How to forecast 6,273 elections, most of which nobody polls

On Tuesday, voters in Texas House District 118, in San Antonio, chose between Kristian Carranza, a Democrat, and Jorge Borrego, a Republican, for an open seat held by Republicans. Two years earlier, Donald Trump had carried the district by five points. As far as we know, no pollster asked a single voter there how they planned to vote. Almost nobody polls races like this one: of the 5,652 state legislative races on the ballot this year, three had a public poll.

Our forecast gave Carranza an 87 percent chance anyway, a confident call for a Republican seat in a district Trump had won. By the unofficial count, she won by 23 points. The forecast put a number on every one of the 6,273 races on the 2026 ballot, from the U.S. Senate down to the state house, and on which party would control the House, the Senate, a majority of governorships and each of the 89 legislative chambers with seats up.

This is an explanation of how that worked: what goes into a forecast, how the evidence is weighed, and why what comes out is a probability rather than a prediction. It is written for anyone who has seen a number like “87 percent” and wondered where it came from, or how seriously to take it.

This project was inspired by the public election models of FiveThirtyEight, The Economist and Nate Silver’s Silver Bulletin, which showed how much a forecast can say, and how honestly it can say what it doesn’t know. Ours uses the same broad ingredients: polling averages that correct for each pollster’s lean, “fundamentals” that describe the candidates, and tens of thousands of simulated elections. What is different is how far down the ballot it goes, to the races where polls run out and most of the forecasting has to be done some other way.

Our final forecast ran in the early hours of Election Day, Nov. 3. Votes are still being counted in many states; once results are certified, we will publish a full accounting of how the forecast did, race by race. This piece describes the model as it ran.

The short version

Every race’s forecast starts from four questions.

1. How is the country leaning this year? The national environment.
2. How does this place usually vote, compared with the country? Its partisan lean.
3. Who is running? Incumbency, experience, money and track record: the fundamentals.
4. What do the polls say, about this race and about races in places like it?

The answers become an expected margin for each race and a measure of how far off that margin could be. Then the computer runs the election 40,000 times, with errors that push races in the same direction the way real elections do, and counts how often each candidate wins.

There are two versions. Horizon, our headline forecast, uses everything above and looks ahead to Election Day. Pulse uses only polls, as if the election were held today. Unless we say otherwise, the numbers in this piece are Horizon’s.

Start with the country

Much of what decides a race happens outside it. In a midterm, the strongest single force is the national mood, which is usually a verdict on the president. The most direct measure is the generic ballot, the poll question that asks which party’s candidate you would vote for in your House district, without naming anyone.

We average those polls the way we average every poll (more on that below). The average moved steadily through the cycle: Republicans led by 2.6 points in January 2025, Democrats pulled ahead that April, and on Election Day they led by about nine points among likely voters.

The generic ballot is one reading of the mood. The races themselves offer a second. If Senate and House polls are running, on average, a few points better for Democrats than those races’ fundamentals would predict, the national mood is probably better for Democrats than the generic ballot alone shows. That bottom-up reading put Democrats 10.4 points ahead. We average the two, for an environment of D+9.6.

Earlier in the year, when polls say less about November, Horizon also leans on history. Since 1946, the president’s party has almost always lost ground in midterms, and by more when the president is unpopular. A model of those elections, fed the president’s approval rating (37 percent in our average of approval polls) and the previous House vote, pointed to Democrats by 9.9 points. Horizon pulls its reading of the national mood toward that history, by up to about a third when the election is six months away, less as it approaches, and not at all on Election Day. By the end, polls and history were within a point of each other anyway.

Then, the place

A good year for one party does not lift it equally everywhere. Each state and district has a partisan lean: how much more Democratic or Republican it voted for president than the country as a whole, counting 2024 three times as much as 2020. Official returns are not broken down by state legislative district, so for districts we rebuild the presidential vote from precinct results on the lines in use this year.

