Methodology
how FutureBallot forecasts the 2026 elections and what its numbers mean
Start here
Inputs vary by component: past votes, candidate records, polls and national conditions. The models combine these inputs with uncertainty measured from historical errors. A win probability describes the chance that the named candidate or side wins under the model, not a predicted vote share. Forecasts combine estimated race distributions and simulated elections. Shared polling misses, sparse evidence and assumptions can all make forecasts wrong. Historical tests and known exceptions are listed in the component evidence table.
Summary
FutureBallot forecasts the 2026 Senate, governor and House races and control of each chamber. For a Senate or governor race between a Democrat and a Republican, the election-day forecast combines the race's polls, where there are any, with an estimate built without them from the state's past results, demographics, economy, candidates and money, and the national environment; a race between any other pair gets no such estimate. The House is forecast district by district from the same kinds of evidence, the professional raters' ratings and any district polls. Chamber control comes from simulated elections in which races move together, because part of any polling miss is shared. Polls are compared with the pollster's own release where it can be found; the Model health page gives how many of the poll records used for Senate and governor races have been checked.
What is forecast
FutureBallot forecasts the general election of November 3, 2026: the Senate and governor races, all 435 House seats in its totals, control of the Senate and of the House, and the number of state governorships each party holds. Races the model withholds carry no published forecast, and some House districts are counted in the totals without a race entry. Territorial governor elections are outside this forecast's coverage. State legislative chambers are forecast by a separate model, described on the state legislatures page. Race polls are collected every six hours and national polls twice a day; the forecasts are recomputed once a day.
Each polled Senate and governor race has two numbers, and so do Senate control and the count of governorships. The nowcast estimates where the race stands today among the people the polls sample; the election-day forecast estimates the result in November, allowing for how far opinion can still move, so its range is wider. They answer different questions and are never averaged. The House has an election-day forecast only.
A race with no polls has no nowcast on its page. The Senate and governor nowcast totals are simulations of their own, and such a seat enters them at its starting estimate (its last result, adjusted for the national environment) where that estimate is informative, with that estimate's measured error; those seats are listed beside the total. In the election-day totals it enters at its ground-up estimate where one exists. A seat with neither a usable poll nor a ground-up estimate enters both totals at its starting estimate, or enters neither total if that estimate is not informative.
Where the data come from
Senate and governor polls come from Wikipedia's article on each race and House district polls from its articles on each state's House elections, each poll with a citation; a few are recorded directly from a published report. A poll is missed when no listing carries it and it has not been recorded from a published report, which is how a campaign's poll reported only in the press can be missed. Excluded by name are campaigns' and party committees' own polling operations, firms found to have fabricated data and firms with no checkable record as a pollster; the pollsters page lists each with its reason. National polls of the vote for the House (the generic ballot) come from each firm's release or a report quoting it. The other inputs are past results, Census Bureau demographics, Federal Election Commission money records, economic series, presidential approval and, for the House, the ratings of the Cook Political Report, Sabato's Crystal Ball and Inside Elections.
A poll's figures are compared with the pollster's own release where it can be found: every stored share must appear in the release tied to that candidate's name. The Polls page gives each poll's status, and Model health gives how many of the poll records used for Senate and governor races have been checked. The comparison marks a poll but does not decide whether the poll is used, and a poll found to be recorded wrongly is corrected or withdrawn.
House district and national polls also pass a sanity check, with limits set from past polls. A poll with impossible shares, dates or sample size is left out. A district poll whose named-candidate result and generic-party result in the same release differ by more than sixteen points is set aside. Another poll of the district decides it: the first, in order of recording, by another firm that is not sponsored by the same side as the set-aside poll. The set-aside poll is kept if the deciding poll's margin is nearer the set-aside poll's named margin than the set-aside poll's generic margin, and left out if not. The deciding poll's fieldwork must end within sixty days of the set-aside poll's fieldwork, and the deciding poll must be recorded within sixty days of the set-aside poll's recording and by election day; if no poll meets those conditions, the set-aside poll is left out. A national poll more than ten points from the median of at least three other firms' polls ending within a week of it is left out while it stays that far from the median, and the median is recomputed as more of that week's polls are published. The poll is kept once the gap is ten points or less, and is marked rejected if the gap is still larger two weeks after its fieldwork. Senate and governor polls do not pass this check; in their trend model a poll far from the others counts for less instead.
