2026 Senate. The road to 51.

Polling data collected by The New York Times, reused under CC BY 4.0. DutyStation filters and aggregates it independently; The Times does not endorse this forecast. Historical results: MIT Election Data and Science Lab, U.S. Senate 1976–2024 and presidential results 1976–2024. Race schedule: U.S. Senate, plus Florida and Ohio special elections. Candidate names are observed polling matchups, not an official ballot roster.

We fit a historical regression from a state’s prior competitive Senate result and previous presidential result to its next Senate result, using a rolling 24-year training window. Current polls update that prior with sample-size and recency weights, one question per poll, and a cap on repeated polls from the same pollster. Effective polling information is calculated across pollsters so repeated releases cannot masquerade as independent evidence. Polls are limited to the selected matchup and the last 90 days, with likely-voter samples preferred.

We simulate 30,000 elections with a shared national error and race-specific uncertainty. No-poll races retain wide historical uncertainty. Expected seats include 34 continuing Democratic-caucus seats and 31 Republican seats. Republicans control a 50–50 split through the vice president.

Independent challengers are modeled separately. We do not assume they will caucus with either party; outcomes requiring their support are shown as unresolved. The scenario slider changes expected seats only.

Loading historical validation…

The historical test covers the results-only baseline, not the polling blend, national-error assumption, or chamber-control probabilities. Candidate changes, third-party vote splitting, runoffs, ranked-choice transfers, and correlated regional errors are simplified. Do not read this beta as a fully calibrated professional forecast.

The main percentage is the current election model’s probability of Senate control. The separate trend scenario asks where the displayed Democratic control probability might move through Election Day, November 3, 2026, if its recent direction continues. It is not a second forecast of the election result.

We use the latest seven consecutive daily snapshots from the same model version, with equal weight per day. We fit a straight line to their log-odds, then extend that slope from the latest published probability. Converting back to a percentage keeps the scenario within 0–100%. The dashed line continues daily through November 3 and stops there. Its endpoint is an Election Day trend scenario, recalculated with each hourly forecast update. Extending a short recent record over several weeks adds uncertainty; this assumes the same daily trend persists, with no adjustment for future events or polling.

Only recorded snapshots are shown as history. The current day’s point changes with hourly refreshes; earlier days use their last saved snapshot. The “past 7 days” change compares today with the same date one week earlier and is measured in percentage points. Chart axes are zoomed and labeled to make small changes visible; the dashed line is not an uncertainty interval.

This extrapolation has not been validated as a predictive model. Daily forecasts are related observations, and changes can reflect new polls, older polls losing weight, or model updates. No new polling evidence is assumed for the dashed line. It may reverse quickly. We suppress the scenario if the snapshot is over eight hours old, the seven-day record is incomplete or spans different model versions, or the election has arrived.

Polling refreshes every hour. A failed refresh keeps the last good forecast and exposes its original timestamp. Historical snapshots are saved daily using UTC dates. Downloads include source timestamps, model version, input hashes, selected polls, uncertainty, and simulation seed. Missing polling stays missing; no polls or trends are invented.

Forecast JSON · Selected polls · Validation JSON · Update health