How the forecast works
The model averages the polls, corrects each pollster for its lean this cycle, anchors the result with fundamentals, and then simulates the election with polling errors that move together across races.
What goes in
Polls, plus fundamentals: each state's or district's partisan lean, the national generic ballot and incumbency. Expert race ratings appear on race and district pages for comparison, but they are not part of the forecast.
The polling average
Each race's average is a Kalman filter: the true margin is treated as a quantity that drifts a little each day, and each poll is a noisy reading of it. A poll's weight comes from its sample size, a 2.5-point allowance for non-sampling error, the pollster's quality grade and how recent it is. A firm's older polls count less once it releases a newer one. The same machinery produces the trend line and band on each race page.
Adjusting polls
Registered-voter polls move 0.7 points toward Democrats, the shift likely-voter screens have produced in 2026. Sponsored and internal polls are shaded 2 points against their sponsor and count half as much. Undecided voters are split in proportion to the decided vote. Each pollster's house effect is estimated across every race it covers, shrunk toward zero for firms with few polls, and removed.
Senate fundamentals
The prior is the state's 2024 presidential margin relative to the nation, plus the current generic ballot and incumbency: one point plus half of how far the incumbent ran ahead of their party's nominee last time. Polls and fundamentals are combined by their precision; with many recent polls the polls carry about 90% of the weight.
House districts
Expected margin = 2 × Cook PVI on the 2026 lines + generic ballot + 3 points for an incumbent (1.5 in redrawn states). District polls are weighted by quality, size and age, discounted for partisan sponsorship, and blended in by precision. 282 seats rated safe by every forecaster are held fixed.
Governor races
The 36 governor races use the same polling average and error model. Their fundamentals differ: in 2018 and 2022, governor results followed state partisanship only about half as closely as federal races, and incumbents ran about 12 points ahead. The fundamentals average half the state's lean plus the environment and incumbency with eight published ratings, which carry candidate information the formula misses. With polls, the polls dominate.
Correlated errors
Each simulation draws one national polling error (2.6 points on election day, more before it) shared by every race, then a state error for every state. State errors are correlated according to how alike states are: education, race and ethnicity, urbanization, religiosity, age and recent voting. Wisconsin, Michigan and Pennsylvania move together; so do Arizona, Texas and Nevada. Those similarities were fit to 2008–2022 poll misses and beat a region-only structure by a wide margin. State errors have fat tails, so big misses happen.
Time to the election
Uncertainty shrinks as the election nears: each race's margin drifts by a calibrated amount per day, and national and state errors widen for every day remaining.
Control rules
Democrats control the Senate with 51 seats; 50–50 goes to Republicans through Vice President Vance. Independents who haven't said which party they'd caucus with are counted separately; when they hold the balance, the result is 'Independents decide'.
Tipping point
In each simulation, races are sorted from most Democratic to most Republican. The race that delivers Democrats their 51st Senate seat (or 218th House seat) is that simulation's tipping point. The page reports how often each race plays that role.
History
Every day since July 1 the model is rerun as it would have looked that day, using only the polls that had finished fielding by the day before. Those runs drive the over-time charts and the rewind slider.
Limits
House districts have few polls and are not backtested. The model does not use fundraising or candidate quality. Georgia runoffs and ranked-choice voting in Maine and Alaska are approximated with head-to-head polling.
How it did on past elections
Senate races, 2018–2024We reran this model on 138 Senate races from 2018, 2020, 2022 and 2024, as it would have looked 90, 60, 30, 14 and 7 days out and on election eve, using only the polls available at each point. The error sizes and how fast they shrink were set from that test. In every cycle except 2018 Democrats ran behind their polls, so the model's Democratic-leaning toss-ups lost more often than forecast; it does not assume the next miss will go the same way, because in 2012 and earlier cycles polls missed toward Republicans.
Calibration
Races grouped by the forecast's chance for the Democratic (or independent) candidate. A well-calibrated model lands near the diagonal.
Accuracy by days before the election
Competitive races (final margin under 10 points). Lower log loss is better; 0.69 is a coin flip.
| Cycle | Races | Called right (eve) | Brier | Log loss | Avg. miss | 80% range held |
|---|---|---|---|---|---|---|
| 2018 | 34 | 94% | 0.044 | 0.143 | 3.0 pts | 97% |
| 2020 | 34 | 91% | 0.069 | 0.213 | 7.6 pts | 50% |
| 2022 | 36 | 100% | 0.023 | 0.096 | 5.2 pts | 75% |
| 2024 | 34 | 94% | 0.040 | 0.133 | 5.3 pts | 76% |
Common questions
Who is favored to win the Senate in 2026?
Democrats win the Senate in 68% of the 20,000 simulations and Republicans keep it in 30%. In the other 3%, independent senators hold the balance. The median outcome is 52 Democratic seats, and 80% of simulations fall between 47 and 56.
Who is favored to win the House?
Democrats win the House in 90% of simulations. The median result is 235 Democratic seats; a majority is 218.
When is the 2026 midterm election?
Tuesday, November 3, 2026. All 435 House seats and 35 Senate seats are on the ballot, including special elections in Ohio and Florida.
Has the model been tested?
Yes. It was rerun on 138 Senate races from 2018 to 2024 using only the polls available at each point. On election eve it picked the winner in 95% of those races, and its error sizes were set from that test. The methodology page shows how well its probabilities matched what happened.
What does a 70% chance mean?
The favored side wins about 7 times in 10 simulated elections. The other side winning is an ordinary result, not a sign the polls failed.
How current is the forecast?
The site checks for new polls every hour and reruns the model whenever one arrives. The model last ran Oct 28, 12:50 AM ET, with polls through Oct 27. The bar at the top of every page shows both.
Can I use the data?
Yes. The race-level forecast is available as CSV files for Senate races and House districts.
Download the forecast
Race-level probabilities and projected margins as of Oct 27, 2026: Senate races (CSV) · House districts (CSV)
Data sources
Poliwave poll database · RealClearPolling generic ballot · Silver Bulletin generic ballot · Wikipedia compilation of House ratings and 2026 Cook PVI · Cook Political Report ratings via 270toWin · AP: final 2026 district maps · US Atlas TopoJSON (Census boundaries) · Ballotpedia approval averages · Kalshi markets via DeFi Rate · The Downballot special-elections tracker · Fundraising: TX · Fundraising: OH · Fundraising: IA · Fundraising: NE · Fundraising: ME · Fundraising: AK · Fundraising: MI · Fundraising: GA · Fundraising: NC
About
Midterm Forecast 2026 was created by Craig Vander Galien, a Chicago-based software developer and the founder of DrillerDB. It is an independent project, not affiliated with any party, campaign or media outlet. The model, data pipeline and site update automatically as new polls arrive.