Democrats have an 8 in 10 chance of securing a majority in the U.S. House of Representatives this fall, according to Cornell political scientists whose model has correctly predicted winners of the previous 14 congressional elections.
Released Sept. 3 at the American Political Science Association (APSA) annual meeting in Boston, the researchers’ midterm forecast – based on data collected at least 100 days before the Nov. 3 election – anticipates Democrats winning approximately 226 seats to Republicans’ 209 (though simulations show as few as 206 Democratic seats or as many as 258 are statistical possibilities). A U.S. Senate forecast is in progress.
“These forecasts aren’t deterministic – there is uncertainty,” said Peter K. Enns, professor in the Department of Government in the College of Arts and Sciences and in the Cornell Brooks School of Public Policy. “But given our model’s impressive historical accuracy, if Republicans hold the House, it likely means either everything has gone their way or something unprecedented has happened.”
The team’s analysis is detailed in “A District-Level Forecast of the 2026 U.S. House of Representatives Midterm Elections,” presented during an APSA panel discussion on midterm forecasting. Co-authors with Enns are Leigh Farah, a doctoral student in public policy, and Thomas Gareau-Paquette and Claudia Miner, doctoral students in government.
Forecasting recent presidential elections, Enns in 2024 predicted the correct outcome in every state and Donald Trump’s Electoral College tally, and in 2020 called the winner in all but one state (Georgia). The House forecast differs from most academic models by predicting outcomes for all 435 congressional districts – updated to reflect recent redistricting – rather than relying solely on state- and national-level indicators.
The model considers, for example, how a district voted previously, expert assessments of the likely winner, and whether the election is competitive or unopposed, as well as campaign donations to the candidates and incumbent status. The model also weighs polling data about state- and national-level voting intentions to capture the broader mood, since the incumbent president’s party almost always loses seats during midterms.
“One of the important factors in the model is voter intentions – who people say they’re going to vote for,” Enns said. “One hundred days out, the electorate is saying they are more supportive of Democrats than two years ago.”
For historical elections, “before-the-fact” forecasts considered only data that would have been available 100 days before each contest. Thousands of simulations were used to generate a 95% confidence interval around the predicted outcomes, and the forecasted outcome in each district fell within simulated predictions 95 times out of 100.
“That gives us a very high degree of confidence in our model,” said Enns, the Robert S. Harrison Director of the Cornell Center for Social Sciences and co-teacher this fall of “Taking America’s Pulse,” a class in which students design, conduct and analyze a national-level public opinion survey.
Enns said the new forecast’s simulations account for the model’s error observed over the past 30 years of elections, encompassing more than 6,500 individual races since 1996. Through 2024, the model achieved an overall accuracy rate of 96%, including correctly forecasting two-thirds of the races experts deemed 50-50 “toss-ups” – the most difficult to predict. That level of accuracy can’t be attributed to statistical chance and outperforms any other model, the researchers said, including popular media sites updated with more recent data, such as FiftyPlusOne.
But producing a forecast is not about bragging rights, Enns said. As a research exercise, forecasts help explain midterm outcomes that the scholar Edward Tufte in 1975 described as “a mixture of the routine and the inexplicable.”
“Forecasts help us better understand election outcomes,” Enns said. “If we can put all these variables into a statistical model in advance and predict the winner, that’s an especially rigorous test.”
For example, he said, if campaign donations prove valuable in correctly forecasting election outcomes, that becomes an important signal, although further investigation would be needed to understand exactly why. Do they reflect a candidate’s popularity, or simply an ability to buy more ads?
Forecasts at a point in time – 100 days out, in this case – also are informative about campaign dynamics, including whether races are determined in the final days and weeks or long before that. And forecasts can aid “post-mortem” diagnoses of anomalies in specific districts, potentially clarifying the significance of issues now in the news: How important are views about the economy or the president’s popularity? Was there confusion about mail-in voting rules? Did progressive candidates outperform or underperform expectations? Did redistricting shift the balance of power?
“Our forecast can help answer these questions,” Enns said. “If the election unfolds as it has historically since at least 1996, the outcomes should be very much aligned with what we predict.”
