Simulating UK polling error: where the chances come from
Abstract. The site's seat ranges and probabilities come from 1,000 simulated elections. Each one replays one of the four real England polling-error patterns of 2015–2024, scaled by a random factor centred on zero, with Scotland and Wales taking their own errors from the same year, plus independent national and seat-level noise. Because errors are drawn as whole patterns, the simulation keeps the strong negative correlation between Conservative and Labour errors, but four patterns are a thin basis for calibration and some sources of error are left out.
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.
- Every outcome (3 October). The seats in every one of the simulated elections are shown as histograms at /uk/outcomes, with the majority line.
- Corrected in this paper: the movement between now and polling day is fitted on the polls before the seven elections from 2001 to 2024 (PollBase, which runs from 1983); the text said "every GB poll since 1983".
From a best estimate to a probability
Our projection is a single set of vote shares and a single seat projection (see Correcting the raw answers: the method behind our UK projection and From votes to Commons seats: UK swing models and their test). A reader who sees one party on a given number of seats will reasonably want to know how confident we are. How likely is a majority? Which seats are close enough to go either way? The lenses give part of the answer by replaying named historical misses, but a replay is a scenario, not a probability. To attach chances to outcomes we need a model of how wrong the inputs might be, and we need to push that uncertainty through the seat model many times.
This paper describes that model, Seats/Simulation.cs. Each of 1,000 runs perturbs the vote shares of the headline lens with a draw of polling error, re-projects every seat, and records the winner. The seat ranges, the majority and hung-parliament probabilities and each seat's win probability are all counts across those runs. We set out exactly how a run is built, why errors are drawn as whole patterns from past elections rather than party by party, what the outputs mean, and the ways in which the result may be miscalibrated. The last point deserves emphasis. The simulation is built from four elections of evidence, and its probabilities are only as good as that evidence.
One run, step by step
A run has five steps. Formal definitions follow in the next sections.
- Choose a past miss. Pick one of the core elections (2015, 2017, 2019 or 2024) at random, each with equal probability, and take the pattern of errors the final England polls made at that election.
- Scale it. Multiply the whole pattern by one random number drawn from a standard normal distribution. The number can be large or small, positive or negative.
- Add national noise. Add a small independent random error to each party.
- Apply it in each nation. Subtract the resulting errors from the headline lens's shares in England. In Scotland and Wales, use each nation's own error pattern from the same election, scaled by the same number. For an election-day forecast, add a random drift for the time left to polling day as well.
- Project every seat with seat-level noise. Run the seat model from each nation's perturbed shares, add a random local error to each party in each seat, and record the winner.
The simulation always runs around a lens. On the home page that is our projection. When a reader chooses another lens on the map, the forecast for that lens is simulated around that lens's shares instead. The reader's own what-if error is shown as a central projection only and is not simulated.
The random number generator is seeded with a fixed value (20260925). Given the same polls, the same lens and the same seat model, the simulation returns exactly the same result. A change in the published chances therefore always reflects a change in the inputs, never Monte Carlo noise from one page load to the next.
National polling error: patterns, not independent parties
The draw
Let \(t = 1, \dots, T\) index the elections in the England error history (\(T = 4\) at present) and let \(E_{t,p}\) be the average error of the published final England polls for party \(p\) at election \(t\), defined as poll minus result, for the five parties with an error history, \(p \in \mathcal{P} = \{\text{CON}, \text{LAB}, \text{LD}, \text{REF}, \text{GRN}\}\). These are the figures in the table below; UKIP and the Brexit Party count as Reform.
| Election | Final polls | Con | Lab | LD | Reform | Green | Lab lead |
|---|---|---|---|---|---|---|---|
| 2015 | 2 | −5.7 | +1.5 | +1.8 | +0.8 | +0.9 | +7.1 |
| 2017 | 4 | +0.6 | −3.9 | +0.7 | +2.1 | +0.6 | −4.5 |
| 2019 | 5 | −1.6 | +0.9 | −0.6 | +0.8 | +0.1 | +2.5 |
| 2024 | 6 | −3.5 | +4.3 | −0.9 | +2.0 | −0.6 | +7.8 |
| Mean | −2.6 | +0.7 | +0.3 | +1.4 | +0.2 | +3.3 | |
| Root mean square | 3.5 | 3.0 | 1.1 | 1.5 | 0.6 |
In run \(r\) the simulation draws
with national noise \(\sigma = 1\) point (the NationalNoise setting in SimulationSettings). The England error for party \(p\) in run \(r\) is then
The same scale \(z_r\) multiplies every party's error in the chosen pattern, so the pattern keeps its shape. A run that draws the 2015 pattern with \(z_r = 0.5\) assumes the polls understate the Conservatives and overstate Labour, as in 2015, but by half as much.
