How the forecast works

Bellwether simulates the 2026 election 40,000 times. Each simulation draws a result for every Senate, House and governor race, with errors that move together the way real elections do. A party’s chance of winning is the share of simulations it wins. Model version 0.2.0; data as of 2026-10-19.

1. Races and district lines

The list of races, candidates, incumbents and each seat’s Cook Partisan Voting Index come from Wikipedia’s 2026 Senate, House and governor pages, which cite Cook’s 2026 PVI on the new district lines. Ten states changed congressional maps after summer 2025 (Alabama, California, Florida, Louisiana, Missouri, North Carolina, Ohio, Tennessee, Texas and Utah), so we never reuse 2022 lines. Hand corrections live in a versioned overrides file with a stated reason for each one.

Special cases: Georgia requires a majority (we estimate the chance of a runoff); Maine and Alaska use ranked-choice voting (we model the final two-candidate round); California and Washington’s top-two primaries can produce same-party general elections (the seat’s party is then settled); Louisiana’s House seats hold an all-party primary on Nov. 3 with a Dec. 12 runoff (we forecast which party ends up holding the seat); independents run against Republicans in some states (Nebraska’s Dan Osborn, for example). Independents who haven’t said which party they’d caucus with are counted separately, which is why a sliver of simulations end with no party at a majority.

2. Polling averages

Polls come from the VoteHub Polling API. After June 2026 that feed stopped carrying most national approval and generic-ballot polls, so from July on we also enter those polls from each pollster’s own release (YouGov/Economist, Reuters/Ipsos, Quinnipiac, Emerson, Marist, CNN/SSRS, Echelon Insights, Verasight, Fox News and others); each one links to its release, and a poll the feed already has is not counted twice. The full list is public. We keep polls that test the two leading candidates actually on the ballot. Each poll’s margin is adjusted three ways before averaging:

• Likely voters. Registered-voter and adult polls are shifted by the average gap between those polls and likely-voter polls taken within two weeks of them, estimated from the generic ballot.
• House effects. A pollster that consistently runs more Democratic or Republican than other pollsters of the same race is adjusted back toward the field. The estimate is shrunk toward zero (as if it had five extra neutral polls).
• Timeline. An older poll is shifted by how much the national generic ballot has moved since it was taken (80% for Senate, 60% for governors, 100% for House).

Weights: recency (a poll’s age is counted from the middle of its field dates; 14-day half-life, stretched up to 60 days when a race has fewer than three recent polls, so one new poll can’t erase the others), sample size (square root, capped at 3,000), pollster rating (0.35×–1.4×), half weight for partisan or internal polls, and divided by the square root of how many polls a pollster has in the race so no single firm dominates.

3. Fundamentals

Before polls, a race’s expected margin is: twice the Cook PVI (PVI is a share-point lean; margin is about double) + the national environment + an incumbency bonus (Senate 3 points, governor 4, House 2.5) + the incumbent’s track record. Governors’ races follow national partisanship less closely, so their lean is scaled by 0.75.

House calibration. Most House districts have no polls, so their fundamentals matter most. Each run we check the House formula against the 2024 results in the 227 contested districts whose lines haven’t changed since then (Wikipedia’s 2024 House results; the ten redrawn states are left out). In 2024, district margins followed partisan lean almost one-for-one (0.96), incumbents ran about 3.6 points ahead, and the typical miss was 4.5 points. We use those values, adding room for this year’s national uncertainty. House members who were already incumbents in 2024 also get half of how far they ran ahead of (or behind) their district that year.

Incumbent track record. Some incumbents consistently beat their state’s lean — Susan Collins won Maine by 8.6 points in 2020 while Biden carried it by 9. For a Senate or governor incumbent on the ballot, we take their last race in the historical file, compare the result with what the state’s lean and that year’s national environment predicted, and carry half the difference into this year (capped at 10 points either way). Added in model 0.2.0.

The national environment blends the generic-ballot average with a structural midterm prior: the president’s party typically loses about 3 points of margin plus 0.4 points per point of net disapproval. Today the generic ballot gets 69% of that blend. The presidential-approval average is corrected the way poll averages are elsewhere: each pollster’s usual lean is estimated against other pollsters polling at the same time (separately for adults, registered and likely voters), shrunk toward zero, and removed before averaging, so a handful of unusually mild or harsh polls doesn’t move the line on its own.

Fundraising. For Senate races and House races within about 20 points, we pull each principal candidate’s cycle-to-date receipts and cash on hand from the FEC and add 0.75 points per doubling of the receipts ratio, capped at ±2.5, when both sides have raised at least $50,000. Money partly follows expected competitiveness, so its signal is mostly already in the polls; the effect is kept small on purpose.