What a margin is
A margin is the leading side's lead in percentage points, shown as D+4.3 (a Democrat ahead by 4.3 points), R+2.0, or, where the two candidates are not a Democrat and a Republican, the leader's name. What it is a share of differs between the two models. In a Senate or governor race it is a share of all votes cast, so votes for minor candidates and write-ins count in the total. In a House district, and in the national environment, it is a share of the Democratic and Republican vote alone; a House seat forecast from polls of its own contest uses that contest's two-candidate margin.
Polls enter both models as published: in a candidate poll the margin is one candidate's published share minus the other's, and in a generic-ballot poll it is the Democratic share minus the Republican share; undecided voters are left unallocated and nothing is renormalized, so poll margins run smaller than results. Undecided voters are not divided between the candidates in any forecast and add no term of their own to any range. For the poll-based Senate and governor forecasts, the stretch described below carries poll margins onto the scale of results. House district polls enter at their published margins, read against the district's lean after the national environment is taken off, with a constant correction for sponsor lean, and generic-ballot polls with a constant correction for the generic ballot's historical error; unlike the Senate and governor forecasts, no fitted slope stretches these margins.
A Senate or governor forecast covers the Democrat and the Republican, unless a third candidate with at least three polls has a higher average poll share than one of them, or the ballot lacks one of them; the forecast then covers the two candidates with the highest average poll share. A race decided by a ranked-choice count or an all-party runoff is forecast for the round and the pair that decide it; a runoff between party nominees is read at its November round. The race's page explains which applies.
How the polls of a race are combined
A poll is usable if it is published by the forecast date and not withdrawn or superseded, covers the whole sample, gives a share to both candidates, and is not from an excluded firm; of one firm's overlapping tracking polls only the newest counts, and a release's likely-voter result is preferred to registered voters, and those to adults.
With four or more usable polls, a trend model reads them as noisy measurements of one line of support that moves a little each day. A poll counts for more the larger its sample and for less the older it is or the further it sits from the others. Each pollster has its own measured error and its own lean against the other firms polling the race (its house effect), held near zero for a firm with few polls. With one to three usable polls, their average is combined with the starting estimate, each weighted by the inverse of its error as measured on past elections. When such a race is between a Democrat and a Republican, it is then also blended with the ground-up estimate described below, so its forecast draws twice on estimates made without its polls; the adjustments page gives the effect. A race with none is forecast from the ground-up estimate alone (below) when its two candidates are a Democrat and a Republican; a race between any other pair with no polls has no published forecast.
Final polling averages have historically understated how lopsided races end up, so the lead of every poll-based Senate and governor forecast, nowcast and election-day forecast alike, is stretched by a factor fitted on past elections (1.07 for 2026: a ten-point lead becomes 10.7). The stretch comes before the blend with the ground-up estimate, which is not stretched; nor are a race forecast without polls and the House. Apart from the generic-ballot correction in the national environment, no correction is made for a past lean of the polls or of the forecasts toward either party; that lean is measured on the Accuracy page.
After a verified scandal about a candidate themselves (a criminal charge or conviction, a specific misconduct allegation or official finding, or a documented false claim about their own record, reported by two outlets doing their own reporting), the election-day margin moves three points against that candidate. In a Senate or governor race the shift lasts only while fewer than three polls have been taken since the event, because later polls already show most of its effect; in a House race the shift applies from the event on. FutureBallot lists each event by hand with its sources. The two outlets must be different news organizations, each with its own reporting, not opinion, a party or campaign release or a rewrite of another outlet's story; trouble of a running mate, relative or staffer does not count, and a report found by the daily news scan is a lead to check, never an entry by itself. The shift leaves the range and the nowcast unchanged. The adjustments page gives the full rule, its evidence and shortfall, and each event with its sources.
The estimates that do not use polls
Each Senate and governor race between a Democrat and a Republican also gets a ground-up estimate: a model fitted on past elections that predicts the state's margin from its past results, demographics and economy, the candidates' incumbency, experience and money, and the national environment. The model is fitted on the two-party vote and scaled to all votes by the share taken in the state's last election for the office by the Democratic and Republican candidates together with any other candidate who finished first or second. A race between any other pair has none, so its election-day forecast is not blended with one.
The election-day forecast's center is an average of the stretched poll-based forecast and this estimate, each weighted by the inverse of its measured variance, so the estimate counts for more where polls are few or the election is further off. The poll-based forecast's whole distribution is then moved to that center without being narrowed. Averaging two estimates narrows the range less the more their errors move together. These two tend to miss in the same direction, so a range narrowed as if they were independent would claim more certainty than past elections support. A Democrat-Republican race with no polls takes the estimate's own measured error as its range. No nowcast includes the ground-up estimate.
