How we read American polls
How we turn the US's polls into a forecast for the Electoral College (538 seats, 270 for a majority): the 3 polls of the last 30 days, corrected for how the US's final polls have missed before. Today's average: Democrats 54.3, Republicans 44.8.
The same rule as everywhere on the site: measure how the final polls missed at past elections, correct only when predicting each past election from the others says the correction helps, and let the simulation miss the way the polls have missed. Every parameter below is fitted when the site starts, from the data in script 37, using elections from 2000 on.
- The generic ballot. Each pollster's latest poll ending within 21 days of the newest (likely voters where published, then registered voters, then adults), on the decided vote: Democrats 54.3, Republicans 44.8, a margin of D+9.5 from 6 pollsters.
- Correct it or not. Leaving out each of the 13 past final generic ballots in turn, taking off half the others' mean miss was closer on average (2.44 points a party as polled, 1.92 corrected). Our national House vote: D+7.5, give or take 4.9 points (one standard deviation: the final generic ballot's miss plus 0 days of drift).
- 435 districts. Each district's margin = 1.0 + 0.98 × its presidential lean + the national House vote + 1.4 for an incumbent standing again, with 19 uncontested seats settled and 109 districts' own polls blended in.
- 35 Senate and 36 governor's races. The pollster-adjusted average of the state's polls, blended with the fundamentals (lean, national vote, incumbency) by their measured errors. Older polls are brought up to date with the national trend since they were taken (our trend line through every generic-ballot poll, now D+11.5), marked ↻ on each race page; on the races of 2008-2024 this made the eve averages more accurate (4.2 to 3.8 points). State polls are not corrected: the leave-one-out test did not show a gain.
- 5,000 simulated elections give the chances: House majority 90% Democratic, Senate control 83% Democratic.
From the national vote to 435 districts
The lean. Every district's presidential margin against the nation's, three parts 2024 and one part 2020, on the lines in force for 2026. The research team placed the 2020 precinct results (VEST) and the 2024 precinct returns (MEDSL) into each district; in the nine states with new lines the 2024 votes are carried to the new districts by their 2020 shares, an estimate.
The fit. On 1142 contested districts in 2022 and 2024 and 2026 (2022 only in states whose lines stayed the same for 2024), each district's House margin minus the national House margin against its 2020 presidential lean and incumbency (a candidate who won the state's previous House election for the same party): intercept 1.0, slope 0.98, incumbency 1.4 points. What's left over has a typical size of 6.0 points: a district's own error.
Uncontested seats. Where only one of the two parties has a nominee in November (including California's same-party top-two races), the seat is that party's.
District polls. Averaged like the state polls below and blended with the district's fundamentals. District polls are few and often paid for by a campaign, so each partisan poll counts half and is moved back by the 2.5 points such polls have leaned; their error is taken as a quarter larger than statewide polls'.
Special elections. Since January 2025 Democrats have run on average about 18 points ahead of the 2024 presidential margin in House specials. We show this but do not use it: the site holds no specials from earlier cycles, so no backtest can show it improves the forecast (the rule for every input).
The tipping point. In each simulation the districts are lined up from the winner's safest to its weakest; the 218th is the one that gave the majority. The House page lists the districts that were it most often.
The poll average
General-election polls naming both nominees (generic "Democrat v Republican" tables count only in House districts). One version per poll: likely voters first; in Alaska, the last ranked-choice round. Each poll's margin is on the decided vote, less:
- the pollster's house effect: how far its final polls sat from the other final polls of the same races since 2000, shrunk toward zero for pollsters with few races;
- for a poll paid for by a campaign, party or allied group, or from a pollster known to work mainly for one side, the 2.5 points such polls have leaned to their side against the independent polls of the same race (652 final polls);
Keeping older polls up to date
National opinion moves, and a state poll taken a month ago was taken in a different national climate. So each poll is brought up to date: it moves by as much as the nation has moved since its fieldwork, read from our trend line through every generic-ballot poll, the daily and weekly trackers included (today D+11.5). A race polled last week and a race last polled in September are then compared on the same footing.
We tested it on the final month of every presidential, Senate and governor race from 2008 to 2024 (419 races): our election-eve averages became more accurate, the average miss falling from 4.2 to 3.8 points, with the biggest gains in 2016 (5.1 to 4.0) and 2020 (5.4 to 4.4). Every poll moved this way is marked ↻ on its race page, with the size of the move.
Weights: recency (a 30-day scale), the square root of the sample (capped at 1,500), the pollster's record (its typical final-poll miss against the typical pollster's), half for partisan polls, and the people sampled: likely voters in full, registered voters 0.7 (0.5 where an independent is the main challenger, since such candidates draw most on the voters least sure to turn out) and all adults 0.5. These last factors are set, not fitted. Several polls from one pollster share one pollster's weight. Polls from a firm found to have made up its data are removed, not down-weighted: Wedgewood Polls (all its polls) and SoCal Strategies (its polls since the firm was dissolved in May 2026). A pollster with no record (no poll in our records before December 2025, no final polls at past elections, not in AAPOR's Transparency Initiative or the Roper Center archive, and not a university) is marked "new, no record" on the race pages but counts in full: we tested counting such firms at half weight on the final polls of 230 races from 2018 to 2025, and the averages were no more accurate (our article on it).
