This forecast gives the odds in every House, Senate and governor race on November 3, 2026, from polls and each race's partisan lean alone. Every poll, national and local, goes into one statistical model, which works out what they say together about the national mood and about each race. The forecast then simulates the election thousands of times, allowing the polls to miss by as much as they have in past elections. Races where one party has no candidate are called outright. Below, the model is built up one piece at a time from its real data, ending with the full fit. On the forecast page, click any race to see its own forecast built up the same way.
The whole model is one line. It gives the probability of every unknown \(\theta\) (each week's national mood, each race's lean, each pollster's house effect, and so on) given every poll \(y\): the product, over all the polls, of how likely each poll's result is, times what is assumed before any polls. The fit does not pick the single most likely \(\theta\). It draws thousands of them in proportion to this probability, so the spread among the draws is the uncertainty. Each draw then becomes one simulated election. Charts show margins in points; the model itself works with the logit of the Democratic share of the two-party vote, which moves almost exactly with the margin in close races.
Start with the generic ballot polls since January 2025, which ask which party voters want in Congress. Each poll \(y_i\) measures the national mood \(\eta_t\) in the week it was taken. Bubbles are sized by sample; hover for details, or click one to open it.
Before any polls, the model assumes only that the mood drifts from week to week. The fit puts the drift at about points a week. So each new poll nudges the estimate, and older polls count for less. The line follows the polls week by week, as a running average would. After today there are no polls, so the range widens until election day: the mood can still drift.
Even a perfect poll is off by chance. Each poll's whisker is its 80% range from sampling alone, \(v_i\), set by its sample size. The line is now the full fit. of the polls miss it by more than their whiskers allow, where chance alone would give 20%. Part of the excess is that pollsters differ.
Some pollsters consistently lean one way. Comparing pollsters on the same weeks and races shows each one's house effect \(h_p\): among regular pollsters, leans furthest toward Democrats () and toward Republicans (). Here each poll moves by its pollster's house effect. Then of polls miss the line by more than their whiskers, widened by a small extra spread \(\tau\) ( points for national polls). House effects are measured against the average pollster, so a miss that every pollster shares cannot show up here.
That shared miss is real. Each bar is how far the average late poll missed in one election, from FiveThirtyEight's archive of polls since 1998. The biggest, in , overstated by points.
No amount of polling can reveal this miss, so it is not in the fit. Instead, each simulated election adds a national miss \(\varepsilon\), drawn so that 80% of the time it is within ± points (shaded), as of these elections were.
Back to the national mood. Adding the national miss \(\varepsilon\) to the fit's own uncertainty widens its range: in 80% of simulated elections the national vote lands between and . Every race carries this same miss.
Each race leans away from the national mood by \(\lambda_r\). Before its own polls, the best guess for that lean is the race's Cook PVI, plus points for an incumbent running again. Take the race: its PVI of moves the national mood's line to where the race should land.
The next steps turn to how wrong that guess could be.
How good is the PVI guess? Against the 2022 and 2024 results, 80% of races landed within points of it for the House, for the Senate and for governor.
This is uncertainty about the race itself: its candidates, its local issues. It is exactly what a race's own polls can reveal.
A race's polls can also all be wrong together, beyond the national miss: a likely voter screen that is off in one state, say. Comparing the late polls of past races with their results, and setting aside each year's national miss and the polls' own sampling error, the polls of a race missed by less than points 80% of the time. That is \(\delta_r\), part of it shared statewide.
This is uncertainty about the polls, not the race. More polls of the same race do not shrink it, because they all share it.
Back to . The faded band is its PVI guess with the range from part 1. Its polls measure the same national mood plus the race's lean and its shared miss, \(\eta_t + \lambda_r + \delta_r\), with the same house effects.
The fit weighs the two kinds of local uncertainty: the race may really differ from its PVI guess (part 1), but its polls may all be off together (part 2). So it lands between the guess and the polls, with a range that stays wide however many polls there are.
Every race poll also measures the national mood, so race polls inform the national line too, and national polls move every race.
The real model does not go step by step: it fits every piece to every poll at the same time. Each line is one draw of all the unknowns together, kept in proportion to how well it explains all polls. Past today the mood keeps drifting at its fitted pace. The highlighted draw moves every race together, because they share the national mood.
Here each line is also shifted up or down by its draw's national miss \(\varepsilon\), so the lines spread wider than the polls do: that is the miss the polls cannot see. Where a line ends on Nov 3 is one simulated election, and the histogram counts each one's House seats. Shown are of the fit's draws and of its races; over every draw, Democrats win the House of the time.