Lean plus the national environment is the simplest forecast there is, and in an era of little ticket-splitting it gets surprisingly far. Texas’s 118th District, for instance, voted about two points more Republican for president than the nation. That makes it close to a toss-up in a neutral year, and a Democratic-leaning seat in a year as good for Democrats as this one.

Three refinements matter. Some places swing more than others between good and bad years, and we measure how much from the Cooperative Election Study, a large academic survey run every election year; District 118, by that measure, swings about 33 percent more than the average place. A national swing also does not reach every office in full: in past elections, House races moved about four-fifths as much as the national vote, and state legislative races about three-fifths. And some regions consistently run ahead of or behind their presidential vote down the ballot; New England’s Republicans, for one, run well ahead of their party’s presidential ticket.

Then, the candidates

The next layer is what forecasters call fundamentals: facts about the candidates that predict how they will run compared with their party. Incumbents usually do better than a newcomer would, and an incumbent who outran the party last time tends to do it again. Candidates who have held elected office before do better than those who have not. In federal and legislative races, so do candidates who raise more from individual donors, especially donors in their own state.

We estimate what each of these is worth from past elections: every Senate and governor’s race since 1976, other statewide offices since 2014, and House and state legislative races since 2016, weighting recent elections and midterms more heavily. Those effects have shrunk as voting has become more partisan, and a model built on candidate traits can find patterns that are not really there. So about half the weight goes to the simple version, lean plus the environment, and when the two disagree sharply, the model trusts the fundamentals less and leaves more of the decision to the polls.

Susan Collins, Maine’s Republican senator, is the textbook case. Maine leans about 8 points more Democratic than the country. In a year this good for Democrats, lean and environment alone would have made her challenger, Troy Jackson, a 15-point favorite. But Collins has a long record of beating her state’s lean: in 2020 she won by about nine points in a state Joe Biden carried by about nine. Count her incumbency and that record, and the fuller version of the fundamentals favors Collins by two points. Estimates 17 points apart are a sign that neither can be trusted on its own, so the model treated Maine’s fundamentals as nearly uninformative and let the polls decide.

Reading the polls

Polls are the most direct evidence there is. They are also noisy, and some are noisier than others. The forecast used 2,767 polls of 2026 races and the generic ballot: Senate, governor and House polls and the generic ballot from The New York Times’s polling database, and polls of other statewide races and of state legislatures from Wikipedia’s compilations. Each poll is weighted by three things.

• How recent it is. Newer polls count for more. In the last weeks of a campaign, a poll from 10 days ago counts half as much as one taken today. Earlier in the year, when the election is far off, a poll keeps half its weight for about a month.
• How many people it reached. Bigger samples count for more, though with diminishing returns. A firm that polls the same race every week does not get a vote every week: its repeated polls are discounted, so no single pollster can dominate an average.
• The pollster’s record. We grade pollsters ourselves, scoring every poll from the final three weeks of past elections, 1998 to 2024, against the certified result and against other polls of the same race. Grades run from A+ to D, and better-graded firms count for more. Each grade is on the pollster’s page.

Then come the adjustments. Polls of registered voters or all adults are converted to likely voters, using the gap in polls that report both; this year, the gap was small. An older poll of a race is moved by half of how much the generic ballot has moved since it was taken. And each poll is corrected for its pollster’s house effect.

Some firms consistently show one party doing better than other polls of the same race taken around the same time. That is not necessarily a sign of bad polling: firms make different, defensible choices about whom to count as a likely voter and how to weight their samples. But an average that happened to include more of one firm’s polls would drift with that firm. So we measure each pollster’s lean against everyone else’s, across every race it polls, and subtract it. A firm with only a few polls is assumed to lean the way it did in past elections until it has polled enough to show otherwise.

Polls paid for by a campaign or a party are a special case. They are real polls, but campaigns tend to publish the ones they like. We treat each firm’s sponsored polls as a house of their own, start from the assumption that they lean four points toward whoever paid for them, and let the evidence move that estimate. By Election Day, Democratic-sponsored polls were running 3.0 points more Democratic than other polls of the same races, and Republican-sponsored polls 3.5 points more Republican.

