How the ElectionsUSA average works

1. Which polls are included

Any publicly released poll of the relevant population (national adults, registered or likely voters, or a party's primary voters) with a stated pollster and fieldwork dates. We don't exclude partisan or internal polls, but they are rarely graded well and so carry low weight. Polls are collected from pollsters' releases and from community-maintained compilations that cite them; the row's link always points to the source we used.

One entry per poll. When a pollster releases one poll for several populations, only one enters the average, in this order: likely voters, then registered voters, then all adults. A poll is identified by its race, fieldwork dates, the firm that ran it and its media partners, so the same poll listed twice in a source table (sponsor and pollster swapped, or "Zogby" next to "Zogby Analytics") counts once, while two different polls by the same firm in the same days (for The Washington Post and for Reuters, say) both count. Every population stays in the open dataset.

2. Weights

Each poll's weight is the product of four factors:

1. Recency. 0.5 ^ (age_in_days / half_life). The half-life is 14 days for approval and the generic ballot, 45 days for 2028 general-election matchups and 60 days for the 2028 primaries while polling is sparse; these shorten as Election Day approaches and every change is logged here. A poll two half-lives old counts a quarter as much as one released today.
2. Sample size. min(sqrt(n / 600), 1.5). Unknown n is treated as 600. Error falls with the square root of sample size, not linearly, and above ~1,350 respondents design effects dominate, so the bonus is capped.
3. Pollster quality. FiveThirtyEight's final letter grade for the firm that conducted the poll, mapped to a weight from 1.0 (A+) to 0.1 (F); unrated pollsters get 0.6. A poll released as "The Economist/YouGov" or "CNN/SSRS" takes YouGov's or SSRS's grade, never the media sponsor's, and a firm's name must match the one 538 rated (or a name 538 rated it under before, such as Marist College for Marist University). We will switch to our own grades once the 2026 scorecard has enough polls.
4. Frequency damping. If a firm has k polls inside the active window, each is divided by sqrt(k), so a daily tracker contributes like roughly three polls, not thirty. Polls count toward the firm that ran them, so YouGov's polls for The Economist, Yahoo News and CBS News are damped together.

Polls weighing less than 1% of the heaviest poll are dropped. The average for each candidate is the weighted mean of the polls that asked about that candidate, and a candidate is shown only if at least one poll inside the active window (about 4.3 half-lives) included them, so a name that pollsters have stopped testing drops out rather than lingering on stale numbers.

3. The trend line

For each date on the chart we recompute the average using only polls whose fieldwork had ended by that date. The line therefore shows what the average was, not a smoothed hindsight curve, and it never changes retroactively when new polls arrive.

4. What we don't do

No house-effect adjustment, no likely-voter conversion, no trend-line extrapolation, no "fundamentals". These can improve accuracy but they also add judgement calls that are hard to audit. Our priority is that a reader can reproduce every number from the table beneath it.

5. Pollster scorecard

Historical accuracy is computed from FiveThirtyEight's open archive (CC BY 4.0) of polls conducted in the final 21 days before an election, 2016–2022. For each pollster we report mean absolute error on the margin, mean signed error (bias), the share of races called correctly, and an adjusted error shrunk toward the field mean with k = 15. After the November 3, 2026 midterms we apply the same calculation to this site's own 2026 polls.

6. Prediction markets

Prices are read from Polymarket's public API once per update and stored as implied probabilities. Closed, inactive and thinly traded contracts (under $10,000 traded) are skipped, and a price that hasn't moved in four snapshots is hidden as a dead contract. Kalshi is linked rather than republished. We show them next to poll shares for the same outcome but do not combine them; they measure different things.

7. Updates and corrections

Data refreshes several times a day; the timestamp on every page is the build date. Errors reported via the corrections page are fixed in the dataset so the correction propagates to every chart, table and the API.

Method changes. November 3, 2026: polls are now de-duplicated before averaging (one population per poll, likely voters first, and copies of the same poll removed). Until then a poll published for likely voters, registered voters and adults counted up to three times, and ten polls were listed twice. The corrections log lists the averages that moved.

November 2026: pollster grades and frequency damping now follow the firm that conducted each poll rather than the first name in the source table, which was often the media sponsor. Earlier, partial name matches also gave some polls another pollster's grade (co/efficient got the A+ of The New York Times/Siena College). Every average and trend line was recomputed with the corrected weights.