing on polling day itself, when only ordinary polling error remains. Second, an insurgent's drift is never allowed to be smaller than a main party's at the same distance: the code uses \(\max(\sigma_{\mathrm{ins}}, \sigma_{\mathrm{main}})\). Third, the insurgent formula applies to Reform UK and the Greens only. Restore Britain, the SNP and Plaid Cymru, which are also carried forward as polled, receive the main-party drift of equation (3).2
The insurgent curve is fitted to few cases, eight at most distances and three at 1,095 days, and its root-mean-square values jump about (from about 2 points at one month to above 8 at six months, then between 5 and 6 out to two years, and 7.5 at three years on three cases). The logarithmic fit smooths these, but the uncertainty about the uncertainty is itself large.
Vote-share bands
On the forecast pages each GB party's 80% vote band is set from its final-poll error in England, \(\mathrm{rms}_p\) (the root mean square of its average published miss at each election, 2015–2024, the last row of the table of headline errors in Measuring British polling error in England, 2015–2024), and the drift, combined as independent sources:
where 1.2816 is the 90th percentile of the standard normal distribution. For the "election today" forecast \(\sigma_p = 0\) and the band reduces to the final-poll error alone. The SNP's and Plaid Cymru's bands come from their own nation's final-poll record (see Scotland and Wales: sparse polls and no steady lean) and, as the code stands, do not include drift.
A worked example at 400 days
Take a forecast made 400 days before polling day, and for simplicity work at GB level with round illustrative numbers rather than today's polls. Suppose today's shares (after the lens) are Conservatives 21.0, Labour 26.0, Liberal Democrats 10.5, Reform UK 22.5, Greens 11.5 and Others 8.5. The 2024 GB results are Conservatives 24.36, Labour 34.64 and Liberal Democrats 12.56.
Step 1: the weight. With the fitted floor of 0.62 and \(\tau = 30\) days, \(w(400) = 0.62 + 0.38\,e^{-400/30} \approx 0.62 + 0.38 \times 0.0000016 \approx 0.620\). At this distance the exponential term has vanished and the weight is the floor.
Step 2: pull the big three back. Each keeps 62% of its change since 2024:
- Conservatives: \(24.36 + 0.620 \times (21.0 - 24.36) = 24.36 - 2.08 = 22.28\)
- Labour: \(34.64 + 0.620 \times (26.0 - 34.64) = 34.64 - 5.36 = 29.28\)
- Liberal Democrats: \(12.56 + 0.620 \times (10.5 - 12.56) = 12.56 - 1.28 = 11.28\)
Reform UK, the Greens and Others stay at 22.5, 11.5 and 8.5.
Step 3: rescale. The shares now sum to 105.34, because the big three together polled 57.5 but took 71.6 in 2024, and the pull-back restores about 38% of that gap. Dividing by 1.0534 gives the projection:
| Party | Today | After pull-back | Polling day (rescaled) | Change |
|---|---|---|---|---|
| Conservative | 21.0 | 22.28 | 21.15 | +0.1 |
| Labour | 26.0 | 29.28 | 27.80 | +1.8 |
| Liberal Democrat | 10.5 | 11.28 | 10.71 | +0.2 |
| Reform UK | 22.5 | 22.50 | 21.36 | −1.1 |
| Green | 11.5 | 11.50 | 10.92 | −0.6 |
| Others | 8.5 | 8.50 | 8.07 | −0.4 |
Table: Worked example, 400 days before polling day, GB, illustrative shares.
Two things stand out. Labour gains most because it has fallen furthest from its 2024 result; the Conservatives, though further below their own 2024 share than the Lib Dems, end almost where they started, since the rescaling takes back most of what the pull-back gave them. And Reform UK loses 1.1 points without any pull-back of its own, purely through the rescaling. The live election-day forecast shows the same pattern with today's numbers, nation by nation.
Step 4: the uncertainty. At 400 days, equation (3) gives \(\sigma_{\mathrm{main}} = 3.85 + 0.56 \ln(13.3) \approx 5.30\) points and equation (4) gives \(\sigma_{\mathrm{ins}} \approx 5.70\). If Labour's final-poll root-mean-square error were 3 points, its 80% half-width would be \(1.2816 \sqrt{3^2 + 5.30^2} \approx 7.8\) points, so the band would run from about 20 to 36. Reform UK, with a smaller final-poll error of, say, 1.5 points but the larger drift, would have a half-width of about 7.6 points. These bands are wide, and they should be: 400 days before 2017 or 2024, a forecaster using the polls would have been wrong by more than this for at least one of the big parties.
Polling day, and how drift enters the simulation
The default date: 15 August 2029
A forecast for polling day needs a polling day. Under the Dissolution and Calling of Parliament Act 2022, a parliament is dissolved automatically five years after the day it first met, unless dissolved earlier, and an election follows. The current parliament first met on 9 July 2024, which makes 15 August 2029 the latest date on which the next general election can be held. The site uses that date by default. The reason is not a prediction that the prime minister will wait until the last possible day. It is that the latest date is the only one fixed by law; any earlier date is a guess about a political decision. It is also the cautious choice, since it gives the longest horizon and therefore the widest ranges. Readers can pick one of several preset dates or any date of their own on the election-day page, and the whole forecast (weight, drift, simulation) is recomputed.
The distance \(d\) is counted from the date of the latest poll in the average, not from the date the page is viewed. At the time of writing the default date is a little under three years away, within the 1,095-day range of the evidence. Earlier in a parliament a reader could ask for a date further out than the data cover; the weight would stay at its floor, and the drift formulae would be extrapolated along their logarithmic curves.
