Methodology · 2026

How the forecast works

Model v15 · last changed Oct 26 · See every change

Bellwether publishes a win probability for every House, Senate and governor race on the 2026 ballot, and a seat distribution for each chamber. Each race blends fundamentals with any polling average, and each chamber is simulated 20,000 times with errors correlated across races. Where certified ballot rules leave only one party eligible, that party's win is certain.

Two estimates per race

Each is a margin with an uncertainty, and the uncertainty sets its weight in the blend.

Fundamentals: Where the race should be before any poll.
Partisan lean, incumbency and the candidate's past overperformance, fitted separately for the House, the Senate and governors. An appointed incumbent counts as running for an open seat.
House. Each side's share of two-party FEC receipts so far this cycle; a race where neither side has filed reads as neutral. A returning incumbent in a redrawn district keeps their 2024 overperformance, tapered by how much the district changed.
Scandals. A House nominee with a publicly known scandal (criminal charges, credible misconduct allegations, an ethics investigation, a fabricated biography, or statements their own party's leaders condemned) loses about four points of margin from the day it became public. The list is reviewed by hand; the term is fitted on 51 races, 2014–2024.
Uncertainty. Widest for governors. The Senate's is measured on its five most recent cycles, which track presidential lean more closely than older ones: a 12-point favorite loses about 11 times in 100.
Polling: A precision-weighted average of the race's public polls.
Weight. Sample size and age. A poll counts in full for 45 days, then fades out by day 75; in a House district, where polls are rare, it counts until day 240 and its age widens its uncertainty instead.
Floor. No average is treated as more accurate than 5.1 points for the House, 5.6 for the Senate or 4.3 for governors. The House and Senate floors are what 2018–2022 averages still missed by in the final two months.
House effects. Each pollster's lean against others in the same races, shrunk toward its historical lean, is removed.
One per firm. Only each pollster's newest poll counts, with one question per matchup: a final round first, then likely voters, registered voters, adults.
Undecided. Split in proportion to each side's support. No rule for which way they break has beaten that since 2014. Race pages show the raw shares.
Sponsored polls. Counted with their fitted lean toward the sponsor removed, and no further discount.
Fitted lean toward the sponsor, in points
House Senate & governor
Campaign's own poll 6.4 5.6
Partisan-sponsored public poll 4.2 3.0
Left out. A sponsored poll whose side cannot be read, and any firm the ratings ban or Bellwether excludes. See which firms.
Quality scores. Carried, but the fit weights them at zero.
Missouri. House polls begun before September 4, 2026 wait for review against the restored 2022 map.

Fitted lean toward the sponsor, in points
	House	Senate & governor
Campaign's own poll	6.4	5.6
Partisan-sponsored public poll	4.2	3.0

Expert ratings are carried through the pipeline but not used: there is no licensed archive of past ratings to fit them against.

Ranked-choice states

In Alaska and Maine, a poll's final round is used when it publishes one, never averaged with first choices. A first-choice ballot with two candidates of one party is read as each side's leader, since votes for a trailing candidate do not all transfer. Transfers themselves are not modeled.

How they are combined

By inverse variance, so a precise estimate counts for more than a vague one. Two rules constrain the result.

Conflict widens: Estimates that disagree beyond their stated uncertainty widen the result.
No new confidence: The blend is never more confident than its most confident estimate.
When that binds, the uncertainty widens; the margin does not move.

The national environment

Bellwether's own generic-ballot average, read as a shift from the 2024 national House margin.

No forecast of drift: Today's reading, carried forward unchanged.
Its uncertainty grows with the days remaining, on a curve fitted across fourteen cycles. About five points of movement from August to Election Day is normal.
By office: The national shift is added one for one, in the frame each fit was built on.
House. A district's relative lean plus the national margin.
Senate and governors. A state's margin relative to the nation's in the preceding presidential election. In midterm governor races, the party out of the White House also runs about two points behind what the national shift alone implies.
Approval: Not a model input: an approval term did not beat the generic ballot alone.

The published polling averages

The Environment tab publishes the generic ballot and presidential approval, under the same rules for which polls count.