The national environment
The national environment is the expected margin of the national House vote. A weighted average of generic-ballot polls, cleared of each firm's lean and corrected by the generic ballot's average past miss, is combined with a regression on whether the year is a midterm for the president's party, presidential approval, income growth and the previous House vote. Each poll's weight in that average comes from its measured precision (from its sample size plus an error that no sample size removes, so a very large poll counts only a few times as much as an ordinary one) and from its age, a poll's weight halving every fourteen days. The national environment feeds the House forecast and every ground-up estimate; the national environment page shows it week by week.
The House
Each district's lean, its margin relative to the national vote, is predicted by a model fitted on past House elections. The model's inputs are the district's past House lean where its lines are unchanged, the presidential lean within its current lines, its demographic, religious and economic makeup, whether the seat is open, each nominee's highest elected office held before the election (an incumbent coded at the highest value), and net outside, party and PAC money. Adding the national environment gives its expected margin.
The lean then moves toward the professional raters. Each rating by Cook, Sabato and Inside Elections is scored toward its party: three for Safe, two for Likely, one for Lean, a half for Tilt, zero for Toss-up. The average is rounded to a whole number, halves away from zero, which puts the district in one of seven categories (Safe, Likely or Lean for either party, or Toss-up; an average of a half counts as Lean), and the lean moves about 15% of the way toward the average past result of districts in that category. An unrated district keeps its lean; the raters page shows each rating.
A district's own polls then move its lean. Each pollster's latest poll counts once, corrected for the average lean of its kind of sponsor (nonpartisan, Democratic or Republican). The poll's weight depends on when its fieldwork ended: full weight in the sixty days before election day, falling in a straight line to a half at 210 days before it and a quarter earlier still, a schedule set by judgement rather than fitted. The lean and the polls' average are combined, each weighted by the inverse of its error variance: for the polls, one poll's variance divided by the sum of the polls' age weights, plus a variance that all of a year's district polls share. The two parts overlap, so this is a weighting tested on past elections, not an exact split of the error; an exact split scored no better. As in the Senate and governor blend, the range is not narrowed.
A district where only one of the two parties has a nominee is counted as won by that party; any other district without a Democrat-against-Republican contest on record is forecast from its lean, except a seat forecast from polls of its own contest, as its page explains. Districts are read on their current lines; in Alabama, Florida, Louisiana, Ohio, Tennessee and Utah the presidential vote on those lines is estimated from census blocks, to within about one to two points, and a district on new lines carries the larger error measured for such districts.
How uncertainty is measured
Every range is built from separate errors, each measured on past elections, and most allow large misses more often than a bell curve would (they are heavy-tailed). A Senate or governor race's range combines:
- the uncertainty of the polling trend itself (for a race with one to three polls, of the combined estimate of its polls and its starting estimate);
- a polling error shared by every race in the country;
- an error particular to the race;
- for the election-day forecast, the movement still possible before election day.
The weights of its ground-up blend come from:
- the ground-up estimate's error, measured separately by the size of the margin it predicts: under ten points, from ten to under twenty, and twenty or more;
- the poll-based forecast's own total error.
A House district's range combines:
- the error of the national environment, including the movement still possible before election day, shared by every district;
- an error shared by the districts of one state;
- the district's own error, sized by its type: an incumbent on unchanged lines, an open seat, or new lines.
A district's polls are weighted against its lean by:
- the error of the district's lean;
- the error of one district poll;
- an error shared by all of a year's district polls.
The Senate and governor totals come from simulated elections that split each race's error into parts shared with other races:
- a national error;
- an error for each Census region;
- an error for each office;
- each race's own error.
When the model is fitted on past elections, the region and office parts come out at or near zero, so races move together mainly through the national part. This split decides only which races move together. Each race's simulated margins are then matched by rank to its own published forecast, so a race's range and chance in the simulation are the ones on its page. The House totals come from one simulation that draws the district terms together. For the Senate and governor totals, the fitted sizes of the national part and of each race's own part are on the parameters page.
Chances, ranges and control
A chance always names the side it favors. A 70% chance means that side wins about seven times in ten elections like this one, and loses about three. Ratings put the favored side's chance, to the nearest whole percent, into words: Toss-up below 60%, Lean from 60% to 84%, Likely from 85% to 94%, Safe at 95% or more. The margin of a Senate or governor race forecast from polls, or of a House race, is shown with a 90% range, which should hold the result nine times in ten, while a state-legislature district page shows an 80% margin range. A party's seat total in a current chamber forecast, and the margin of a Senate or governor race forecast without polls, have 80% ranges, which should hold the result eight times in ten; an older state-legislature forecast can show a labeled 90% seat range. The House model's tests and its record on the Accuracy page use a 90% seat range instead.