Order of operations
The election-day forecast is built as follows (ForecastService.BuildElectionDay):
- Take today's polling averages in England, Scotland and Wales (see Scotland and Wales: sparse polls and no steady lean for the last two).
- Apply the chosen lens in each nation, exactly as for the "election today" forecast.
- Apply equations (1) and (2) in each nation, pulling the big three towards that nation's 2024 result, and rescale to 100.
- Compute the drift standard deviation \(\sigma_p(d)\) for each party.
- Project seats with the chosen swing model (see From votes to Commons seats: UK swing models and their test) and run the simulation with the drift switched on.
Drift in the simulation
The simulation (see Simulating UK polling error: where the chances come from) runs 1,000 elections. In each run \(r\) it already draws a historical polling-error pattern, scaled by a standard normal factor, with independent national noise and seat-level noise. For the election-day forecast it also draws a drift for each party,
independently across parties. The same drift is added to the party's share in England, Scotland and Wales, on the view that a national movement in opinion between now and polling day moves a party everywhere. The run's share for party \(p\) in nation \(n\) is then
where \(\hat{v}_{n,p}(d)\) is the projected share from step 3 and \(e^{(r)}_{n,p}\) is the polling error drawn for that run. Drift is applied only to parties with a positive share in the nation, and the shares are rescaled to 100 before the seats are projected. The rescaling introduces a mild negative dependence between parties (if one rises, the others fall a little), but the drifts themselves are drawn independently, which is a simplification: in practice one party's gain is often a specific rival's loss.
The drift is centred on zero. All the directional expectation, the pull back of the big three, is in the central projection \(\hat{v}(d)\); the simulation adds spread around it, not further movement.
Limitations
The first limitation is the sample. Seven elections (ten in the data, the three before 2000 set aside), eight insurgent cases, and a curve with two parameters fitted to both. The leave-one-out test guards against the grossest overfitting but cannot manufacture evidence that is not there. The drift back could weaken, or the next parliament could be one in which the big three's mid-term position proves durable. The 2017–2024 period already contains some of the largest movements in the series.
Second, the prior is the previous result and nothing else. The model has no economic fundamentals, no leader ratings and no local-election results. This is deliberate: each would add parameters that seven elections cannot estimate. But it means the model's view of where the big three are heading is simply "some of the way back", with no view of why.
Third, the treatment of the insurgents is binary. Reform UK is either pulled back or not, and on the evidence it is not. But Reform's position in 2026 is unlike UKIP's in 2014: it is polling alongside the two old parties rather than behind them. Whether its support will behave like an insurgent's or, increasingly, like a main party's is a question the data cannot yet answer. The same is true, less acutely, of the Greens.
Fourth, the evidence is GB-wide but the model is applied nation by nation. We have assumed that Labour in Scotland drifts back towards its Scottish 2024 result with the same strength as Labour in Britain as a whole. In Scotland especially, where the 2024 result was itself an unusual swing, that assumption carries weight. It also produces some effects through rescaling that deserve watching: pulling Labour back in Scotland shrinks the SNP's projected share even though the SNP is carried forward as polled.
Fifth, the uncertainty is probably somewhat too wide rather than too narrow. The drift standard deviation is the root-mean-square error of the method against the result, which already contains the final-poll error at the end of the campaign. The vote bands then add the final-poll error again, in quadrature. At long distances drift dominates and the double count is small; at one or two months out it is not. We prefer bands that err on the wide side, but the choice is not calibrated and is recorded here as such.
Sixth, the drifts are independent across parties and identical across nations. Both are simplifications. The polling error in the simulation has a realistic correlation structure because it replays whole historical patterns; the drift does not.
Conclusion
A poll three years before an election describes the present, not polling day, and British elections since 2001, like those before, are clear about the direction of the difference for the three traditional parties: they finish closer to where they started than their mid-term polls suggest. A simple two-parameter curve that counts a little under two thirds of their change since the last election, and a little more in the final weeks, has predicted their results better on average than the polls as they stood at every distance from one month to three years, though by less than on the older record. The same treatment made insurgent parties worse, so they are carried forward as polled. Drift uncertainty that grows with the logarithm of the time left widens the ranges accordingly. The result, on the election-day page, is deliberately kept apart from the "election today" headline: it is a forecast of a different quantity, built on a different and thinner body of evidence.
References
Erikson, R. S. and Wlezien, C. (2012). The Timeline of Presidential Elections: How Campaigns Do (and Do Not) Matter. Chicago: University of Chicago Press.
Gelman, A. and King, G. (1993). Why are American presidential election campaign polls so variable when votes are so predictable? British Journal of Political Science, 23(4), 409–451.
Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions. Journal of the Royal Statistical Society: Series B, 36(2), 111–147.
Wlezien, C., Jennings, W., Fisher, S., Ford, R. and Pickup, M. (2013). Polls and the vote in Britain. Political Studies, 61(S1), 66–91.
-
PollBase is compiled by Mark Pack and is available from his website. We use it only to average published GB voting intention at fixed distances before past elections; nothing from it enters the "election today" forecast. The horizon points were built by
research/horizon/build.pyand loaded by database script 12 (generated byresearch/method/make_12.py); the Python scripts are kept as the record of how the stored data were produced, while the model itself is C#.↩ -
The decisions log describes the wider formula as applying to "the challengers" in general. The code is narrower: only Reform UK and the Greens, the two parties whose history the formula was fitted on, receive it. Restore Britain and Your Party have no history of any kind, and giving them the insurgent drift would be as defensible as the main-party drift; we note the choice rather than defend it strongly.↩
How to cite
Polling Forecast (2026). From today to polling day: how far British polls move before a general election. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-election-day. Live figures change as new polls arrive; the date above is when the text was last revised.


