Senate, Governor and House on one map. Slide the dial to hear from conservative-leaning or liberal-leaning pollsters, or drop it into the center notch to lock my best prediction: only the most reliable polls, blended with a prior calibrated to how this cycle has actually moved. Every result is shown for both parties, and the pollster mix is balanced so the estimate does not depend on which firms are in the pool. Polls through Oct 24, 2026.
Which pollsters do you want to hear from?
This page is meant to state what the evidence says, not to favor a party. These are the checks I ran on Oct 19, what I found, and what I changed.
How the dial works. Each stop keeps or drops whole groups of pollsters: conservative-leaning only, drop liberal-leaning, all polls, drop conservative-leaning, liberal-leaning only. A state's margin is the plain average of the polls that survive (Democrat minus Republican). The count is how many polls survive, out of all polls collected for that chamber.
By affiliation uses the label a source gave a pollster or, for InsiderAdvantage, Quantus, Big Data Poll, co/efficient and Cygnal, my assumption from general reputation. Democratic-aligned firms are rare in public polling and I have only two such polls, so the liberal end is thin on purpose, not by accident.
By measured results ranks pollsters by how far their results sit above or below other polls of the same race this cycle (a third each). It selects on results, so it shows who ran high or low, not who is right. Pollsters with fewer than two comparisons count as middle.
Coverage. Senate: polls ending Sep 15 or later. Governor: the latest RealClearPolling list (Sep 25 to Oct 18), where field dates and sample sizes were not shown. Many Senate polls are likewise dated by RealClearPolling's list date. House: national generic-ballot polls ending Sep 1 or later; no district-level public polls are in my data, so the House map shows Cook's Oct 9 seat ratings and redistricting, and only the national reading responds to the dial. Cook lists only competitive House seats, so states with no flagged seats are not necessarily uncompetitive.
October surprise. A late shock (a news event, a scandal, a market move) is the thing no poll can see yet. History says the polling average moves little after this point: in the four elections where I have a snapshot two weeks out or less, it moved 0.1 to 1.1 points before Election Day, and the gap between the final average and the result was usually larger than that. The model therefore adds an explicit allowance to its uncertainty even though no surprise has happened: Typical adds 1.0 point of shared national uncertainty (about the largest late move on record) and Large adds 2.0 (about twice that, since a shock 13 days out has more time to act). It is added in quadrature to both the shared and the per-race error, so it pulls win chances toward 50% and widens ranges without favoring either party; it is on Typical by default and None removes it. A surprise that helps one side can be tested with the Turnout switch.
Rewind. The slider re-runs everything using only polls published by the chosen date (poll ages are measured from that date). It shows how the picture moved, not what anyone knew at the time about pollster house effects, which are measured on the full set.
Group and sector factors. The state panels show five measures of who lives and works in the state: union membership (BLS, 2025), veterans, farm and resource workers (agriculture, forestry, fishing and mining), manufacturing workers and median household income (Census ACS, 2023). I tested them out of sample against the 2026 poll averages and they did not improve the forecast beyond the 2024 margin, which already carries them, so they are context rather than inputs. The test and what is not covered (Hollywood, active-duty military, group-level movement) are in the Do group factors help? section.
Immigrants and origin. Immigrants cannot vote until they naturalize, so the page shows only naturalized citizens (about 10% of eligible voters nationally, with state shares where Pew reports them). Group-level facts about how origin groups voted in 2024 are listed with their sources and caveats in the Immigrants and the electorate section. None of it feeds the estimate: group averages do not describe individuals or whole communities, standard data do not record a person's region within their birth country, and the surveys disagree with one another by several points.
Pollster balance. Pollsters differ in how Democratic or Republican their results run compared with other polls of the same race, and the most-reliable pool happens to run a little Democratic (about 0.9 points). To keep the pool from tilting the estimate, each poll is shifted by its tier's measured house effect before averaging, so the center is the consensus of all pollsters. It is on at the center settings and locked on in the prediction gear; it is off in the camp-by-camp settings because those show each camp's raw view. Balancing assumes the all-pollster consensus is the right center; the backtest shows bias flips direction from cycle to cycle, so I do not pick a side.
