2026 Midterm Forecast

The model is built to be simple enough to check by hand, and it treats both parties exactly the same way. Every number on this page comes from the data files listed below.

1. Polls. Only polls that finished in the last 30 days count. A poll's weight halves every 10 days, and bigger polls count more (by the square root of the sample size). If one pollster has several polls, they are averaged first so that one pollster can't take over the average. Likely-voter versions are used when a pollster released both. Polls that don't publish a sample size are left out.
2. Senate baseline. For Senate races with a Democratic challenger, the polls are blended with a baseline: the state's 2024 presidential margin, minus the national 2024 margin, plus today's national generic ballot. With k polls, the polls get k ÷ (k + 2) of the weight. Races with an independent challenger (Idaho, Montana, Nebraska, South Dakota) use polls only, since a party baseline doesn't fit them.
3. Governors. Same method, but when a governor race has polls, the polls are used alone. Voters split their tickets for governor far more than for Senate (Vermont keeps electing a Republican governor in a state Harris won by 31), so a presidential baseline would mislead. The baseline is used only for the 7 races with no public polls.
4. House. Almost no House districts get public polls, so each district starts from its 2026 Cook PVI (how it votes compared with the whole country, measured on the new district lines) plus today's generic ballot. Each of 10,000 simulations adds a shared national miss (3-point spread) and a separate miss for each district (6 points). Districts with only one party on the ballot go to that party. 218 seats is a majority.
5. Not everyone votes. The polls used are mostly likely-voter polls, which screen out people who probably won't vote. For election night, each state's expected vote is its voting-eligible population times the average of its 2018 and 2022 midterm turnout. That is how the map knows what share of the vote has been counted.
6. Chance to win. Each margin becomes a chance to win using a bell curve with a 5-point spread, or 7 points if the race has fewer than 3 polls.
7. Control. 10,000 simulated elections. Each one adds a shared national polling miss (3-point spread) plus a separate miss for each state (4 points, or 6 with fewer than 3 polls), so states can miss in the same direction like real polls do. In the Senate, 51 seats is a majority, and a 50–50 Senate goes to the Vice President's party (Republican). Independent winners are counted on their own.
8. Election night. Projected margin = (counted margin × share counted) + (pre-election margin × share still out). The spread shrinks as the count goes up. Some states count mail ballots first and others count them last, so early leads can swing either way. That's why the uncounted votes use the pre-election estimate instead of the current lead.
9. Calls. The model never calls a race. A race only shows as called after it's marked by hand from an AP or Decision Desk HQ call.

Known limits. The baselines ignore candidate quality and incumbency, which matters most in Maine and in House races with long-serving members. Alaska and Maine use ranked-choice voting, so the winner may not be known on election night. Georgia needs over 50% or it goes to a runoff on Dec 1. Expected vote totals are estimates, so "% counted" can run a little high or low. The House outlook rests heavily on the generic ballot; if national polls are off, every district moves together, which the simulations allow for.