Election Win Probability Charts Aren't Vote Shares

When a forecast says Democrats have an 85 percent chance of winning the House, that is not a prediction that they will win 85 percent of anything. It is a statement about how often one side ends up with a majority when the model plays the election out thousands of times. The vote itself is expected to be much closer.

How much closer? FiftyPlusOne's 2026 forecast, published on August 3, put Democrats at an 85 percent chance of House control and, in the same post, projected the national House popular vote at Democrats plus 7 points with an uncertainty band of plus or minus 5. So the headline number is 85 and the margin behind it could plausibly be anywhere from a 2-point race to a 12-point one. Those are two different measures on the same scale, and the scale is the trap.

A win probability is a counting exercise. The model takes polling, fundamentals like presidential approval and the economy, and an estimate of how wrong the polls might be, then simulates the election over and over. Decision Desk HQ runs a million simulations of every district. FiftyPlusOne renders 40,000 of them, one square per simulation, so the spread stays visible. The probability is simply the share of those runs where a party clears 218 seats.

That counting step is where the compression happens. Every simulation that ends with a Democratic majority counts the same, whether it is a 1-seat squeaker or a 40-seat landslide. A lead that is small but consistent produces majorities in most runs, and "most runs" prints as a large percentage.

Jessica Hullman and Steve Franconeri at Northwestern have made the point directly in their work on forecast visualizations: a small difference in vote share can translate into a big difference in win probability. The two numbers move together, but not at the same speed, and a chart that shows only the second one hides how little movement it took.

Set this cycle's figures side by side and the gap is easy to see.

The probabilities are high and they do not agree. Nate Silver has pegged the Democratic House win probability at roughly 85 to 90 percent, above prediction markets at 70 to 85 percent, which he argues are hedging toward 50/50; his Silver Bulletin model was last updated on September 21. Decision Desk HQ, as of September 22 at 2:06 PM, had Democrats at 71 percent and Republicans at 29, with the summary line "The House leans in favor of Democrats." That is a wide spread for the same election, and worth understanding on its own terms if you are reading several dashboards; we covered why these forecasts disagree so much separately.

The vote numbers behind them are ordinary. Newsweek reported in August a YouGov/Economist generic ballot of Democrats 40, Republicans 34, and a Reuters/Ipsos reading of 42 to 37. Republicans hold the chamber 218 to 212, so the seats in play are few.

And the seat forecasts carry real range. Decision Desk HQ projects about 231 Democratic seats against 204 Republican, with individual districts sitting at coin-flip odds: Pennsylvania's 8th at 50 percent, North Carolina's 1st at 52, Ohio's 7th at 54. FiftyPlusOne's median is 230 Democratic seats with simulations running from 211 to 253, which spans both a Republican hold and a rout.

The historical version of this gap is 2016. Forecasters gave Hillary Clinton a 70 to 99 percent chance of winning, according to the Annenberg School's summary of the research, while the same models projected her electoral college vote share at only 56 to 60 percent.

Sean Westwood, Solomon Messing and Yphtach Lelkes tested this in "Projecting Confidence: How the Probabilistic Horse Race Confuses and Demobilizes the Public," published in the Journal of Politics. In the first study, 4,151 participants saw a hypothetical Senate race described either as a win probability or as a projected vote share. Same race, different unit. The people shown the probability came away more certain of the result. A second study, an economic voting game with 1,171 respondents, found that higher win probabilities reduced participation, while varying the vote-share projection alone had no detectable effect.

The clearest evidence of the misreading is blunter than that. In the summary of the experiment, 8.6 percent of respondents reported the win probability back as a vote share, treating a 71 percent chance of winning as though the candidate would take 71 percent of the ballots.

This is not a one-study finding. A preregistered YouGov experiment with 2,934 French voters just before the 2022 presidential election, published in Public Opinion Quarterly in May 2025, found that the vote-share format moved people's expectations closer to the model and the actual result, while the probability format moved them about 5 points in the wrong direction.

The numbers are also everywhere. Pew Research Center counted Google News mentions of probabilistic forecasting rising from 907 in 2008 to 3,860 in 2012 to 15,500 in 2016, with cable news referencing them an average of 16 times a day.

It is also not a confidence interval. Silver's model publishes a 90 percent range for seat outcomes rather than one number, and a summary of that forecast put the range at a Democratic gain of 5 to 45 seats. That range is the uncertainty. The probability is what you get after the uncertainty has been collapsed into a yes or a no.

When you see a probability bar, find the two other numbers before you form a view: the projected margin and the range around it. If the page does not publish them, treat the probability as incomplete. Decision Desk HQ's overview, for instance, leads with 71 percent and does not put a national vote margin next to it.

Then ask what the probability is a probability of. Chamber control is a threshold, 218 seats, and a forecast can be very confident about the threshold while being genuinely unsure about half a dozen districts underneath it, several of which sit between 50 and 54 percent right now.

And remember what a single-event probability does to the brain. As David Leonhardt put it in Pew's coverage, people understand that rolling a die 100 times produces some 1s, but shown a probability for one event they ask whether it is going to happen or not. The election happens once. A 29 percent chance is not a rounding error; it is roughly the odds of drawing a heart from a shuffled deck.