Methodology

How we collect and review polls

1. Collection. Polls are entered by our editors or imported automatically from polling institutes’ published releases.
2. Pending review. Automatically imported polls enter a review queue, flagged with anything that could not be read from the source (for example, missing fieldwork dates). Only sources our editors have verified as consistently reliable are published directly, and they are still checked for duplicates and implausible figures.
3. Editor approval. An editor checks each figure against the original source before approving it. Possible duplicates of an existing poll are flagged rather than added.
4. Publication with attribution. Published polls show the institute, publication date, fieldwork period, sample size and methodology, with a link to the original source that is never altered.

Automated tools may only record what the source literally states. Missing values are left blank and flagged for a person to check; they are never estimated.

Changes and trends

The change shown next to each figure compares a poll with the previous poll by the same institute for the same election, in percentage points. Comparing different institutes can mislead, because their methods differ.

The polling trend chart plots each published poll as a point. Each line is the Pollytic polling average, worked out again for every day. For each party or candidate it is a weighted mean of the polls published in the 60 days up to that day; older polls are left out. Each poll’s weight is the product of four parts:

• Recency. A poll’s weight halves every 7 days, so last week’s polls count far more than last month’s.
• Sample size. Weight grows with the square root of the number of people interviewed, up to 3,000; beyond that a bigger sample adds nothing. A poll that does not publish its sample is given the median sample of the other polls in the window, or 1,000 when none of them publishes one.
• Institute. Each institute counts as much as its best-weighted poll in the window, and that weight is shared among its polls in proportion to their own weights. An institute that polls every week therefore counts about as much as one fresh poll, never more and never much less, so a firm that polls daily cannot drown out one that polls monthly.
• Track record. Institutes ranked on the pollster ratings (at least 3 past elections compared with the result) are weighted by the median ranked institute’s error divided by their own, kept between 0.5 and 1.5: an institute twice as accurate as the median counts 1.5 times, one twice as far off counts half. Every other institute counts 1.

Each party or candidate is averaged only over the polls that include it. One that appears only in a few old polls, carrying less than 5% of the window’s total weight, is left out rather than shown at a stale figure. A new poll moves the average a little instead of the line jumping to its figures, so one unusual poll cannot dominate the picture. The average is recorded every day, and the last value before election day is kept as the final average to compare with the result.

The average is not a forecast. It describes what the polls say now, not what will happen on election day, and it inherits any error the polls share. The written summary on each country page (“What the polls say”) is generated from the same figures: who leads the average and by how much, and the largest change over the past month.

Accuracy

After an election, the final Pollytic average (the last one recorded before election day) is set beside the result on the election page. The error for each party or candidate is the average minus the result, in percentage points; the mean absolute error is taken over those in both; the error on the winner’s margin compares the average’s gap between the result’s top two with the real gap.

Each institute is scored the same way on its own polls whose fieldwork ended in the 21 days before each election (the publication date when fieldwork dates are missing): the mean absolute error per poll, averaged over its polls in an election, then over elections, plus the same for the top-two margin. Institutes with at least 3 elections are ranked by mean error on the pollster ratings page; the others are listed as “not enough data”. The scores are recalculated every week.

Pollytic does not publish its own win probabilities yet. A forecast model will only be shown once it has been tested against past elections; until then pages show the Pollytic average, and any probabilities on the site are third-party forecasts, labelled as such.

US forecasts over time and their accuracy

The third-party US forecasts are recorded once a day: every forecaster’s rating and chances for each race, and its projected seat totals. The forecast page charts how the model’s chance of a majority has moved and compares each forecaster’s totals now, a week earlier and on the first day recorded; race pages chart each forecaster’s chances for that race and list every rating change. These appear once there are at least two days on record.

After the count, each forecaster is scored on the last ratings it published before election day, for every Senate and governor race with a stored result (House races are not scored: district results are not stored). A call is correct when the party it rated Tilt, Lean, Likely or Safe won; a Toss-up is not a call and counts as not correct. Forecasters are ranked by the share of their rated races called correctly. For forecasters whose percentages are win probabilities, the Brier score is the mean of (1 − p)² over those races, where p is the chance given to the winner’s party: 0 is a perfect forecast and 0.25 is what a coin flip scores in a two-way race. It breaks ties in the ranking.

Live results, swing and race calls

While a count is in progress (from the day before election day until the count is complete, two weeks at most), the results and the election-night board are updated every two minutes. Otherwise they are shown from Pollytic’s own copy of the results feed, updated twice a day.

On US election night, the control board adds each party’s seats: seats are called when the results feed marks a winner and leading when votes are in but no winner is marked. Senate seats not up this year come from the latest model forecast (its projected totals minus the seats up that it rates to each party); governorships not up are the current governors of the states without a race. The majority lines are 218 of 435 House seats and 51 of 100 Senate seats.

The county swing on a race page compares each county’s margin (the Democratic share minus the Republican share of all votes counted) with the same county in the state’s 2024 presidential race, for counties with votes in both. The statewide figure is the average of those county shifts weighted by the votes counted now, so it covers only the counties reporting in both races. County names are matched to the map by name; a county that cannot be matched with certainty is listed under the map rather than placed.

Race calls are the results feed’s, not Pollytic’s. The time shown with each call is when Pollytic first saw it while updating the live pages, so it can trail the call itself by a few minutes.

Governments and ideology

Government composition and ideology labels are editorial information maintained by our team. They are never derived from polling figures.

Data sources

Pollytic draws on Wikipedia for polls, governments and election results, and on civicAPI for US forecasts, early vote figures and live election results.

Polls are found through the lists of published polls that Wikipedia’s editors keep for each election, and for Germany through Wahlrecht.de. Each poll page links to the publication those lists cite for it, usually the institute’s own release or the outlet that commissioned it, so any figure can be checked against where it was first published; for Germany the link is Wahlrecht.de’s table for that institute.