Centred on zero
Because \(z_r\) has mean zero, a pattern is as likely to be applied in reverse as forwards. A run that draws the 2024 pattern with \(z_r = -1\) assumes the polls understate Labour by as much as they overstated it in 2024. This is a deliberate choice, and the code comments say so: the simulation "assumes no direction". The headline lens is already our best estimate of where the polls are wrong. If the simulation then shifted it further in the historical direction, it would count that correction twice. Centring on zero means the simulation represents uncertainty around the headline, not a second correction.
Two properties of equation (1) follow directly. Since \(z_r\) is independent of \(k_r\) and \(\varepsilon_{r,p}\), and \(\mathbb{E}[z_r^2] = 1\),
where \(\text{RMS}_p\) is the root-mean-square past error for party \(p\), the last row of the table above. So the spread of simulated national error for each party matches its historical root-mean-square error, widened a little by the noise term. And for two parties \(p \neq q\),
which is the average cross-product of the two parties' past errors. The simulation therefore reproduces the way parties' errors have moved together, not only how big they have been.
Why patterns rather than independent errors
The obvious alternative would be to draw each party's error independently, with a spread set by its past errors. That would be wrong in a way that matters for seats. Polling errors are not independent across parties. Vote shares add up to 100, so a poll that overstates one party must understate others. More substantively, the causes of the big British misses (chiefly unrepresentative samples, with too many politically engaged respondents; see Sturgis et al., 2016, and Mellon and Prosser, 2017) push the main parties in opposite directions at once. In England, the Conservatives were understated and Labour overstated in 2015, 2019 and 2024; 2017 went the other way for both.
With the four patterns currently in the record, equation (3) gives a correlation of about −0.6 between the simulated Conservative and Labour errors, once the national noise is included, and about −0.4 between the Conservatives and Reform. Independent draws would set both to zero. The difference matters because seat outcomes depend above all on the gap between the two leading parties in each seat. If Conservative and Labour errors were drawn independently, many runs would move both up together or both down together, and the Labour–Conservative lead would vary less than it has. With correlated errors, a run that raises the Conservatives usually lowers Labour, and the lead swings the way it has at real elections.
A pattern-based draw is also simple to explain and to check. Every run corresponds to a real miss, larger or smaller, forwards or backwards. The price is that the simulation can only produce errors in the directions spanned by the past patterns, plus noise. We return to this in section 10.
Scotland and Wales
Scotland and Wales have their own polling-error records, measured from Scotland-only and Wales-only final polls at the same four elections (Scotland and Wales: sparse polls and no steady lean). Their polls missed by about as much as England's, but not in the same way. The simulation uses them as follows. Let \(E^{(n)}_{t,p}\) be nation \(n\)'s own average error for party \(p\) at election \(t\). When the run's chosen election \(k_r\) has a record for nation \(n\), the error applied in that nation is
The nation uses its own pattern from the same year, scaled by the same \(z_r\). A bad year for the polls is therefore bad in all three nations at once, but each nation misses in its own way. The GB-wide parties share the national noise term \(\varepsilon_{r,p}\) with England, while the SNP and Plaid Cymru each get their own independent noise. This is how the SNP and Plaid obtain proper ranges: their uncertainty comes from their own nation's record, not from England's. If a nation has no record for the chosen year, it takes England's errors from equation (1) and the SNP and Plaid are left unperturbed. At present both nations have records for all four elections.