Most reliable polls. This switch keeps only tier-1 pollsters. I define them from sources I could read: Silver Bulletin's January 2026 ratings name the Washington Post, Marquette and NYT/Siena as top-rated; ABC/538 gave NYT/Siena and ABC/Post its only three-star ratings; Media Bias/Fact Check rates Beacon/Shaw (Fox News), Emerson, Marist, SurveyUSA, Suffolk and YouGov High for factual reporting. I could not open Silver Bulletin's full grade table, so the tiers are my reading of those sources, not its grades. Trafalgar and Rasmussen are tier 3 because Silver Bulletin says they have worked with Republican campaigns without fully disclosing it; Democratic-aligned firms are tier 3 too. Everyone else is tier 2. If a race has no tier-1 poll, the best estimate falls back to all of its polls and says so.
Best estimate. For each race: every poll gets an error from its sample size plus a systematic error by tier (2, 3 or 4 points) and is down-weighted by age; polls are averaged by those weights; that average is blended with a fundamentals prior (the state's 2024 presidential margin plus 90% of the national swing, where the national swing is the generic-ballot average minus the 2024 national margin of D-1.5). The blend weights each by its precision. Win chance assumes a 4.5-point Election Day error (wider on request). Senate control comes from 4,000 simulated elections with a 3-point error shared across states (wider on request). No candidate-specific adjustments are applied; a Collins sensitivity is shown under the Senate chart. All of these numbers are my assumptions and are listed in the database. The estimate uses demographics only indirectly: 2024 results already carry each state's makeup, and the group crosstabs shown here are not fed into the number, because I could only retrieve them for four states.
Turnout. The 2024 result turned mostly on who showed up, not on who switched. Pew finds 89% of Trump's 2020 voters returned against 85% of Biden's, and each point of gap between the parties' returning voters moves the national margin about half a point, so 2024's gap was worth about 2 points. This switch moves every race and the House scenario by 2 points either way. The Republican edge repeats 2024 (it also matches Pew's finding that Democrats lead by 6 among validated 2024 voters but 8 among all registered voters). The Democratic edge is its mirror image, supported by Pew's motivation gap (59-35 among the extremely motivated against 44-36 overall) and Reuters/Ipsos enthusiasm (66% of Harris voters against 51% of Trump voters), but it is a scenario, not a measurement. The gear locks this switch at as polled.
Polling error. Senate polls from 2018 to 2024 averaged a 3.7-point error toward Democrats across 3,312 polls, and the direction has flipped from cycle to cycle, so I widen the error instead of adjusting the polls. Base uses a 3-point error shared across states and a 4.5-point error per race. Wider and Widest add 1 and 2 points to both. The grid under the dial shows Senate control under every turnout and error setting at once.
District map. "By district (map)" draws the real district boundaries for the 40 states whose maps did not change for 2026 (Census-based 119th Congress boundaries from a public dataset). The ten states with new 2026 maps (AL, CA, FL, LA, MO, NC, OH, TN, TX, UT) are hatched and show one placeholder hexagon per district, because I could not obtain verified boundaries for the new maps and the older shapes would be wrong. "Hex (all 435)" puts every district on one equal-area grid grouped by state. Inside a hatched state or the hex grid, a hexagon's position is arbitrary. Competitive seats come from Cook's Oct 9 list. Other seats are shown by the current member's party (unitedstates/congress-legislators, current through Sep 2026). Those parties add up to Cook's solid totals (185 D, 174 R), but in states with new maps a member's district number may not match its new territory, so those seats are faded.
The projection layer. Seat margins start from Cook's ratings at the all-poll generic-ballot average (Toss-up 0, Lean ±4, Likely ±9, Solid ±25 points, my assumption). Every seat then shifts by how far the polls you are consulting sit from that average (uniform swing). Seat chances use a normal error of about 7 points, 4 shared nationally and 5.5 per seat, over 3,000 simulations. It is a scenario tool, not a forecast: at the all-polls setting it gives Democrats about 223 seats and a 65% chance of reaching 218, lower than the models and markets cited elsewhere, because Cook's ratings are its only input and its errors are wide. Read the change between dial settings, not the level. A seat's colour is its simulated chance, stretched so close seats stand out.
Sources. RealClearPolling (governor and generic-ballot lists, Senate averages), Newsweek poll roundup, Cook Political Report (Senate Oct 6, Governor Oct 13, House Oct 9 via 270toWin), redistricting trackers. Full source list and the database are in the Cassandra repository.