One reading per survey: The broadest population it published: adults, then registered, then likely voters.
The likely-voter series uses only likely-voter samples, so the two can move apart.
What is left out: Excluded firms, including polls already stored.
See which firms are left out.
Approval also drops internal and partisan polls. The ballot keeps a sponsored poll only when its side is known, with that side's lean removed.
Presidential approval: VoteHub's polls from the last 90 days, fitted with a local linear curve.
30- and 90-day windows blended equally. Larger samples weigh more, and house effects are removed, shrunk toward zero.
The range is where nine in ten new polls should land. No number without three polls from three pollsters; stale after 14 days without new fieldwork; thin below 4.5 effective pollsters.
Generic ballot: New York Times and VoteHub polls, a survey carried by both counted once.
Each firm's newest poll, weighted by sample size and age, with house effects removed against the scorecard's prior. The range is the 80% interval; stale after 30 days.
The forecast reads its own blend of these polls, not this published cut.
Against past presidents: Approval beside every president since Truman, at the same day of the term.
FiveThirtyEight's historical averages are built differently, so compare level and direction, not exact points.

From one seat to a chamber

20,000 draws

Every run draws a national, a state and a local error, so seats move together. That is why a chamber's range is wide even when most races look safe.

Errors move together: A polling miss in one direction tends to be a miss everywhere.
The national term is the environment's uncertainty; the share reaching a House district is fitted.
The levels split a seat's uncertainty without adding to it, so each seat's simulated win rate matches its published probability.
On the page: A chamber's hundred dots are a hundred runs, sampled by percentile.
Governorships are reported as a count out of fifty.

Independent of markets

The model never sees a market price. Prediction markets are neither used in the forecast nor shown with it.

What is fitted, and what is not

Every fitted constant comes from a script that records its standard error and provenance. The table names what is not fitted rather than letting a precise-looking number imply it.

Layer	Status
House fundamentals	Fitted on 2,223 contested races across six cycles, 2014–2024.Held out by state and by cycle, with each cycle's national environment estimated separately. The fundraising term entered only after it improved the held-out error.
Senate fundamentals	Fitted on 695 contested races across 25 cycles, 1976–2024.Scored on 17 held-out cycles, 1992–2024. The 9.8-point uncertainty combines the five most recent cycles' held-out spread (8.5) with the shared national error.
Governor fundamentals	Fitted on 233 contested races, 2006–2024.Scored on 6 held-out cycles, 2014–2024.
Polling precision	Fitted on 2,542 House and Senate polls across 480 resolved races.Plus 1,257 governor polls for that office's floor. Covers sample size, age, house effects, the quality-score weight and the accuracy floor; the House and Senate floors are fitted on the size of the miss, the governor's still on the winner call. The 45–75-day fade is a policy choice, with everything else fitted around it.
Sponsored polls	Lean fitted on 309 internal and 303 partisan-sponsored polls, 2018–2024.Against official results, net of same-race nonpartisan polls. Held out on 2024, where the lean is smaller but in the same direction.
Pollster accuracy and herding	Computed, not fitted, and not in the weights.Each firm's House and Senate polls from the final 21 days of 2018–2024, scored for error, bias and closeness to the field. Published on the pollster scorecard.
National environment	Uncertainty fitted across fourteen cycles.The central estimate is today's generic ballot, not a predicted shift.
Error hierarchy	National, state and local shares fitted from presidential residuals.The governor pass-through is constrained to one, selected in its own walk-forward.
Expert ratings	Unfitted and unused.

Where the data comes from

Polls: The New York Times and VoteHub public feeds. A poll carried by both is counted once.
Pollster ratings: Silver Bulletin's January 2026 edition, used with attribution: the ratings, each firm's signed historical lean, and the resolved polls behind them.
Historical polls: FiveThirtyEight's archive under CC BY 4.0, joined to official returns.
Historical approval: FiveThirtyEight's archived daily averages, Truman through Biden.
Races and calendars: Reviewed static data.
Candidates and money: FEC filings.
Results: Official canvasses for past results; official state results pages for 2026, at the most detailed level each state publishes.
District profiles: U.S. Census Bureau district-to-county relationships and 2024 county population estimates. Approximate in states that redrew their lines for 2026.

What an empty section means

Missing data is shown as missing, not filled in. It is expected early in a cycle.

Empty: No reviewed source row, poll or model output yet.
It does not mean the value is zero.
Primaries: Read from each state's official results file.
Results pending until that file has been read; reporting while it is still incomplete.
No usable input: Drawn at even odds in the simulation.
The chamber page says how many such races there are, and a chamber with no simulation says so instead of drawing an empty chart.