The Senate forecast simulates all 35 contests, and the 65 seats not on the ballot, 40,000 times.

How a race is forecast

For each contest with a named Democrat and a named Republican, the model forecasts the margin between the two in November. It combines two sources of evidence.

The first is a structural estimate, built from the state's presidential lean, whether a sitting senator is on the ballot, whether that senator was appointed rather than elected, and the national environment. The national environment comes from the generic congressional ballot, adjusted by the historical tendency of the president's party to lose ground in a midterm. The structural coefficients are estimated from the 2012 to 2024 Senate elections.

The second is race polling. Only polls of the two named candidates feed a race estimate. Each poll is weighted by its age, sample size, voter screen and sponsorship, and adjusted for movement in the national generic ballot since its fieldwork. The two sources are blended in proportion to their estimated precision, so a race with many recent polls leans on the polls and a race with none rests on the structural estimate. The weight the blend can give to polls is capped at about 0.72.

Seats where an independent leads or no Democrat is on the ballot

In Idaho, Montana, Nebraska and South Dakota no Democrat is on the ballot, or full-field polling places an independent among the two leading candidates. The Democrat-against-Republican model does not apply there. These four seats use a separate external-seat model, C-EXT-2026.1 stage one: the leading candidate's chance of winning is derived from the margin between the two leading candidates in full-field polling, with the frozen polling-error components of the main model. It has no structural estimate and no historical validation: only three comparable past contests exist. The four seats are drawn independently of each other and of the rest of the chamber; that independence is a simplification.

An independent winner is counted as caucusing with neither party, following each candidate's own dated statement. Nebraska matters most: if its independent candidate were counted with the Democrats, the Democratic-majority probability on the published run would be about 5.2 points higher. He has said he will caucus with neither party.

Georgia

Georgia requires a majority in November or a runoff on 1 December. Until the certified ballot fixes the share of the vote going to other candidates, the forecast assumes it is 4.2%, the conservative end of a range set before the forecast was run. In each simulated election where the leading named candidate has 50% or less, the seat goes to a runoff. The runoff itself is not modelled: the party that led in November wins it with probability 6 in 9 (about 0.67), the rate at which the November leader won the nine comparable Senate runoffs from 1992 to 2022. On the published run, moving the other-candidate share to the other end of the range (1.9%) changes the Democratic-majority probability by about 0.3 points.

How uncertainty is handled

Polling error is not independent across races. The model carries three components: a national polling error common to every race; a race-level error shared by that race's polls; and a sampling error that shrinks with the effective number of weighted poll observations. These were estimated on 720 race-checkpoints from the 1998 to 2024 Senate elections and are fixed for 2026.

The chamber forecast is produced by simulation. Each of 40,000 simulated elections draws a national shock, regional shocks, state shocks, race shocks and a shared polling error, so a national miss moves every race together. The four external seats are then drawn and combined with each simulated result. Chamber figures come from counting seats across the simulations, never from adding up race probabilities.

What the chamber figures mean

The forecast publishes four chamber outcomes, each defined by seat count alone; together they cover every result.

• Democrats win 51 or more seats: Democrats and the independents who caucus with them hold 51 seats or more.
• Republicans win 51 or more seats.
• Republicans win exactly 50 seats: Democrats therefore hold 50 or fewer. The Vice-President, a Republican, breaks tied votes.
• Independents hold the balance: neither party reaches 51 and Republicans have fewer than 50. No organisation rule is assumed: who organises the chamber then depends on votes and agreements the model does not forecast.

Each figure is rounded to one decimal place on its own, so the four may not add to exactly 100%; where they do not, the page says so. The seat headline is the expected composition of the chamber, averages across the simulations rather than medians, rounded to whole seats. No race probability is shown as 100% or 0%: above 99% it reads ">99%" and below 1% it reads "<1%".

Limits of this forecast

• The historical results are development evidence. 2026 is the first cycle in which this methodology is used prospectively.
• On the historical record there was no statistically significant difference between this model and the previous Senate model.
• The blend caps the weight on polls at about 0.72.
• All pollsters are treated as equally reliable, and there is a single poll source.
• The error components are applied at the point 46 days before the election.
• 7 of the 35 races have no qualifying polls and rest on the structural estimate alone.
• Alaska's ranked-choice transfers are not modelled.
• The external-seat model has no structural estimate and no historical validation.
• The Georgia runoff share is held at a conservative fixed value until the certified ballot is known.

Data

Race polls and the generic ballot come from the New York Times polling dataset, processed through the versioned live-source adapter. The adapter resolves each candidate's identity against the certified race set and applies the frozen named-pair rule: a poll enters a race estimate only when it tests the two named nominees. It keeps every excluded row, with the reason for excluding it. Candidate and ballot information comes from state election authorities where their records could be read, and from documented secondary sources otherwise; the source is recorded for every contest.

Version history