The table below shows the Scotland and Wales errors the simulation draws on.
| Nation | Election | Polls | Error by party |
|---|---|---|---|
| Scotland | 2015 | 3 | SNP +0.3, Lab 0.0, Con +0.6, LD −1.6, Green +0.5, Reform −0.3 |
| Scotland | 2017 | 3 | SNP +4.1, Lab −0.9, Con −2.9, LD −1.2, Green +0.8, Reform −0.2 |
| Scotland | 2019 | 3 | SNP −1.3, Lab −1.6, Con +2.2, LD +0.6, Green +0.2, Reform −0.4 |
| Scotland | 2024 | 6 | SNP +2.4, Lab −0.5, Con +1.0, LD −1.6, Green −0.6, Reform −0.7 |
| Scotland | Leave-one-out | as polled 2.10, corrected 2.25 (mean abs. error) | |
| Wales | 2015 | 1 | Lab +2.1, Con −2.2, Plaid +0.9, LD +1.5, Reform −1.6, Green −0.6 |
| Wales | 2017 | 1 | Lab −3.0, Con +0.4, Plaid −1.4, LD +0.5, Reform +3.0, Green −0.3 |
| Wales | 2019 | 1 | Lab −0.9, Con +0.9, Plaid +0.1, LD 0.0, Reform −0.4, Green 0.0 |
| Wales | 2024 | 1 | Lab +8.0, Con −0.2, Plaid −2.8, LD −1.5, Reform −4.0, Green −0.7 |
| Wales | Leave-one-out | as polled 1.77, corrected 2.31 (mean abs. error) |
Drift for election-day forecasts
Everything so far describes uncertainty about an election held today. The election-day forecast adds a second source: the vote may move between now and polling day. The size of that movement is estimated from British polls at fixed distances before the seven elections from 2001 to 2024 (From today to polling day: how far British polls move before a general election). For each party it is a standard deviation \(d_p\) that grows with the logarithm of the time left, using a formula fitted to the Conservatives, Labour and the Liberal Democrats for most parties, and the larger of that and a second formula, fitted to the challengers, for Reform and the Greens. Under 30 days it shrinks linearly to zero on polling day.
In each run the simulation draws one drift per party, shared by all three nations:
and adds it to the party's share before the polling error is subtracted. Drift applies only to parties with a positive share in that nation, and a share cannot be pushed below zero by it. For "election today" forecasts no drift is set and this step does nothing. Drift reflects the long-established finding that polls become more predictive of the result as polling day approaches (Erikson and Wlezien, 2012; Wlezien et al., 2013), and that campaign polls can move a great deal even when the eventual vote is fairly predictable (Gelman and King, 1993).
Applying the error and projecting the seats
Nation shares in a run
Let \(V_{n,p}\) be the headline lens's share for party \(p\) in nation \(n\). The perturbed target for run \(r\) is
where the drift term is omitted for parties with a zero share and \(\mathcal{N}\) sets negative values to zero and rescales to 100, exactly as in the lenses (Lenses: reading the same British polls several ways). For England \(e^{(n)}_{r,p} = e_{r,p}\). Parties without an error history, Restore Britain and Others, are not perturbed directly; they move only through the rescale and, for Restore Britain, through drift.
Seat-level noise
Each GB seat is then projected with the chosen swing model \(M\) (uniform swing by default) from its 2024 result, its nation's 2024 baseline and \(V^{(r)}_{n}\). For every party whose projected share \(\hat{s}_{i,p}\) is above 0.05 points, a local error is added:
with \(\tau = 3\) points (the SeatNoise setting). The square-root term makes the noise proportional to the square root of the share, loosely like sampling variation: a party on 25% has a local standard deviation of 3 points, one on 4% of 1.2 points, one on 50% of about 4.2 points. The shares are not rescaled afterwards; the winner is simply the party with the largest perturbed share. The seat noise is independent across seats, parties and runs. Northern Ireland's seats and the Speaker's seat keep their 2024 winners in every run.
The seat noise stands for everything local that the swing model cannot see: candidates, incumbency, local campaigns, tactical voting that departs from the pattern of 2024. Its size of 3 points at a 25% share is a working setting. The decisions log records it but no test that estimated it from data.
What comes out
After 1,000 runs, every output is a count.
- Seat ranges
- For each party, the seat totals across runs are sorted and read off at the 10th and 90th percentiles (the 80% range), the 2.5th and 97.5th (the 95% range) and the median, using the nearest rank.
- Majority probability
- The share of runs in which a party wins at least 326 seats.