One more rule: a firm shown to have fabricated its results is removed entirely. Its polls are not averaged, graded or shown anywhere on the site.

Weighing the evidence

For each race, the model now has up to three estimates: its own polling average, its fundamentals, and a reading borrowed from polls of similar places, which we come to in the next section. Each comes with a sense of how wrong it tends to be. The model combines them in proportion to their precision: an estimate whose typical error is half as large gets four times the weight. When the estimates disagree, the combined forecast is less certain than any one of them would suggest, and its range widens.

In Maine, 33 polls put Jackson ahead by an average of 3.4 points. The fundamentals’ range was so wide that they got 1% of the weight to the polls’ 91%, and the forecast landed close to the polls: Jackson by 3.5, a 76 percent chance.

The races nobody polls

In most races, there is no polling average at all. Of the 6,273 races on the ballot, 223 had a poll. Another 1,823 were not contested by both parties, which settles them: the model simply counts those seats. That leaves more than 4,200 races, nearly all of them for the state legislature, where the forecast has to infer what polls would have said.

It does that in two ways. The first is similar places. For every race, the model finds the polled races most like it, in partisanship, education, race and ethnicity, religion, age, income, population density and geography, with extra weight on races in the same state and region. It then asks how those polls were running against their own fundamentals, and carries that over. If polls across the Southwest are running two points better for Democrats than expected, an unpolled district in San Antonio probably is too. Special legislative elections held during the cycle count as evidence the same way.

The second is the state environment. Candidates for governor or attorney general share a ballot with legislative candidates, and their polls say something about how a state’s voters feel about state government that the national mood does not. Each state’s environment is read from its own governor’s race first, then other statewide races and any legislative polls, measured against what their fundamentals expected. It is pulled toward the national environment, so a state with little polling stays close to the national mood. In Texas, state-level polls ran about 0.6 points more Democratic than the national mood.

Neither is as good as a poll of the race itself. A reading from similar places measures how a kind of place is moving, not how a particular candidate is doing. So the forecasts for unpolled races are much less certain than for polled ones, and that is the honest answer.

In District 118, though, everything pointed the same way. The fundamentals saw an open seat that swings hard, in a good year for Democrats, with Carranza having raised about $480,000 from individual donors to Borrego’s $78,000; they put her ahead by nine points. Polls of similar places, most of them elsewhere in Texas, agreed. Two wide estimates that agree make a confident forecast: 87 percent. She won by 23, beyond even the top of the forecast’s 80 percent range, part of the statewide polling miss described in our analysis of the Texas results.

Elections move together

If you flip 435 coins, the number of heads barely varies from one try to the next. If the 435 House races were like coins, each with its own independent chance, the forecast would have been nearly certain how many seats Democrats would win: four times in five, between 236 and 247.

Races are not coins. When polls miss, they tend to miss in the same direction in many places at once. In 2016 and 2020, they underestimated Republicans almost everywhere. In Texas this year, as our analysis of the results shows, they underestimated Democrats by about nine points. A forecast that treats races as independent will be confidently wrong in exactly the years that matter most.

So each of the 40,000 simulations draws its errors in layers. There is a national error, shared by every race, which is larger the further away the election is and has fat tails, leaving room for a miss bigger than any in recent memory. There is an error for each region, one for each state that moves every race in it from the Senate to the state house, and one for each legislative chamber. There are errors tied to demographic groups, so that if polls misjudge, say, Hispanic voters, every heavily Hispanic district moves together. And each race gets an error of its own. The sizes come from how wrong forecasts have been in past elections; in our rerun of 2022, races landed inside their 80 percent ranges 80 percent of the time.

