The House forecast simulates all 435 districts together, 40,000 times.

What the model produces

For each of the 434 House districts contested by a Democrat and a Republican the model produces a distribution of possible two-party margins. Those 434 distributions are simulated together 40,000 times. Alaska at-large, where the main challenger is an independent, is calculated separately and added to each simulation. From that simulation come the seat totals, the seat intervals, the probability that each party wins 218 seats, and each district's win probability. Every published figure comes from the same simulation run, identified by its run timestamp and seed.

Published figures and intervals

• Control probability: the share of simulations in which a party wins 218 or more seats. Each party's figure is counted from the simulations directly, never as one minus the other, and a small share in which neither party reaches 218 is shown on its own.
• Expected seats are the average across simulations; the median is the middle simulated result. They are labelled separately.
• 80% range: the 10th to the 90th percentile of simulated Democratic seats. The run record also holds the 50% and 95% ranges.
• District margin: the district's mean simulated two-party margin, Democrat minus Republican, in points, shown with its 80% interval.
• District win probability: from the chamber simulation's own draws. No probability is shown as 100% or 0%: above 99% it reads ">99%" and below 1% it reads "<1%".

The model does not produce candidate vote shares, so none is published.

Inputs

The model reads four things and nothing else.

• Generic-ballot polling for the current cycle, from the platform's own poll database.
• House district polling for the current cycle, from the same database.
• Each district's partisan lean: its margin at the most recent presidential election on the district lines in use this cycle, minus the national margin of that same election.
• Each district's incumbency status and whether the incumbent is serving a first term.

The model does not read fundraising, endorsements, candidate quality, advertising, local news, prediction markets, or any other forecaster's output. There is no editorial adjustment and no manual override of any district.

Method and parameters

The structural estimate. Each district starts from a relationship fitted across seven completed elections, 2012 to 2024: district margin relative to the nation as a function of partisan lean, incumbency, and first-term status. The three coefficients and the residual spread are estimated from those elections, not set by hand.

The national environment. The generic-ballot average uses a 21-day half-life, shortened to 10 days in the final 14 days before the election, and a 365-day maximum age. It weights by sample size and sample type, and halves the weight of partisan releases and of tracking polls. Every pollster is given the same default reliability score of 0.5, so no pollster is weighted above another. That average is then adjusted in two ways, both fixed in advance: it is shrunk towards the long-run average, by half at 240 days or more before the election and falling in a straight line to zero on election day, and in a midterm year a term is added for the president's party, sized by how far the election is away.

The district-poll blend. Where a district has eligible polls, their weighted average is combined with the structural estimate in proportion to the precision of each. A district poll's assumed error is SIGMA_P1² / n_eff + SIGMA_HB², where n_eff is the effective number of polls and SIGMA_HB is a floor that more polling does not remove. In version 3.1 those two constants are 8.343241 and 4.355848, estimated from every district poll in the historical record against what those districts did; in version 3.0 they were 6.0 and 3.0, set by hand. A poll is eligible when its fieldwork ended, and it was available, on or before the run date, and it is no more than 365 days old. A poll fielded earlier in the cycle is adjusted for how the national environment has moved since.

Uncertainty. Four error layers are simulated separately: a national error, drawn from a heavy-tailed distribution and scaled by each district's elasticity; a regional error; a state error; and a district error, also heavy-tailed. The first three are correlated across districts; the fourth is specific to one race.

Simulation. 40,000 draws, from a fixed seed derived from the cycle and the number of days to the election, so a given input set and date always produce the same output.

Alaska at-large (external seat)

The core House model is a Democrat-against-Republican model, and it is unchanged. In Alaska's at-large seat the main challenger to Nick Begich, the Republican, is Bill Hill, an independent, so a Democrat-minus-Republican margin does not describe the race. Alaska is therefore calculated separately, under method AK-EXT-2026.1.

The Alaska estimate uses direct Begich-against-Hill poll questions only, including a ranked-choice final-round two-way result. It uses the same poll weighting, eligibility rules and poll-error size as the core model, but no structural prior, no generic-ballot adjustment and no national swing, and it is drawn independently of the other 434 seats. Questions that list the full field are not converted into a two-way result, because that would need an invented rule for where other candidates' voters go.

In each simulation the House is counted as Democratic, Republican and independent seats. If Hill wins and the parties split 217 to 217, neither party reaches 218. The published probabilities are raw simulation frequencies, so the Democratic figure, the Republican figure and the share in which neither party reaches 218 sum to 100% only together. No Democratic probability or Democrat-minus-Republican margin is published for Alaska.

Its limitations: few direct Begich-against-Hill polls exist, five when the method was released; one recent poll carries most of the weight; and treating Alaska as independent of the national environment is a simplification.

What this model does not do

• It does not know which seats will go uncontested. Ballot composition for a future election is not settled, so every district is simulated as contested.
• It does not use any previous House result. The structural estimate reads the presidential lean, incumbency and first-term status only.
• It does not correct the spread of its own structural estimate. That spread is measured to be wider than the historical record supports, and version 3.1 leaves it unchanged.
• It does not adjust for pollster house effects. Those are computed for the UK market and are not computed for the US.
• It does not carry any adjustment for a national polling bias. Measured across seven cycles and 35 checkpoints, the model's national expectation has sat +2.24 points on the Democratic side of the result on average, with a mean absolute error of 2.80 points. In 2024 the district estimates sat about 2.5 points Democratic. No correction is applied.
• It does not treat one withheld election as a track record. The model was selected after examining 2018, 2020 and 2022; 2024 was held back and opened once, under rules written in advance. 2026 is its first fully prospective cycle.

Coverage of the district polling

Most districts are never polled. At the 17 October run, 102 of the 434 two-party districts carry an eligible poll. Where no poll exists, the district estimate is the structural estimate alone.

Earlier House figures

House figures published before 20 September 2026 come from an earlier model and are not directly comparable with current figures. The US methodology overview carries the historical data note.

Version history