The road to November.

Current release: polling monitor + experimental model

The catalogue contains 435 House districts, 100 Senate seats and 50 governorships. Only seats scheduled for the 2026 election count as contests. Selected original polling records have been manually reviewed. Coverage is incomplete, and the model has undergone only a limited retrospective archive validation, which found undercoverage.

Experimental estimates

When enabled, a two-party margin model weights polls by recency (30-day half-life selected on historical training data), voter population and sponsorship. A published candidate-share margin of error is conservatively converted to a margin error; an additional pollster error floor and shared election error widen uncertainty. Repeat polls from one institute have a precision cap. Decay and residual error were fitted on a historical training archive; no pollster quality ratings are inferred from the 2026 outcomes.

Intervals contain the central 80% of the model’s assumed distribution. A 60% win chance still means a 40% chance of losing under those assumptions. Poll shares include undecided voters; the model does not silently redistribute them.

Missing data and independent candidates

Grey means no estimate, not a toss-up. Historical baselines and 2026 district geography are connected. Reviewed listings cover all 506 contests, with official sources taking priority over informational Ballotpedia listings. A fully certified national register and future independent caucus affiliations remain incomplete. The data schema supports party and caucus separately. Named-candidate models preserve independent and other-party candidates. Conditional ranked-choice finalist signals remain separate from national seat allocation.

National seat projections

The overview separates current holdings from experimental baseline and named-candidate scenarios. All held seats are included. Unmodelled independent and ranked-choice contests remain unallocated; the displayed control range considers the possible allocations of those races. Correlated national error is shared across races. These scenarios are conditional on the model assumptions, including unvalidated historical baseline settings. A 50–50 Senate tie is not counted as Democratic control.

Backtesting and historical snapshots

A retrospective archive stress test now uses 1,288 training races and 226 holdout races. It is not a full release-time backtest: the archive lacks original release dates, published MOEs and population metadata. Detailed results and limitations appear below. History begins when real snapshots are saved. Today’s inputs are never presented as an earlier forecast.

Admission policy

Require an original source and verified published results. Standard admission requires field dates, population, published uncertainty and disclosed methodology. The owner-approved NYT/Siena exception permits original toplines with unavailable detailed methods, retaining null metadata and labelling any conservative working assumptions. Prefer LV over RV from the same survey, and count each survey once per race. Label sponsored polls. An unusual result alone is not grounds for deletion. Pollster accuracy adjustments will require historical evidence.

Election night and certification

Election mode switches automatically on November 3 at 18:00 US Eastern time (midnight November 4 in Oslo). Reported votes, provider calls, model projections and certification are separate fields. Official result adapters require live-source activation and coverage remains incomplete. A real-time projection model is not connected. Precinct reporting is never treated as an estimate of votes counted.