Forecast guide
Where a chance comes from
- An estimate without polls. Every race starts from what you would expect before any poll of it. That begins with the national mood, from nationwide polls of which party people plan to vote for. Then come how the state or district voted in 2024 compared with the whole country, and the seat itself, such as whether the incumbent is running.
- An estimate from the polls. The race’s own polls, weighted and adjusted (below), make a second estimate, corrected for how far polls of that chamber have missed in past elections.
- The two, blended. The more polling a race has, the more the polls count. A race with no polls rests on the estimate without them.
- Thousands of simulated elections. In each one, races can miss together, the way they do when the whole country moves. A side’s chance is how often it wins, and the same simulations give each party’s chance of controlling the Senate and the House.
See it worked through for a close race
Governor races follow a simpler version: an estimate without polls, moved toward the polls the more of them there are.
Reading a race
- Chance
- How often each side wins in the simulations. A 70% chance is not a call that the leader wins: the trailing side still wins about three times in ten.
- Margin and range
- The margin is the expected gap between the two sides, in points. The range beside it is where the result should land about eight times in ten.
- Rating
- Solid
- Likely
- Lean
- Toss-up
- Lean
- Likely
- Solid
- Decided by its ballot
- Only one candidate is running, or every candidate is from the same party, so there is nothing between the parties to forecast.
- A party’s chance, not its candidates’
- Two candidates from one party are both on the ballot and nothing yet separates them, so the forecast gives the party’s chance rather than splitting it between them.
- No forecast yet
- There isn’t yet enough to estimate the race from, so it’s left out of the chamber totals until there is.
How a poll counts
Not every poll counts the same. Each poll on the Polls page shows how much it counts, or why it doesn’t.
- Newer polls count more
- A poll’s weight halves every four weeks after it finishes, so the average follows the race as it moves.
- Larger polls count more
- Within limits, so one very large poll can’t drown out the rest.
- Likely voters count most
- Then registered voters, then polls of all adults, since the closer a poll gets to who will actually vote, the better it predicts.
- Campaign and party polls count less
- A poll paid for by a campaign or a party counts 60% as much as other polls, and all the polls from one sponsor together count no more than its strongest single poll.
Raw and Adjusted
- Raw
- The result the pollster found and released, exactly as published.
- Adjusted
- The same poll after taking out that pollster’s measured lean. This is the figure the forecast uses.
Some pollsters’ results run a little more Democratic, or more Republican, than other polls of the same races, again and again.
We measure each pollster’s lean two ways: against other pollsters’ polls of the same race taken within two weeks, and by how its final polls compared with the results in past elections. A pollster with only a few polls this year is judged mostly on its past record; the more it polls, the more its polls this year decide.
A poll paid for by a campaign or a party is also moved toward the other side, by how far such polls have leaned toward their sponsor in the past, until the firm’s own record takes over.
On the Polls page, switch between Adjusted and Raw to see both. Governor polls are averaged as released, so they show Raw only.
What if the nation moves?
The forecast assumes the national mood stays where the polls put it now. To see what a shift would do, choose from Much more D to Much more R on the Forecast pages or in How it works, and every race moves with it. Now goes back.
How well it has done
- Races with polls
- 78%
- Races without polls
- 84%
We tested the method on every Senate and House race from 2018 to 2024. Each election was forecast by a version of the method fitted without it, so it couldn’t learn the answer. Its 80% ranges held about eight times in ten, with polls and without. A late scandal or a turnout surprise can still beat it.
Limited estimates
Some races aren’t a straight contest between the two major parties: an independent against one of them, a ranked-choice vote, or an all-party primary with a runoff. They still get a real estimate, marked Limited estimate, with a wider range, since there is less to go on.
Betting markets
Each race page shows the betting market’s price beside the forecast, so you can compare the two. The forecast never uses it.