How sure the average is depends on how many firms it rests on, not how many polls: one firm's polls share its habits. Each firm's share of the weight gives k, the effective number of firms (two firms with equal weight make 2; one firm with three quarters of the weight and three small ones make about 1.7). The average's error is √(4.7² + s²/k) points, where s is 5.7 points, the single-poll error fitted on every race's final polls since 2000, or how far this race's firms disagree with each other if that is more. A firm with no past final polls adds the uncertainty of its house effect (1.9 points, the typical established pollster's), less as its record grows. Then the drift polls taken 30-45 days out showed before polling day (3.2 points). So a race carried by one new firm's poll has a wide error and leans on the fundamentals, and a race polled by several established firms that agree rests on its polls.
The fundamentals
Fitted on every contested race since 2000 (margin = a + b × state lean + e × national House margin + i × incumbent):
| Office | Races | a | b (lean) | e (nation) | i (incumbent) | Typical miss |
|---|---|---|---|---|---|---|
| Senate | 444 | 2.9 | 0.84 | 0.74 | 9.2 | 15.0 |
| Governor | 355 | -1.3 | 0.39 | 0.61 | 14.5 | 14.0 |
The state lean before an election is three parts the last presidential result and one part the one before, against the national margin. Incumbents: the winner of the same seat's previous race standing again (surname and party). Independents who are the main challenger (Nebraska, Idaho, South Dakota) are treated as the Democrat would be, with half as much uncertainty again.
The blend
The poll average and the fundamentals are averaged with weights in inverse proportion to their errors, so a well-polled race rests on its polls and an unpolled one on its fundamentals. Each race page shows both and the weight.
Vote shares
The model forecasts the margin. Each race page and list also shows it as each candidate's share of the vote, beside the same polls as published (with undecided voters left out, before any allowance for pollsters' leans or past misses). Third parties take their share first: their share in the polls times 0.57, because they get less than they poll (in the Senate and governor races of 2018, 2022 and 2024 they polled 4.0% on average and got 2.7%), or 2.7% where a race has no poll. The two candidates share the rest by our margin. Tested on those races, our candidate shares on the eve of the election missed the result by 2.2 points on average, against 2.4 for the polls; half our misses were within 1.8 points. The range shown is our 80% range for the margin, in shares.
Missing together
Polls miss in the same direction across a whole election (2016 and 2020 overstated the Democrats almost everywhere). Each of the 5,000 simulated elections draws:
- one national error shared by every race: 4.9 points on the House vote, and for statewide polls the part every race of a cycle has shared, 2.9 points;
- a census division error (2.6 points), a state error (3.9) shared by the state's House districts and its Senate and governor races, and a demographic error (2.5 points in all) along the share of white non-Hispanic residents, of graduates and of Hispanic residents, so similar places move together. These sizes come from how far each state's presidential swing strayed from the national swing in the elections since 2000 (House districts take half);
- each race's own error: what is left of its uncertainty.
Senate control: 51 seats for the Democrats, 50 for the Republicans, since Vice President Vance breaks ties until January 2029. An independent counts for the party they have said they will caucus with, as Angus King and Bernie Sanders count for the Democrats. An independent who has said neither is counted apart, and if neither party reaches control without them, they hold the balance; the chances of Democratic and Republican control then add to less than 100%. This year's independent candidates, Dan Osborn (Nebraska), Todd Achilles (Idaho), Brian Bengs (South Dakota) and Seth Bodnar (Montana), have each said they will caucus with neither party, so each is counted apart. Our projection uses the same rule: the party it names as winning the Senate is the one with control in it.
Readings of the polls
Beside our projection, the maps and tables offer the polls as they stand and each past cycle's average statewide miss replayed on every race, like the "misses like 2020" lenses of the other countries.
2027 and 2028
The 2028 presidential projection has its own method page: how the 2028 forecast works, a baseline from the fundamentals blended with the head-to-head polls as they come in, with each candidate's evidence and range. The early projections for the 2028 House and Senate and the governors of 2027 and 2028 rest on the same national environment.
The Electoral College: 538 votes on the 2020 census apportionment, winner takes all except Maine and Nebraska, which also give one vote to the winner of each congressional district. A 269-269 tie goes to the House (one vote per state delegation) and is shown as its own outcome.
Publication rules
No American law bans publishing polls before or on polling day (the First Amendment protects it), so the US pages have no blackout switch and stay live until the counts begin. Networks by custom hold projections for a state until its polls close; the election-night page lists closing times.
What we do not use
No forecaster's or aggregator's averages, ratings or forecasts (RealClearPolitics, 538, Silver Bulletin, Decision Desk HQ, 270toWin, Cook, Sabato, Inside Elections, Split Ticket, VoteHub, Race to the WH, The Economist's model, betting markets).
Notes
- 2024 presidential results on the nine redrawn states' new lines are estimates; Louisiana's 2024 figures rest on state-level scaling.
- Only Fox News's past final generic ballots are held, so the national correction test has five elections.
- The 2018 House backtest starts from 2016 House margins, since no presidential result by district is held for those lines.
Figures as of . Published by Polling Forecast.




