US midterms 2026

Who controls Congress after Tuesday 3 November? A model of every Senate and House contest, race by race, beside what the prediction markets say.

What is being decidedAll 435 seats in the House of Representatives and 35 of the 100 seats in the Senate. Senators serve six-year terms, so only about a third come up each time; two of this year's 35 are special elections to fill vacancies. Midterms fall two years into a presidential term.
Who holds what nowRepublicans hold the House, the Senate and the presidency. Democrats need 51 Senate seats for a majority, Republicans only 50, because the Vice-President breaks ties. In the House, 218 is a majority for either side.
Our modelA statistical model fitted to every Senate and House race since 2012. It runs this election thousands of times with the kind of polling error past elections have had, and the percentages are the share of those runs each side wins. Every race moves with the national mood, so misses tend to be shared.
The marketsKalshi and Polymarket, where people bet real money on the result. A contract that pays $1 if Democrats win the Senate trading at 72 cents means the crowd puts that at about 72%. We show them beside the model; they are not an input to it.

Who ran ahead of Harris or Trump?

Looking ahead: 2028 presidential candidates and nomination markets →

Compare each party's 2026 vote share with its presidential candidate's 2024 share in the same place. House figures use the 2026 district boundaries; Senate figures use the whole state. Choose vote-share change, D–R margin change, or the separate comparison with our recorded forecast.

Example: a Democrat getting 52% where Harris got 47% ran 5 vote-share points ahead. If the Republican fell from 52% to 47%, the D–R margin shifted 10 points. Shares and margins use total votes, including other parties. This cross-election change also reflects turnout and the national mood; it is not a clean measure of candidate quality. House source: The Downballot's map-matched 2024 totals.

The Senate-versus-House option compares the Senate candidate’s share with their party’s pooled House vote across the same state in 2026. We sum votes, not district percentages. Every House district and the Senate count must be complete; uncontested, same-party or unresolved fields leave the comparison unranked. Raw completed House shares can still be shown with that warning. This measures a same-election difference, not a pure candidate effect or individual split-ticket voting. District-level Senate comparisons will require finer returns.

Partial returns can be unrepresentative: a count-so-far gap is not a final performance estimate or a nowcast. Only explicitly completed counts enter the ranking and margin summary. Winner calls, count completion and certification are separate. This reporting subset is not a national vote estimate.

Forecast accuracy on called and certified races

Called outcomes are provisional and can be retracted. Brier and log loss score whether a Democrat wins; lower is better. Compare models only within the same paired race set. Independent winners are not assigned a caucus. No chamber-control call is inferred here.

The Senate races that decide it

Democrats carry 34 seats that are not up this year; Republicans carry 31. Of the 35 on the ballot, these are the ones the model cannot call, with the market's price for each.

Model: chance the Democrat winsKalshi: what bettors pay for the Democrat to winUnder each race: the model’s expected margin, with the range it thinks covers 80% of outcomes

What if the polls are off, everywhere at once?

Polling misses tend to be shared across races. Drag to shift the Democratic margin in every modelled contest by the same amount and watch the chamber odds move. Same-party fields stay fixed; open contests stay at the extremes of the range.

Every contest

All 35 Senate seats and 435 House districts: the model's chance for the Democrat, its margin range, and the market price where one exists. Search by state, district or candidate.

Candidate lists are the reported fields at the model's input cutoff, not certified ballots. Market and model dates differ; a gap between them is a snapshot comparison, not evidence of an edge.

What the markets say

Kalshi and Polymarket prices for control of each chamber, recorded every six hours since July. Prices are not inputs to the model.

Democrats control the Senate

Democrats control the House

Balance of power

Each outcome is its own contract on each platform; the four need not sum to 100%.

How many seats?

Kalshi seat-count contracts for Democrats. Majority: 51 Senate, 218 House.

Senate race prices

ContestCandidates (provider labels)Kalshi DPolymarket DKalshi R30-day changeKalshi D, last 30 days

Governor race prices

Governors run the states and are not part of Congress; 36 are elected on the same day. Market prices only, no model.

Exact control-price observations

How this works

The model

A Bayesian model fitted in Stan to every D-versus-R contest since 2012. For the House it starts from how each district voted at the last presidential election, re-tabulated onto this year's map, plus the national generic-ballot polling. For the Senate it blends state polls with those fundamentals. Errors are shared: one national swing, one per state, then noise for each race, so the races move together the way they do in real elections.

It does not use House district polls, candidate quality, incumbency or market prices. Open contests, where the field is not a straight D-versus-R matchup (five Senate, eleven House), are outside the registered study and appear in its published range at both extremes. For the conditional scenario they are resolved by a stated rule: no Democrat on the ballot means not Democratic; otherwise the same fitted model predicts the race from its fundamentals and polls, sharing each simulated election\'s national swing. This is an unvalidated conditional scenario, not calibrated chamber-control odds. National errors are correlated across chambers; state errors are shared by races within each state. Extra correlation between neighbouring states or regions is not modelled.

Does it work?

Expanding-window test: fit on earlier cycles, score 2020, 2022 and 2024. Brier loss, lower is better; a coin flip scores 0.25. Accepted D/R contests only, so this is a race-level test, not a chamber-control track record.

The original Senate polling study, recorded daily since September

Latest poll per firm within 45 days, equal firm weights, shared election error and race-specific noise. It covers only races with recent polls and is scored against the fixed three-point polling baseline it did not beat historically.

Historical population, exclusions and Stan diagnostics
Sources and provenance