In short
- Every House, Senate and governor race gets a forecast margin built from three signals: the district's partisan lean, expert ratings and public polls.
- The model then simulates the election 20,000 times. Errors are shared across races, so a polling miss in one state usually shows up in similar states as well.
- A party's chance of winning a chamber is the share of those simulations in which it reaches a majority. On election night, official vote counts gradually replace the pre-election forecast.
Current forecast: loading…
The model
Each race: three signals
Margins are measured as the Democratic (or left) candidate minus the Republican (or right) candidate, in points. Each race combines three estimates. Each one is weighted by how accurate it has been in past elections. The combined uncertainty is widened by 10%, because the signals overlap: raters read the polls.
- Fundamentals. The race starts from the 2024 presidential margin within the 2026 district lines. It is shifted by how far this year's national environment differs from 2024 (R+1.5), scaled by how swingy the seat is. Incumbents add 3 points in House races, 2.5 in Senate races and 5 in governor races. An incumbent's own last result also counts, with a weight of 0.2 in the House, 0.25 in the Senate and 0.5 for governors. That is why popular incumbents like Maine's Susan Collins or Vermont's Phil Scott are not read off their state's lean alone.
- Expert ratings. The average of Cook Political Report, Sabato's Crystal Ball and Inside Elections is converted to a margin. Toss-up counts as 0 points, Tilt 2.5, Lean 6, Likely 11 and Solid 20. These values were set against Cook's own record from 1984 to 2022 (see "How it was checked"). Seats the outlets do not rate rely on fundamentals instead.
- Polls. The poll average gives a poll half its weight after 21 days and drops polls older than 120 days. Larger samples count more, up to 2,000 respondents. Likely-voter polls count fully, registered-voter polls 85% and polls of all adults 60%. Polls paid for by a party or campaign are shifted 2 points against their sponsor and weighted down. Repeat polls by the same pollster count less each time. On top of sampling error, the average is assumed to miss by about 4.5 points, which is roughly the historical miss of Senate and governor poll averages.
The national environment
The national House vote is estimated as 75% the generic-ballot poll average and 25% midterm fundamentals. The fundamentals are the 2024 House vote plus the average swing against the president's party in the 2006–2022 midterms. That national number can miss in either direction. Its error follows a Student-t distribution (5 degrees of freedom, standard deviation 3 points), which matches how far final generic-ballot averages have missed in past midterms and allows for occasional large misses.
Simulating the election
Each of the 20,000 simulated elections draws one national error, one error for each of 8 regions (SD 1.5 points), one for each state (SD 1.8), and a fat-tailed error for each race. Races in the same state or region therefore move together. That shared movement is what gives the seat totals an honest width.
- House: control takes 218 of 435 seats.
- Senate: 35 seats are up. Democrats need 51 seats, because Vice President Vance breaks a 50–50 tie for Republicans, so Republicans need 50. If an independent with no stated caucus holds the balance, the forecast reports that separately as "an independent decides".
- Governors: 36 races. The count of Democratic governors includes the 14 governorships not up this year.
- Special rules:
- Georgia requires a majority, with a runoff on 1 Dec.
- Louisiana's House vote on 3 Nov is an all-candidate primary, with a runoff on 12 Dec.
- Maine uses ranked-choice voting for its US Senate and House races (not for governor, which its constitution decides by plurality), and Alaska uses top-four ranked choice. For those ranked-choice races, poll respondents who choose minor candidates are redistributed to the finalists, using Alaska's 2022 transfers as the guide.
- California and Washington run top-two generals, so some races have two candidates from the same party.
- Ballot measures are deliberately light. The chance of passing comes from the latest poll's yes share, less 2 points for the usual late drift toward no. Unpolled measures sit at 50%, and no measure goes above 99% or below 1%.
Election night
From 6 AM ET on Election Day, the poll and rating refreshes stop and the forecast follows the official count. Results come only from state and county election offices: live feeds where a state publishes one, and official downloads or pages entered by hand where it does not. No media outlet's private feed is used.
- Where counties report by vote type (early, mail, Election Day, provisional), each type's outstanding vote is projected from that type's own counted margin. The share of each type is learned from counties that have finished counting. This handles the "red mirage" and "blue shift" from the data rather than from assumptions about who votes how.
- Where only statewide totals exist, the uncertainty is sized from past count nights. It is very wide while little has been counted, because the first counties to report can be very unrepresentative.
- What is learned about one race also updates related races, through the same shared errors the pre-election model uses.
- A race is marked as projected only when the model gives the leader at least 99.9% with 60% of the expected vote counted, or at least 99% with 95% counted. Calls entered by hand from official sources override the model.
How it was checked
Do the probabilities mean what they say?
There is no 2026 result to score yet, so race probabilities were compared with the longest public track record covering every race type: how often Cook Political Report's favorites won, 1984–2022.
| Rating | Races | Model: average chance for the favorite | Cook favorites' historical win rate |
|---|---|---|---|
| Solid | 113 | 99.5% | 99.97% |
| Likely | 37 | 95.4% | 97.1% |
| Lean | 36 | 84.8% | 93.7% |
| Toss-up | 17 | 60.8% | D won 48% |
The model is deliberately less sure than the raters about "Lean" seats. With a D+8 national environment, fundamentals make many Republican-held "Lean" seats closer than the ratings suggest, and 145 districts are on new maps with no election history.
Election-night count model
The count model was rehearsed on real 2024 count nights, using only 2020 and 2022 information as its baseline:
- 1,489 snapshots of 15 statewide contests in Virginia, Georgia and North Carolina, from the official results pages as archived by the Wayback Machine.
- Real 2024 county results by vote type from 8 states, released county by county on a simulated night.
Across both rehearsals the projection rules would have made 52 projections, with none wrong. In the county-level rehearsal, with 20–60% of the vote counted, the model's estimate was 2–3 points off the final margin, where the raw count was 7–9 points off.
The pre-election model has not been back-tested against past midterms, so no back-test numbers are shown here.
Data sources
- Races, candidates, incumbents and ratings: Wikipedia's dated tables of the Cook Political Report, Sabato's Crystal Ball and Inside Elections ratings.
- District partisan lean: 2024 presidential results within the 2026 district lines, from The Downballot. National 2024 results come from the Federal Election Commission.
- Polls: public polls as published, each linked from the Polls tab. The averages shown there for comparison (Decision Desk HQ, FiftyPlusOne, NBC News, RealClearPolitics, Silver Bulletin) are those outlets' own published numbers.
- Election-night results: official state and county election offices only.
- Maps: U.S. Census Bureau boundaries via us-atlas.
- Headlines: RSS feeds from NPR, Ballotpedia News, Votebeat, The Christian Science Monitor, the North Carolina State Board of Elections, and Google News search. Only the headline, source, time and link are shown, never article text.
Data refreshes run every hour until Election Day morning: polls, ratings, the generic ballot, presidential approval and special elections. After each refresh the forecast is re-run. Headlines are checked every 15 minutes.
Downloads
CSV files of the current forecast, generated in your browser from the same data as the charts.
Loading the current forecast…