Methodology

How the Senate, House, and governor Bayesian forecasts are built.

The forecast combines what polls say with what the fundamentals of each race suggest, then runs tens of thousands of simulations to turn those estimates into probabilities. Every margin below uses the convention that a positive number favors the Democratic candidate.

1. Poll averages

Each poll is weighted by how recent it is (an exponential decay with a three-week half-life), its sample size, the population surveyed (likely voters count more than adults), and the pollster’s quality rating. A house-effect adjustment removes each pollster’s persistent lean before the average is taken.

Polls commissioned by a partisan sponsor (a campaign, party, or aligned group, marked (D) or (R)) tend to lean toward whoever paid for them, so their margin is pulled back toward the center and they are given less weight than independent polls.

2. Fundamentals

Before polls, each race has a prior expectation based on the state’s partisan lean, whether an incumbent is running, candidate quality, fundraising, and the national environment (the generic congressional ballot and presidential approval, with a midterm penalty for the party in the White House).

A few races feature a credible independent, rather than a Democrat, as the main challenger (for example Dan Osborn in Nebraska). Because such a candidate consolidates the anti-incumbent vote and runs well ahead of a generic Democrat in a lopsided state, their prior adds an overperformance adjustment, widens the uncertainty, and ties them less tightly to the national partisan tide.

Montana is a three-way race: a Democrat and a credible independent are both on the ballot against one Republican. The seat still goes to whoever finishes first, so the model follows whichever challenger leads and treats the other as splitting the same pool of voters. That split carries a penalty, which is what lets a Republican win on a plurality even when the two challengers together outpoll them. Only the leading challenger carries a win probability; polls that offer just a head-to-head are set aside, because they describe a ballot that does not exist.

3. Bayesian updating

Each race starts with a Gaussian fundamentals prior. Its corrected poll average is a noisy observation of the final margin. Combining the prior and polling likelihood produces a posterior mean and uncertainty. The poll variance is an error floor squared plus a scale squared divided by the effective poll count; this prevents many polls from creating false certainty. A race with no usable polling keeps its fundamentals prior.

The reviewed floor/scale parameters, in margin points, are 3.94/8.97 for the Senate, 7.13/4.67 for the House, and 4.94/7.74 for governors. They are pinned in the live model, not refitted automatically each day. Effective poll counts still change as polling arrives and ages. House and governor calibration used 577 and 279 historical races, respectively, at a 36-day-before-election cutoff. The Senate retains its existing 403-race calibration.

4. Simulation

Forty thousand correlated Gaussian draws from the race posteriors produce win probabilities and seat distributions. Shared national, regional, and partisan-lean error structure means races can move together; outcomes are not independent coin flips. Another race’s polls do not directly shift a race’s posterior mean. The previous model’s heuristic blend and fat-tailed Student-t simulation remain available as a legacy comparison, but are not the default live forecast.

The House requires 218 of 435 seats. Governor summaries count all 50 governorships, including the 14 not up this cycle; either party needs 26 for a majority, and a 25–25 split is no majority. This is a count of state offices, not control of a chamber.

Democratic control requires 51 seats because the Republican Vice President breaks 50–50 ties. Thirty-five seats are contested in 2026; the other sixty-five are held at their current party.

When a winning independent has said they would caucus with neither party, their seat counts toward neither majority. A strong enough independent can therefore leave the chamber hung, with neither party reaching a majority on its own — an outcome the forecast reports alongside each party’s chance of control.

Data & limitations

Holding out entire election cycles, the House polling likelihood covered 80.2% of outcomes in its 80% intervals and 95.1% in its 95% intervals. Governor coverage was 83.2% and 93.5%. These checks validate polling uncertainty, not the complete Bayesian posterior, fundamentals priors, or chamber correlations. Unpolled, independent-led, and three-way races are not represented by the D-versus-R calibration sample.

The historical archive ends in 2022 and contains polling within 61 days of elections. Its cutoff uses median field dates, not verified publication dates. Historical population types and contemporaneous ratings are unavailable and use defaults; some sample sizes were imputed by the provider. Applying the fixed calibration at different horizons, or to very stale polling, remains an extrapolation.

Historical projection points from before the Bayesian release are retained, not recomputed. A change around the transition can reflect methodology as well as new polling. New forecast files identify the model and its calibration version.

Live polling is compiled from Wikipedia’s race tables and 270toWin’s published poll listings (individual pollster results only), together with hand-entered overrides. Pollster ratings and historical polling come from the FiveThirtyEight data archive (CC BY 4.0); certified past results come from the MIT Election Data + Science Lab; fundraising comes from the Federal Election Commission. Third-party aggregator forecasts and polling averages are used only for private cross-checking and are never republished. This is an independent project and is not affiliated with any of these organizations.