How the forecast works
Every day the model collects polls, ratings, prediction-market prices, fundraising reports, economic data and news, then simulates the election 40,000 times. Each simulation draws one plausible set of polling errors and late shifts, and plays out all 435 House races, 35 Senate races and 36 governor races together.
1. The national environment
The starting point is the expected national House vote margin. It blends four signals, each weighted by how accurate it has been: our own adjusted average of generic-ballot polls together with the published averages; prediction-market prices on the national vote (Kalshi and Polymarket); a fundamentals model of how the president's party fares at a given approval rating, fit on every midterm since 1946; and special-election swings when available.
2. Each race
Each race combines four estimates. Fundamentals are the national environment plus the district or state's partisan lean (twice its Cook PVI), incumbency, candidate experience, scandals and FEC fundraising. Polls come from Wikipedia, VoteHub, 270toWin and Silver Bulletin, de-duplicated. They are adjusted for partisan sponsors (internal polls historically lean about 4.5 points toward their sponsor), pollster house effects, likely-voter screens and national trend, and weighted by recency, sample size and 538 pollster grade. Ratings are the consensus of up to ten forecasters: Cook, Sabato, Inside Elections, Split Ticket, Decision Desk HQ, The Economist, Silver Bulletin, FiftyPlusOne, RealClearPolitics and Fox News. Markets are vig-free prices from Polymarket, Kalshi and PredictIt. The four are combined by precision. Ratings and markets get a discount because they partly reuse the same polls.
3. Correlated uncertainty
Polling errors are not independent. They move together nationally, by office, by region, and by how red or blue a place is, as in 2016 and 2020. The simulation draws all of these jointly with fat-tailed distributions, so a polling miss that hurts one party in Ohio also tends to hurt it in Michigan and Pennsylvania.
Margin of error, honestly
What news does, and doesn't, do
Headlines for every competitive race are collected daily and shown with each race. News feeds the model through what it reports: new polls, candidate changes, rating changes and market moves. Raw media tone is not used to move margins, because it has no demonstrated out-of-sample predictive value.
Known limits
- District-level polling is sparse and often partisan; most House seats lean on fundamentals, ratings and markets.
- Alaska uses a top-four primary with ranked-choice voting, Georgia requires 50% to avoid a runoff, and Louisiana's House races use a Nov. 3 all-candidate primary with Dec. 12 runoffs. The model forecasts the eventual two-candidate result.
- Nebraska's Dan Osborn is an independent who has not committed to either caucus; he is counted on the non-Republican side but not as a Democrat for Senate control.
- Nine states use new congressional maps in 2026 (TX, CA, NC, OH, UT, FL, TN, LA, AL); incumbents running on redrawn lines get a smaller incumbency bonus.