- Largest party
- The share of runs in which a party wins strictly more seats than any other. A run in which two parties tie for most seats counts for neither.
- Hung parliament
- The share of runs in which no party reaches 326.
- Seat win probabilities
- For each seat, the share of runs in which each party wins it. These drive the win-chance colouring on the map and the probabilities on each seat page.
For today's headline forecast, the table below shows the vote shares and seats under our projection and the polls as they stand, with the simulated 80% range of seats for each party. Across the 1,000 runs, no party reaches a majority in 66% of them.
| Party | Polls % | Our bet % | Seats (polls) | Seats (our bet) | 80% range |
|---|---|---|---|---|---|
| Lab | 28.1 | 27.2 | 329 | 317 | 237–372 |
| Con | 21.6 | 22.6 | 143 | 154 | 88–217 |
| LD | 10.1 | 10.2 | 72 | 72 | 62–75 |
| Reform | 21.8 | 21.2 | 37 | 36 | 29–81 |
| Green | 9.5 | 9.8 | 4 | 4 | 4–6 |
Two points of interpretation. First, the ranges are ranges of seats given our vote-share uncertainty. They are not confidence intervals in the sampling sense, and the 80% range is not guaranteed to contain the result 80% of the time; section 10 explains why we cannot yet check that. Second, with 1,000 runs, a reported probability has a Monte Carlo standard error of at most \(\sqrt{0.25/1000} \approx 1.6\) points. Differences of a point or two between probabilities should not be read as meaningful.
Vote-share bands
The site also shows an 80% band around each party's vote share. This is not taken from the simulation. It is computed directly from the past errors:
where 1.2816 is the 90th percentile of the standard normal distribution, so that \(\pm 1.2816\) standard deviations covers 80% of a normal distribution, and \(d_p\) is the drift standard deviation (zero for "election today"). For the SNP and Plaid Cymru the band uses the root-mean-square error of their own nation's record instead of England's.
Treating the root-mean-square past error as a standard deviation is a simple rule. It assumes errors are centred on zero around the headline, as the simulation does, and roughly normal. Three details are worth noting. The band leaves out the 1-point national noise that the simulation adds, so for parties with small past errors the simulated spread is noticeably wider than the band: for the Greens, whose root-mean-square error is well under a point, adding the noise nearly doubles the standard deviation. The SNP and Plaid bands do not include drift even on the election-day page. And the decisions log notes that the Green and Liberal Democrat errors were measured when those parties had lower vote shares than they poll now, so their bands are probably too narrow. Making error proportional to share was proposed for a later model version and has not been done.
What the simulation does not capture
The simulation represents one kind of uncertainty well: national polling error of the size and shape seen at recent elections. Several others are missing or only roughly handled.
Seat-model error. The seat-level noise in equation (7) is independent from seat to seat and centred on zero. The backtest of the seat model (From votes to Commons seats: UK swing models and their test) shows that its errors in 2024 were neither. Given the exact 2024 nation shares, uniform swing under-called Labour by 82 seats and the Liberal Democrats by half their total, because tactical voting and uneven swing moved many similar seats the same way at once. Independent noise averages out across hundreds of seats; systematic error does not. The simulated seat ranges are therefore narrower than the true uncertainty about how votes turn into seats.
Regional and demographic correlation. A polling miss that falls unevenly, larger in the North than in London or among some kinds of voter than others, is not represented. Within a nation every seat receives the same national error in each run.
New parties. Restore Britain has no error history and no seat history. It receives only seat noise (and drift on election-day forecasts), and its vote is spread evenly across seats. The simulation cannot express the real uncertainty about where its vote is concentrated.
Uncertainty in the correction. The headline lens itself rests on a correction estimated from the same four elections (Correcting the raw answers: the method behind our UK projection). The simulation takes that lens as its centre and does not add uncertainty about the correction.
Headline errors around a raw-answer estimate. The patterns are the errors of the pollsters' published final polls, while our projection is built from their raw answers and corrected. In backtests our projection has missed by less than the published figures on average, so using the published errors is, if anything, cautious about the size of the error. The shape may differ, though. Our method's weak spot is Labour, which it left too high in 2019 and 2024, and the published patterns are not a direct measure of that.