Control means a majority. The House majority is 218 of 435 seats; the Democratic and Republican chances are counted separately, and where seats won by others leave both short, that outcome is shown as no majority for either party, with its own chance, rather than assigned to a side. The Senate majority is fifty-one seats, or fifty with the Vice President's tie-breaking vote; the Vice President is a Republican, so a fifty-fifty Senate counts as Republican control. A majority of governorships is more than twenty-five. Where an independent's choice of caucus is not settled, the Democrats' and the Republicans' chances of Senate control are each given as a range, one the complement of the other: one end counts each such seat with the Republicans, the other with the Democrats.
What the forecast cannot tell you
Whether the polling industry will miss in one direction all at once: that is allowed for as uncertainty measured on past elections, not detected in this year's polls. Who will turn out. And much about a race with little evidence, whose forecast rests mainly on the estimates built without polls and whose range is wide accordingly.
How the model is tested
Historical forecasts aim to respect the information cutoff for each past date (walk-forward testing). Some economic series use later revisions and some poll release dates are inferred; these are reconstruction limits, not a live track record. Test polls dated or recorded after a forecast are planted to confirm that it cannot see them. Most historical polls have no recorded release date, so a poll counts as public two days after its fieldwork ended (on its last day of fieldwork in the House tests).
The Senate and governor model is tested on the five elections of 2016 to 2024. The standard for a change is that it scores better across them by more than a set minimum, under a rule written before the test; the testing document gives the rule. A new House model must, on 2016 to 2022, have a lower Brier score (the average squared gap between its chances and what happened) and a lower seat error (how far the Democrats' expected seat total is from the seats they won). The new model must also hold the Democrats' seat total in its 90% seat range in at least three of the four elections, and pass a single 2024 check.
Four components in use do not meet that standard: the House rater blend, the House poll-age discount, the Senate and governor scandal rule, and the way a thin race's poll-average error is counted. The adjustments page gives, for each, its evidence, how it falls short and why it is used; each is scored again on the 2026 results. The rule that counts a House district with only one party's nominee as won by that party has no test: past House tests assigned those seats from their actual winners.
Five elections give only weak evidence, and those of 2018 to 2024 have been used to compare many versions of the model, so they are not an untouched test. Chamber control has one outcome per chamber in each election, so its tests rest on very few outcomes. How far opinion can still move before election day is tested only on forecasts made on the eve of past elections, where that term is near zero, so the width of ranges further out is not tested. The House forecast is published as a research model, to be read as provisional, because no person outside the project has checked its model. Each day's forecasts are archived with an outside timestamp for scoring against the results.
Questions
Does the forecast model turnout, early voting or mail voting?
No. Turnout enters only indirectly: through the pollsters' likely-voter screens; through past results, demographics and the national environment; and through error ranges measured on elections whose turnouts varied; the adjustments page compares House models built on turnout.
How accurate has it been?
The 90% ranges held the result in 93% of the election-day forecasts for 256 past Senate and governor races, each forecast as it would have stood at the time. The sample is every past race for which the published model made a forecast from at least one poll, where one of the two candidates that forecast covered won and the two took at least 80% of the vote. Races outside that sample, such as one won by a third candidate, are not scored. The Accuracy page has the full record, the House included.
Does a language model write any of these numbers?
No language model produces a forecast, a chance or any other number on FutureBallot. Every forecast, chance and other computed figure comes from the forecast code run on stored data (the poll shares, results and other inputs are stored records), and every run can be repeated exactly.
The glossary defines the technical terms this page uses.
Reading different ranges
Compare intervals using both their coverage level and the quantity they describe. Senate, governor and House race pages show 90% ranges for margins, except that a Senate or governor race forecast without polls shows an 80% range; state-legislature district pages show 80% margin ranges. Current chamber seat-count ranges use 80% coverage. Older state-legislature forecasts can show a labeled 90% seat range. For the same forecast, a higher coverage level includes more simulated outcomes. None of these intervals is a candidate's chance of winning; Senate alignment bounds describe different assumptions and are not coverage intervals.
Technical documents
- The statistical model
Every distribution, every prior and every error term, with the assumption behind each stated beside it.
- Every adjustment and its evidence
Each step between the raw polls and a published number: what it does, the values it uses, and the evidence for it, including the steps tested and not used.
- Parameters in force
The fitted numbers this page reads from the artifacts the forecast uses, what each is estimated from, and how well it fits.
- How it is tested
The walk-forward protocol, the leakage guards, and what counts as evidence that a change is an improvement rather than a coincidence.
Every dated change to the method, with its evidence, is on the change log: How the model has changed.