What Does a "90% Chance" Really Mean? How to Read Midterm Election Forecasts

The 2026 U.S. midterm elections take place on Tuesday, November 3. Forecast sites publish probabilities. One independent model put the Democrats' chance of winning the House at 94.7% and the Senate at 62.1% as of October 24, while the prediction market Kalshi showed about 90% and 63% on October 16. A model from Nate Silver's team had the House near 93% on October 12. Are these predictions that one party will win? No, and the difference matters.

A probability is not a prediction

A 94.7% chance for one side means a 5.3% chance for the other, roughly 1 in 19. Run the same election 19 times and you would expect the underdog to win about once. A 62% favorite, as in the Senate figures, loses about 3 times in 8. Neither forecast says the favorite will win; each says how often it would win across many hypothetical repeats of the same situation. Forecasters are judged over many forecasts: of all the events given 90%, about nine in ten should happen.

Many 90% events, one upset

Because even 90% leaves room for failure, a long list of 90% favorites almost always contains a surprise. If ten independent events each had a 90% chance, the probability that all ten happen is 0.9¹⁰ ≈ 34.9%, so there is about a 65% chance that at least one fails. The binomial distribution calculator gives this at once. Midterm races are not fully independent, since a national swing moves many districts together, which is why a model's seat range matters as much as its headline probability. The two events probability calculator shows how combining events changes the answer.

A 90% favorite still loses one time in ten. Over many races, that is not a surprise; it is the forecast working.

Read the range, not just the probability

The same independent model that gives 94.7% for the House projects 236 Democratic seats, with a 90% range from 217 to 259. A majority needs 218, so the bottom of the range falls just short of it. A high probability and a range that touches the line are consistent: most simulated outcomes clear 218 and a few do not. Seats are also not equally sensitive to opinion. In one published generic-ballot model, a one-point shift in the national margin is associated with about 2.6 House seats but only about 0.36 Senate seats, so a five-point swing would move roughly 13 House seats and fewer than 2 in the Senate.

Why models differ

Models use different inputs: polling averages, expert ratings, district lean, fundraising, or traders' bets. A market price is not a poll, and a poll average is not a seat forecast. Agreement between models is also weaker evidence than it looks, because many draw on the same polls and ratings, so their errors are likely to be correlated. A Senate probability near 60% is a modest edge, not a verdict. Our article on reading election polls explains the margin of error that sits underneath every forecast.

Questions to ask of any forecast

What is the date? What exactly is forecast: control of a chamber, seats, or vote share? What does the model assume about polling error? Is the probability for one race or for a combination? And remember what forecasts are not: they are not endorsements, and they should not be treated as outcomes. This article is non-partisan and does not predict the result.