How the forecast works

Each race is estimated four ways, and each estimate produces a margin and an uncertainty rather than a single number.

Fundamentals

District or state lean, incumbency, and past overperformance. Fitted separately for the House and the Senate.

Polling

A weighted average of public polls, discounted for sample size, staleness and pollster quality.

Markets

Prices from each mapped prediction market, corrected for the favorite–longshot bias that makes long odds systematically overpriced.

Expert ratings

Carried through the pipeline but not currently used in production — the layer is unfitted, because we have no licensed archive to fit it against.

The estimates are averaged by inverse variance, so a confident input counts for more than a vague one. Two rules then constrain the result:

Conflict widens

When the inputs disagree more than their stated uncertainty allows, the combined uncertainty is inflated. Disagreement can never make the forecast more confident.

No new confidence

The blend never claims more confidence than its most confident input, and the combined probability stays inside the range the inputs themselves imply.

Chamber forecasts are not the sum of independent races. Every run draws a national error, a state error and a local error, so seats move together the way they do in reality — a polling miss in one direction tends to be a polling miss everywhere.

That is why the seat distribution is wide even when most individual races look safe, and why a chamber can flip on a national shift rather than on a list of upsets. Each seat's simulated win rate is held equal to its published probability: the hierarchy splits a seat's uncertainty across those three levels, it never adds to it.

Every race carries two forecasts. The model probability never sees a market price. The blended probability — the one shown by default — includes them. The gap between the two is stored daily and alerted on when it persists, because a standing disagreement between the model and the money is information rather than a bug.

There is one such disagreement right now, and we consider it legitimate: the model prices fourteen cycles of drift between August and November, and the market prices less of it. Movement in that gap is worth more attention than either number alone.

Every constant in the model is generated by a fitting script and carries its standard error and provenance. Some are not fitted yet, and this page names them rather than letting a confident-looking number imply otherwise.

Race metadata and primary calendars are reviewed static data. Candidate filings and campaign finance come from the FEC. Polls come from public aggregators. Prediction-market prices come from the venues we have mapped. Prior results come from official canvasses.

District profiles — county makeup and approximate population — come from U.S. Census Bureau public files: the 119th Congress district-to-county relationship file and the 2024 county population estimates. States that redistricted mid-decade for 2026 are approximate until the Bureau publishes updated boundaries.

Missing data is expected, particularly early in a cycle, and it is shown as missing rather than filled in. An empty section means no reviewed source row, poll, market or model output has been stored for that race yet — not that the value is zero.

A primary whose date has passed can still read as pending until an official source row has been reviewed and added. A chamber with no simulation says so instead of drawing an empty chart.