4. Blending polls and fundamentals

Each estimate is weighted by how precise it is. A polling average’s uncertainty is its sampling error plus how much averages typically drift before Election Day (0.6 points × √days) plus a 3-point allowance for systematic polling error. Fundamentals carry an 11-point (Senate), 14-point (governor) or 7-point (House) typical error. As the election approaches, drift shrinks and polls automatically gain weight.

Three versions are published: Polls only (polls wherever they exist, fundamentals only where there are none), Polls + fundamentals (the default), and + Expert ratings, which adds the average published rating (Solid = 22 points, Likely = 11, Lean = 5, Tilt = 2.5) as a third estimate with a 6.5-point error. Ratings are other outlets’ work; we show them with attribution only.

5. Correlated simulation

Every simulated margin = forecast margin + shared shocks + a race-specific shock. The shared shocks: a national swing (about 2.4 points today, governors at 0.75×); an industry-wide polling miss of about 2 points, applied in proportion to how poll-driven each race is; regional (Northeast, Midwest, South, West) and state swings of 1.5 points each (so a state’s Senate, governor and House races move together); and swings tied to each state’s college-degree, Hispanic and Black population shares. National and polling-miss draws use a fat-tailed t-distribution, because big misses (2016, 2020) happen more often than a bell curve says. The race-specific shock fills whatever variance is left so each race’s total uncertainty matches its forecast.

Senate control: Democrats need 51 seats counting the 34 Democrats and allied independents not up this year; Republicans hold control at 50 because the vice president breaks ties. The tipping-point race is the one that supplies the decisive seat when every simulation’s races are sorted by margin.

6. Backtest

We ran the election-eve version of the model on 172 Senate and governor races from 2018, 2020 and 2022, rating pollsters without the tested year. It called 158 of 172 correctly, with a Brier score of 0.0625 (lower is better; always saying 50% scores 0.25). The two uncertainty settings above were chosen by this backtest (systematic polling error 3 points; fundamentals error ×1.6).

Limits of this backtest

• That file only holds polls from each race's final 21 days, so this tests the election-eve forecast, not the forecast five weeks out.
• Incumbency isn't in the file, so the fundamentals here omit it (the live model includes it).
• Pollster ratings are recomputed without the tested cycle (no peeking).
• 2024 isn't in the file; a 2024 backtest needs a 2024 poll archive with certified results (planned).

7. Charts over time

Seats likely to flip

A seat is marked likely to flip (a thick yellow outline) when the party favored to win it is not the party that holds it today and the chance it changes hands is 50 in 100 or better; could flip (a thin outline) is 25 to 49 in 100. It is computed the same way in both directions and for every office. “Holds it today” is the party of the current officeholder. In the ten states that drew new House maps for 2026 (AL, CA, FL, LA, MO, NC, OH, TN, TX, UT) it is the party of the sitting member listed for that seat, whose district lines have changed, so a flip there means the member’s party is not favored in the new district. Seats with no clear holder (vacancies, independents) are not marked. The yellow marks are a reading aid, not a separate forecast: they come straight from each race’s odds.

Odds-over-time charts go back to September 2025. The site launched on Sept. 28, 2026, so earlier points are a backcast: the same model rerun with only the polls that had been released by each date (weekly, then daily for the last 60 days, using fewer simulations per point). They use today’s candidate list and district lines, so they show how the evidence evolved rather than what a forecast published at the time would have said. From launch on, each point is the live forecast as published. Polling-average charts go back to each race’s first poll; presidential approval starts with the second Trump term.

Use the range buttons (1W to 1Y, All, or Custom dates) to zoom. There’s no one-day view: forecasts update three times a day and polls arrive a few times a week, so a day is too short to show movement. Election night will get its own minute-by-minute view.

8. Sources

• VoteHub Polling API — 2026 race, generic-ballot and approval polls
• Pollsters' own releases (collected by Bellwether) — Generic-ballot and approval polls from July 2026 on that VoteHub doesn't carry; each links to its release
• Wikipedia 2026 election pages (CC BY-SA 4.0) — Races, candidates, Cook PVI on 2026 lines, published expert ratings
• FiveThirtyEight raw polls (CC BY 4.0, ABC News) — Historical polls for pollster ratings and the backtest
• Polymarket and Kalshi public APIs — Market odds shown for comparison only
• Census data via Wikipedia tables — State education, Hispanic and Black population shares (correlation factors)

We never invent polls, results or candidates. When a race has no polls, the page says so and the forecast rests on fundamentals.

9. Data status

Every source is cached. If one is down during an update, the forecast uses its last good copy instead of failing, and this table says so.

10. What’s next

Coming in later phases: geographic district outlines for the 2026 lines (the Census hasn’t published them yet, so House maps use equal-size hexagons, one per district), live election-night results with a rehearsal mode, and 2028 and state-level coverage.