The site is offline. Visitors see the offline message. You’re seeing the site because you’re signed in. Bring it back in the Assembly Room
Preview with sample data. The candidates and numbers on this page are placeholders, not real polling, until the model is connected.
Forecasts
There is only one November, but thousands of ways it could unfold. We simulate 10,000 of them to see who's favored, what could change, and how much we still don't know.
Beta testing. Only your team can see this forecast until it's opened to the public.
The 2026 Senate
Closest races
ClickTap any colored state to see its full forecast.
Each state opens to its polls, its history and every piece of our forecast. Gray states don't have a Senate race this year.
Beta testing. Only your team can see this forecast until it's opened to the public.
Who wins the nomination
Chance of winning
Select a candidate for detailsThe range of outcomes
Each dot is 100 of our 10,000 simulations, placed at the share of the national primary vote the candidate wins in them. Where the dots pile up is where the result most often lands. The dark line marks the middle.
What 10,000 simulations look like
Each dot is one simulated primary season, colored by the candidate who won it. The more dots a candidate has, the more often the race breaks their way.
Point at or tap any dot to see that simulation.
How the odds have changed
Each candidate's chance of winning, recalculated every week as new polls come in.
How sure are we?
A forecast is a range of possibilities, not a single answer. These numbers show how much room there is for surprise.
State by state
Which party is favored to carry each state in November, and who leads each state's primaries. Darker blue means Democrats are more likely to win a state, darker red means Republicans are, and purple means it could go either way.
Point at a state to preview it, or select it for details.
Commentary
Analysis of the race from the people behind the forecast.
How it works
How our Senate forecast works, in four steps.
-
Start with history
Each state starts from how it voted in the last two presidential elections. Then we adjust for the candidates: sitting senators have had about a 4-point edge since 2016, and some carry a personal edge from how they ran last time.
-
Measure the national mood
We estimate how far the country will swing against the president's party, using his approval, including among his own voters. Since 2016 that swing has stayed between 3 and 6 points, even when the president was unpopular, because so few voters cross party lines.
-
Weigh the polls
We use every public poll, adjusted for each pollster's usual lean, measured against past results. Polls count for more when a race has many of them. They can also miss in the same direction everywhere, by about 2 points in a typical midterm, so we don't take them at face value.
-
Play it out 10,000 times
Each simulation draws one plausible version of the election, with misses linked across states, so a surprise in Michigan makes one in Ohio more likely. A candidate's chance of winning is the share of simulations they win.
Open any state on the forecast map to see every piece of its forecast
Why we're cautious about waves
How we tested it
We hid each election from 2016 to 2024 from the model, one at a time, and had it forecast the hidden year. It picked the winner in 95% of Senate races, and in 92% of the close ones. We also tested gas, grocery and electricity prices, and whether some states have more persuadable voters. Neither made the forecasts better, so neither is used, for now.
Where the data comes from
Election results, 2006 to 2024, as compiled by FiveThirtyEight. Past polls and pollster ratings from FiveThirtyEight, used under the CC BY 4.0 license and adapted for our model. This year's polls from the pollsters themselves, gathered from Wikipedia, 270toWin, RealClearPolling and Electoral-vote.com. Presidential approval from Gallup and the RealClearPolitics average. Prices from the U.S. Bureau of Labor Statistics.
Reading the numbers
A 25% chance is not a prediction that a candidate will lose. It means they win about one time in four. If the forecast is working, candidates we give a 25% chance should win about a quarter of the time.
The 2028 presidential forecast (in beta testing)
The 2028 forecast runs on the same model as the Senate forecast. We built it from every Senate race and every state's presidential result from 2006 to 2024, then tested it on the 2016, 2020 and 2024 presidential elections without letting it see them first. It called the winner in 95% of states and missed the final margin by about 3 points on average.
-
Start with history
Each state starts from how it voted in the last two presidential elections. In presidential races, states hold almost all of that lean from one election to the next.
-
Measure the national mood
We estimate how far the country will swing toward or away from the party in the White House, using the president's approval, including among his own voters. Since 2016 that swing has been small, because so few voters cross party lines.
-
Weigh the polls
We use every public poll in every state, adjusted for each pollster's usual lean. Polls miss in the same direction more often in presidential years, by about 3.5 points against 2 in midterms, so we trust them a little less.
-
Play it out 10,000 times
Each simulation runs all fifty states and D.C. together, with misses linked across states, and counts the electoral votes. A candidate's chance of winning is the share of simulations they win.
The 2028 forecasts will run on this model once there are candidates and polls to run it on. Until then, the numbers on the 2028 page are placeholders.
Why we built this
We're trying to understand why Americans vote the way they do.
The way people engage with politics today is fundamentally different than two decades ago. The traditional forces still matter: the economy, party loyalty, incumbency. But new ones have emerged alongside them, and they're making races closer and harder to call. That's why we start with the past. Studying how Americans used to vote is how we measure what's different now, and how much of the old rules still hold.
We're building a model that captures both: not just a better average of the polls, but a clearer picture of what's actually moving voters.
We're data scientists and trained researchers, so we expect to get some of it wrong at first. Understanding everything that shapes a vote takes time. This site is where we work it out in public: we'll show our work, learn from every miss, and keep improving the model as we go.
Ten Thousand Novembers is a non-commercial, self-funded project. We don't run ads, take sponsorships or charge for anything on this site.
Press, feedback or anything else: letters@tenthousandnovembers.com
