To the left, Republicans beat their polls; to the right, Democrats do. The error is added to the margin of each race, in points of margin between the two candidates.
Active scenario:
With this error fixed, what remains is the uncertainty of each state, race and district.
House: the 435 districts, 2026 boundaries (Census Bureau).
Nebraska, Montana, Idaho and South Dakota (Senate): the Republican faces an independent, shown in blue like the Democrats.
Races where the leader is no longer the forecast’s favourite.
| Race | Change | Forecast | With the error |
|---|
With the model: the chosen error is added to the forecast margin of each race, in points of margin for the Senate and governors, in half points of vote share in each House district. With the national error fixed, each simulation only draws the remaining uncertainty: the demographic state factor and the error of each race and district.
Polls only, no model: plain mean of the latest wave of each pollster, fieldwork ended within 30 days (60 if none), with no correction at all. Races without a recent poll keep the model’s margin (dotted outline). For the House, as most districts have no polls: the district’s 2024 presidential result, moved by the gap between the raw generic-ballot average (D +7.9) and the 2024 national result (R +1.5); district polls where there are some.
Historical scenarios. Presidential years: result minus final poll average in each state (FiveThirtyEight for 2012, 2016 and 2020; plain mean of the last polls in 2024), the year’s mean error for states without polls. Midterms: final generic-ballot average versus the national House vote.
Examples of correlation between states: Michigan and Wisconsin 0.89, Georgia and North Carolina 0.80, Nebraska and Kansas 0.97, Texas and Arizona 0.67.
Project director: William Thay
Deputy project director: Sean Scull / Pierre Clairé
Political and data analysis: Matthieu Hocque / Augustin Bataille
Data analysis, mathematics and validation: Florian Gérard-Mercier
© 2026 Le Millénaire SAS, 47 rue de Vivienne, 75002 Paris. All rights reserved. The simulators, their results, maps and texts belong to Le Millénaire. Any reproduction, even partial, is prohibited unless Le Millénaire is explicitly cited as the source: “Source: Le Millénaire, forecast.lemillenaire.org”.
The public data used remain under their own terms: U.S. Census Bureau, Association of Statisticians of American Religious Bodies (2020 Religion Census), FiveThirtyEight, MIT Election Data and Science Lab, Federal Election Commission, Clerk of the House.