(Credit: © scoutori - stock.adobe.com)
A 50-50 Senate Forecast Could Really Mean 43% or 58%
In A Nutshell
- A research team tested 120 versions of a Senate forecasting model, and 117 fit past elections about equally well.
- Those 117 models gave Democrats anywhere from 43.2% to 57.6% odds of winning Senate control in 2026.
- One assumption, how much the national mood sways races, drove about 89% of the differences, and only eight election cycles exist to test it.
- The authors recommend publishing a range of results next to a single best guess.
Two researchers compared 120 versions of a forecasting model for the 2026 U.S. Senate, all tested against the same historical election data. Of those, 117 predicted past elections about equally well. Yet when pointed at 2026, those 117 models disagreed on who is favored to control the Senate, with Democrats’ odds ranging from 43.2% to 57.6%.
Jack T. Rametta of the Mercatus Center at George Mason University and Christopher Hare of the University of California, Davis ran the experiment using Senate races from 2010 through 2024, in a paper accepted by PS: Political Science & Politics. Their single best-guess model put Democrats at a 50.6% chance of winning control, essentially a coin flip. That figure assumes Vice President Vance breaks any 50-50 tie, so Democrats need 51 seats. It counts a win by independent Dan Osborn in Nebraska as a Democratic seat. Without that assumption, the number falls to 47.9%.
Election forecasts often read like sports scores, precise figures that are easy to repeat as settled fact. Rametta and Hare argue that a figure like 50.6% reflects one modeler’s choices among many defensible options, such as how strongly the national mood should sway individual state races.
Only Eight Election Cycles Inform a Key 2026 Senate Forecast Assumption
Race-by-race Senate forecasts rest on a few judgment calls: how much being the incumbent helps, how much a state’s usual leaning matters, and how hard the national mood, such as presidential approval or which party voters favor overall, pushes individual races.
It might seem like 269 Senate races from 2010 through 2024 would be plenty of data to settle those questions. Not quite, because every race in a given election year shares the same national conditions. If the climate in 2018 favored Democrats, it affected every Senate race that year at once. Each election year offers one reading of the national mood. Once that overlap is accounted for, the researchers estimated that the 269 races hold about as much information as 37 independent ones. And the national-mood question has to be answered mostly by comparing only eight election cycles.
117 of 120 Forecast Models Fit Past Senate Races Similarly Well
Rametta and Hare’s best-guess model blends a state’s political history, national conditions, incumbency, and polling, then simulates the whole Senate. As of August 19, 2026, it projected Democrats would hold 50.5 seats on average, with a likely range of 45 to 56.
To test how much that prediction leaned on modeling choices, the researchers built 120 alternate versions, varying the size of the incumbency bonus, how strongly national mood reaches individual races, and how much a state’s overall partisan lean counts compared with its own past Senate results. Each was scored on how well it would have predicted the Senate races from 2010 to 2024.
Of the 120, 117 scored within a close range of the best performer, so past results offered no strong reason to rule any of them out. The researchers call this group a “Rashomon set,” after the Akira Kurosawa film in which characters give conflicting accounts of the same event. Models can fit the same history equally well yet disagree about the future.
National Mood Assumption Drives Most 2026 Senate Forecast Differences
Across the 117 similar models, expected Democratic seats barely moved, staying between 50.0 and 51.0. Because the 2026 Senate is projected to land so close to the 51-seat line, that small wobble was enough to turn Democrats from underdogs into favorites, depending on the model.
One assumption stood out as the cause: how much the national mood sways individual Senate races, the same question eight election cycles cannot settle with confidence. It accounted for about 89% of the differences in control odds across the 120 models, compared with 6% for incumbency and 3% for a state’s partisan lean.
At the race level, 32 of 35 tracked races had the same predicted winner in every retained model. Three did not, led by Maine, where Democratic win odds ranged from 41% to 75%, plus Alaska and Texas.
Incumbency was the one assumption the historical data could pin down better than the others. The estimated edge of holding a seat averaged about 2.3 points of vote margin from 2010 through 2018, then fell below one point in 2020 and 2022, before partly rebounding in 2024. A few cycles cannot establish a trend, but the dip lines up with other published research pointing to a shrinking incumbency advantage.
Authors Urge Senate Forecasts to Show Ranges, Not One Number
Rametta and Hare are not calling election forecasting pointless. Their point is narrower: when several reasonable modeling choices cannot be ruled out, one confident percentage hides how much the answer depends on the forecaster. They recommend publishing the range from many reasonable models next to the best guess.
That range is not a statistical confidence interval. It carries no weighting and only shows how much the forecast moves across similarly accurate models. For a 2026 Senate election that looks like a toss-up, a range of 43.2% to 57.6% may be more informative than any single number.
Disclaimer: This article summarizes one research paper for general informational purposes. The paper is a preproof version that may change before final publication, and its forecast figures reflect the authors’ models as of August 19, 2026. They are not predictions or guarantees of any election result.
Paper Notes
Limitations
Rametta and Hare describe their findings as sensitivity within a declared set of modeling choices, not all possible sources of forecasting error. The 120 versions varied only three weights (the size of the incumbency edge, how strongly national mood affects races, and how much a state’s partisan lean counts), while personal-vote persistence, the polling-bias prior, and the polling and error setup were held the same across all of them. Several settings were chosen or checked against all eight cycles, including 2024, so the historical comparisons are retrospective tests rather than independent evidence of future accuracy. The range across the 117 retained models is unweighted and has no statistical coverage, so it is not a confidence interval and does not mean every retained model is equally likely to be right. The 37-race estimate rests on only eight cycles and is itself uncertain. The best-guess model’s edge over a model using only background factors was modest and overlapped zero in the authors’ historical tests. The incumbency trend is tentative because it rests on only a few cycles, including one rebound year. The 50.6% figure also depends on stated assumptions about Vice President Vance breaking ties and about how Nebraska’s independent candidate is counted. The paper is a preproof accepted manuscript that may change during production.
Funding and Disclosures
No funding source is listed in the paper. The authors declare no ethical issues or conflicts of interest. A footnote states that the forecast has been published daily since April 4, 2026, and that model development was supported by structured LLM (large language model) research agents, with the authors taking full responsibility for all modeling decisions and outputs.
Publication Details
Paper title: “How Model Choice Obscures Forecast Uncertainty: The Rashomon Effect and the 2026 Senate.” Authors: Jack T. Rametta, an emerging scholar at the Mercatus Center at George Mason University, and Christopher Hare, an associate professor of political science at the University of California, Davis. Journal: PS: Political Science & Politics, published by Cambridge University Press on behalf of the American Political Science Association, in a symposium on forecasting the 2026 U.S. midterm elections. The accepted manuscript was published online October 15, 2026. DOI: 10.1017/S1049096526103059. Replication data are available through the Harvard Dataverse at DOI 10.7910/DVN/WHLJNA, and the authors’ forecast updates daily at jackrametta.com/senate-forecast. The paper reports the information set as of August 19, 2026.
