Researchers employ radical data transparency in 2026 midterm forecasts

A team of experts, working at the intersection of statistics and political science, has created an interactive website with forecasts for all 470 congressional elections set for November 3. As of the morning of Friday, October 23, the Democratic Party was given a 94 percent probability of winning the U.S. House and a 72 percent chance of flipping the Senate.

The project is unique “because all data, methods, and other procedures are public — something that is not true for any other public forecast we know of,” said co-creator Gary King, Albert J. Weatherhead III University Professor and faculty director of Harvard’s Institute for Quantitative Social Science.

Powering the forecasts is a novel statistical model, introduced in a working paper King co-authored with IQSS Research Associate Danny Ebanks and Caltech professor Jonathan N. Katz. It generates probabilities based on an exhaustive data set, incorporating what the researchers characterize as “every bit of political science knowledge about the election” they could systematize with the help of machine learning. That includes the latest polling results, a “candidate effects” metric that gauges individual appeal, and history on every congressional seat on the ballot next month.

For example, site visitors can explore this year’s 35 senatorial elections via the site’s build-your-own Election Day tool. Clicking on Texas yields citations of dozens of polls on the state’s 2026 Senate race, plus a series of equations the statistical model factors in. The Texas Senate seat at play this year has been held by Republicans since the 1960s. This year, an upset by Democratic nominee James Talarico over Republican Ken Paxton was given a probability of 56 percent Friday morning, or “about a coin flip’s chance,” according to Ebanks.

The researchers’ data-rich model, validated by simulating more than 15,000 Senate and House elections as far back as the 1950s, upends a number of traditional assumptions about federal elections. One notable finding concerns popular incumbents affiliated with a sitting president’s political party. Their chance of losing during midterm elections has, it turns out, been underestimated by traditional forecasts. In the state of Maine this year, King et al.’s model produced a 75 percent probability of Senator Susan Collins’ seat flipping to Democrat Troy Jackson (as of the timestamp above).

The website went live October 15, with updates made regularly as polling and other new information becomes available. The team invites fellow data scientists to scrutinize the methodology and test forecasts for accuracy. As the co-authors have written at the top of their site: “Academics learn by making themselves vulnerable to being proven wrong.”