Northern Ireland. Its 18 seats are fixed at their 2024 winners.
Everything outside polling error. Events, changes of leader, and turnout shocks on polling day are represented only in so far as they resemble past polling misses or, for election-day forecasts, historical drift.
Calibration with four patterns
A probability forecast is well calibrated if events given 20% happen about 20% of the time. Checking that needs many forecasts with known outcomes. We have four elections, which is not enough to test the calibration of anything. The best we can do is ask whether the design is reasonable and look for ways it could mislead.
The first concern is the small number of patterns. Every run is one of four shapes, scaled and with noise added. With \(T = 4\) the covariance in equation (3) has rank at most four, and it reflects four particular elections. If the next miss has a new shape, for instance one driven by a party with no polling history, the simulation gives it little weight. Adding the extended-era elections (2010, and 2005 as a robustness check) would give more patterns; the decisions log plans to bring 2010 forward, but it is not yet part of the error history.
The second is the scale distribution. Scaling by a standard normal means the average absolute scale is \(\sqrt{2/\pi} \approx 0.80\), so a typical run replays a pattern at about four-fifths of its original size, and only about 32% of runs apply it at full size or larger. Combined with the sign flip, this makes the simulation's national error a mixture of normal distributions whose variance matches the past errors, as equation (2) shows. It does not make the tails right. Whether a miss half again as large as 2015's deserves the probability the normal tail gives it is something four elections cannot tell us.
The third concern is the one 2024 illustrates. Suppose the simulation had been built from 2015, 2017 and 2019 only, as it would have been before the 2024 election. Using equations (1) and (2) with those three patterns and 1 point of noise, the chance of the England polls overstating Labour by as much as they did in 2024 (4.3 points) comes out at about 5%. The chance of overstating the Labour lead over the Conservatives by as much as in 2024 (7.8 points) comes out at about 6%.1 Neither is negligible, and both come mostly from the reversed 2017 pattern and the 2015 pattern respectively, which is the design working as intended. But a 1-in-20 event happened at the most recent election, and the next record will be judged with it included. The decisions log states plainly that no variant of our correction predicted the 2024 Labour miss from the elections before it.
The fourth is that all the seat-level choices, the 3-point noise, its square-root scaling and its independence across seats, are settings, not estimates. They affect the width of seat ranges and the win probabilities of close seats more than the national error does.
Taken together, we think the simulation's ranges are more likely too narrow than too wide, chiefly because of seat-model error and the small number of patterns. We show them because a range built from real past errors is more honest than a single number, and because the alternative of assuming sampling error alone is known to understate the total error of election polls by a wide margin (Shirani-Mehr et al., 2018). They should be read as a guide to how the forecast could go wrong in familiar ways, not as a complete account of the uncertainty.
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.
Mellon, J. and Prosser, C. (2017). Missing nonvoters and misweighted samples: explaining the 2015 great British polling miss. Public Opinion Quarterly, 81(3), 661–687.
Shirani-Mehr, H., Rothschild, D., Goel, S. and Gelman, A. (2018). Disentangling bias and variance in election polls. Journal of the American Statistical Association, 113(522), 607–614.
Sturgis, P., Baker, N., Callegaro, M., Fisher, S., Green, J., Jennings, W., Kuha, J., Lauderdale, B. and Smith, P. (2016). Report of the Inquiry into the 2015 British General Election Opinion Polls. London: Market Research Society and British Polling Council.
Wlezien, C., Jennings, W., Fisher, S., Ford, R. and Pickup, M. (2013). Polls and the vote in Britain. Political Studies, 61(S1), 66–91.
-
These are analytical results, not simulation runs. With three patterns, each chosen with probability one third, a party's simulated national error is a mixture of three normal distributions with mean zero and variance \(E_{t,p}^2 + \sigma^2\). The tail probabilities are averaged over the three. For the Labour lead the variance is \((E_{t,\text{LAB}} - E_{t,\text{CON}})^2 + 2\sigma^2\), since the lead difference carries two independent noise terms. The 2015–2019 errors are those in the table in section 3.↩
How to cite
Polling Forecast (2026). Simulating UK polling error: where the chances come from. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-simulation. Live figures change as new polls arrive; the date above is when the text was last revised.


