The polls behind the forecast.

Provider-added records October21 through 2026-11-02 plus checked primary supplements; not an exhaustive census of all public polls. Added dates are not publication dates.

VoteHub’s generic-ballot feed stopped at October 4 when checked; checked original national releases supplement the retained provider data. A provider’s added date is kept separate from the original publication date. An unknown publication date remains unknown. LV, RV and adult samples are not interchangeable. Alternate screens and initial/leaned ballots from one survey are not separate polls.

Remaining verification gaps: The Honest Poll, Tufts, separate Ipsos mail and OnPoint national releases remain pending original-table verification. TIPP November 2 release and a newer Morning Consult week ending November 1 are reported by aggregators; original retrieval still shows the older Morning Consult release. Not substituted. Reuters/Ipsos November 2 generic ballot reported by Reuters; subgroup/table verification incomplete, so retained October26 checked release is used. VoteHub November2 additions included; provider metadata not individually primary-checked unless indicated. Unknown publications, pre-cutoff rows, duplicate ballot versions and candidate mismatches are excluded. This is not an exhaustive census.

VoteHub data is used under CC BY 4.0; VoteHub API and documentation ↗. Primary publications and corrections are linked per row.

The prior is the dated October 22 forecast, which already incorporates older polls. Only releases or provider additions after October 22 enter this update; older rows are retained for inspection. When publication is unknown, fieldwork must also end after the cutoff to avoid reusing prior evidence. This is a new PolityIQ update model, not a rerun of the prior source’s proprietary model. Changes reflect both new polling and a method change, and are not a clean measure of campaign movement.

Each firm contributes only its latest eligible release per race. Polls receive a 14-day recency half-life, square-root sample weighting capped at 1.5, and lower weights for RV/adult populations and disclosed internal or partisan sponsorship. Unknown subgroup samples remain blank and receive conservative weight. Known primary ballot versions replace conflicting aggregator versions. Candidate and race joins must be unique; mismatches are excluded.

The new national margin adjusts the House prior one-for-one and Senate priors at half strength. Race polls blend into that prior with weight W/(W+1.5), capped at 80%. Undecided shares are excluded when normalizing matched party shares. Thirty thousand seeded simulations include correlated national error (2.5 margin points), state error (1.5 points), and independent race uncertainty with a minimum 3.5-point margin standard deviation. These assumptions are transparent but not fitted to historical election error. Probabilities are experimental and not back-tested.

Senate control groups 32 Democratic holdovers and two Democratic-caucusing independent holdovers together; Republican holdovers total 31. New independent winners remain unaligned. Governor polls are archived but no complete governor forecast is produced. No model run is a live result or a winner declaration.