How our numbers are made

Since 22 October 2026, the UK average is model uk-ssm-v2.0, an average of British voting-intention polls. It uses eligible polls from 5 July 2024 onwards, and recent polls count most. The darker band is the range of the average (90%). In simulations, this range contained the true average 88% to 91% of the time. In tests on real polls, the model's 90% range for a new poll contained that poll's figure 86% of the time for Labour and 88% to 91% for the other main parties. The lighter band shows how far results have differed from the final polls at past elections, as an illustration, not a forecast.

Before that the UK average was wra-v1.0. A weighted average of published Westminster polls. Each poll counts according to three things multiplied together: how recent it is (its weight halves every 14 days), its sample size (the square root of the sample over 1,000, capped at three times), and the pollster's reliability score below. The shaded range is a 90% band. That series stays available to download: the previous average.

The previous method (wra-v1.0) in full is below.

Until 22 October 2026 the UK average was wra-v1.0. It is still computed every day and can be downloaded. It is a daily weighted polling average of UK voting intention, not a prediction. It is designed to smooth out the noise from individual polls while still responding to real movement in opinion.

The model reflects the polls. If polls are systematically biased, as they were in the 2024 general election, when they collectively overestimated Labour by about 7 points, the model will reflect that bias. No ad-hoc correction is applied for past biases, because the direction and magnitude of polling error is unknown in advance.

The three weighting factors

Recency. Exponential decay with a 14-day half-life: a poll conducted 14 days ago receives half the weight of one conducted today. There is no hard cutoff; a 60-day-old poll carries about 4% of a fresh poll's weight. Validated by backtesting against the 2024 general election result. Formula: w_recency = exp(−ln(2) × days_ago / 14).

Sample size. Proportional to the square root of the sample size, normalised to 1,000 respondents and capped at 3×, so very large polls cannot dominate. Formula: w_sample = min(√(n / 1000), 3.0).

Pollster reliability. Each pollster has a score bounded between 0.3 and 0.7, shown in the table above and calculated from four inputs: election accuracy across multiple elections with systemic-bias deflation (40%), consensus alignment against an EMA-smoothed peer trend (20%), methodological trust (25%), and house-effect magnitude and stability (15%). Each input is shrunk toward an informed prior before composition, with the prior conditioned on BPC membership and methodology type. The bounds of 0.3 and 0.7 are limits by construction, not the observed range: the scores in use today run from 0.42 to 0.57, so the most trusted pollster carries 1.4 times the weight of the least trusted. No pollster is excluded, because the floor is 0.3. The score does not decide how much of the average a pollster accounts for: that depends mostly on how many recent polls it publishes and their sample sizes, and the score moves it only modestly.

A poll's final weight is the product of the three factors.

Uncertainty bands

The shaded ranges are 90% uncertainty bands: the weighted standard deviation of polls in the recency window, floored at 1.5 points and multiplied by 1.645. The floor is applied to the standard deviation before the multiplication, so a band is at least about ±2.5 points on each side (1.5 × 1.645 ≈ 2.5), before it is cut off at 0% or 100%. They are called uncertainty bands rather than credible intervals because the model is not fully Bayesian: they reflect pollster disagreement, not a formal posterior. Smaller parties often show wider bands: pollsters disagree more on smaller-party shares because of turnout assumptions, question prompting and sampling variability.

House effects

A house effect is a pollster's systematic tendency to read a party higher or lower than the consensus of other pollsters at the same time, measured in a ±14-day window. House effects are reported for transparency and are not used to adjust the wra-v1.0 average.

What this model does not do

It does not predict elections, and does not account for likely-voter turnout, tactical voting or campaign effects. The seat projection on the UK page uses uniform national swing, applying the same change to every constituency; it is illustrative, not a forecast.

Data and limitations

Polling data is sourced from publicly available records, primarily Wikipedia's UK opinion-polling tables, which aggregate polls published by British Polling Council members and other pollsters. Raw and cleaned versions are both stored, every entry carries source attribution, and polls failing validation (party totals above 105%, very small samples) are excluded with the reason recorded. Known limitations: unpublished or embargoed polls cannot appear; a source format change can break the scraper until fixed; the multi-election accuracy framework currently has full data for 2024 only; all polls are treated as measuring the same thing despite differing question wording and turnout filters; and GB-only and UK-wide polls are combined, which may slightly affect SNP and Plaid Cymru figures.

Reliability scoring is at v6.0 (March 2026). The full version history, v1.0 to v6.0, is in the repository's methodology record; the parameters of wra-v1.0 are the ones stated on this page.

The method of the previous UK average (wra-v1.0): recency with a 14-day half-life, sample size and pollster reliability, with a 90% band. The Swedish reliability scores are not yet published.

Full methodology: fredcornell.com/methodology/sweden

A daily average of national approval polls from the New York Times file of approval polls. Every pollster with enough published releases is counted, with the same quality weight; polls are weighted by recency and sample. Polls the New York Times labels as having a partisan sponsor are excluded. The approval page lists every poll, whether it is counted and the weight it carries. No uncertainty range is published at present.

Full methodology: fredcornell.com/methodology/us-approval

A forecast, not an average: the model simulates the election and reports how often each party wins 218 seats. It uses district polls where they exist and a structural estimate where they do not. Pollster quality does not enter the model; every pollster carries the same weight. The generic ballot average on the midterms page is a separate product, explained on the same methodology page.

Full methodology: fredcornell.com/methodology/us-house

A forecast of the 35 contested seats, reported as four seat-count outcomes. Pollster quality does not enter the model.

Full methodology: fredcornell.com/methodology/us-senate

Polling averages for each party's nomination, shown beside prediction-market prices.

Full methodology: fredcornell.com/methodology/us-2028