With shared errors, the 80 percent range ran from 227 to 256 seats, 2.6 times as wide. With independent races, a Republican majority would have been all but impossible (under one in a thousand); with shared errors it happened in 1.4% of simulations. The Senate forecast (81% for Democrats) and the governors’ forecast (82%) were built the same way, and because legislative and statewide races are drawn in the same simulations, a state’s governor’s race and its legislature rise and fall together.

The simulations also handle each state’s rules: runoffs in Georgia and in Louisiana’s House races, ranked-choice voting in Maine and Alaska (modeled as the final round), districts that elect several members, chambers where only some seats are up, independents running as a party’s de facto nominee, and the lieutenant governors who break ties in some state senates.

Two forecasts, not one

Horizon and Pulse answer different questions. Horizon asks what will happen on Election Day, using polls, fundamentals and history. Pulse asks what the polls say today, and nothing else: races without polls borrow only from polling of similar places, its national mood is not pulled toward history, and because it makes fewer assumptions, its uncertainty is 15 percent wider.

When the two disagree, it is usually because a race has few polls and the fundamentals point somewhere else. In District 118, Pulse, which ignores fundraising and the other candidate traits, was less sure, giving Carranza 71 percent; in Maine, where it never saw Collins’s record, it gave Jackson 81 percent. Nationally, they ended close: the House at 99% in Horizon and 97% in Pulse, the Senate at 81% and 77%, governorships at 82% and 84%.

Over the fall, Horizon’s chance of a Democratic House rose from 89% to 99%, the Senate from 68% to 81%, and a Democratic majority of governorships from 62% to 82%. Each day’s forecast is saved as it ran and never recomputed, so the daily updates show what the model said at the time, not what it would say now.

How often is it right?

A forecast can be tested before an election by pretending it is an earlier one. We reran the model as if it were mid-October 2018 and mid-October 2022, using only what was known then (the polls, candidates and past results of the day), and scored it against what happened in every Senate, governor, House and state legislative race with a result.

Calling 92 to 94 percent of winners sounds impressive, and is less so than it sounds: most races are not close, and a forecast that knew only each district’s lean would get most of them right. The better test is whether the probabilities mean what they say. A candidate given an 80 percent chance should lose one time in five, and in both years, favorites given 75 to 90 percent won 85 percent of the time. The chamber-level misses in 2018 (the Michigan House and Senate, the Pennsylvania House and the New Hampshire Senate) were close calls. 2018 is also the harder test: presidential results by district only start in 2016, so its fundamentals learned from one election, where this year’s learned from five. Two midterms are a small sample, which is why we will keep checking.

What the forecast does not know

• Experts and markets. The forecast uses no expert race ratings and no betting markets. Everything comes from polls, past results and candidate data.
• Turnout. The model forecasts margins and winners, not vote counts.
• Outside money. Fundraising means money candidates raise from individual donors, as reported to the Federal Election Commission or a state’s disclosure office. Spending by super PACs and party committees, which our race pages show, does not enter the forecast, and neither do loans candidates make to their own campaigns. Where a state publishes no usable campaign-finance data, the model estimates a candidate’s money from their other traits.
• Exact district results. Presidential results by district are rebuilt from precincts, and can be slightly off where precinct and district lines do not match.
• Local quirks. Places that vote very differently down the ballot than for president, like the parts of West Virginia that still send Democrats to the legislature, are poorly served by a model built on presidential lean.
• The latest news. Candidate lists come from public sources and can lag a withdrawal or a ballot change by a day.

What comes next

A forecast is a statement about uncertainty, and no single election proves one right or wrong. A 76 percent favorite who loses is not a failure; a forecast whose 76 percent favorites lose half the time is. That is the test we will apply once results are certified: across thousands of races, did the favorites win about as often as their odds said, and did the ranges hold? The early returns from Texas, where polls missed by more than any in a competitive race there since at least 1998, suggest the shared errors described above will be at the center of that accounting.

Until then, every race’s final forecast stays on its page as it was on the morning of Nov. 3, alongside the polls and the fundamentals behind it.