Why 2026 Midterm Forecasts Disagree So Much About Who Wins the House

Three forecasts of the same midterm elections are live this month and they look irreconcilable. A Cornell University model presented on September 3, 2026 gives Democrats an 8-in-10 chance of winning the House. Silver Bulletin's FLIPR model, updated September 18, puts Democrats at 59% to win the Senate. Polymarket, where people bet real money, had Republicans favoured for the House at about 62% on September 11 and Democrats at roughly 90% a week later.

The reason those numbers do not line up is not that two of them are broken. A win probability is not a prediction of who wins. It is a statement about how often a forecaster expects to be wrong across many forecasts, which means you cannot grade it against one election. Once you read the numbers that way, most of the apparent gap dissolves: some of these figures describe different chambers, some were produced on different days from different information, and one of them is not a forecast at all but a price.

A modern election forecast does not calculate a winner. It builds a world, runs it thousands of times, and counts.

The Cornell model, led by government professor Peter K. Enns, forecasts all 435 House districts individually rather than working down from national indicators. Each district gets its voting history, expert competitiveness ratings, campaign donation data, incumbency status and any redistricting changes, and the whole map gets nudged by state and national polling on voting intentions. Then the team runs the map thousands of times, letting the uncertain parts fall differently each run.

The central result is a 226 to 209 Democratic seat split, with simulations ranging from 206 to 258 Democratic seats. The 80% is simply the share of those runs in which Democrats end up with a majority. One run in five did not.

Silver Bulletin's FLIPR does the same thing at larger scale, about 40,000 simulations, and models the House, Senate and governorships jointly so it can answer combined questions such as the chance Republicans hold the Senate while losing the House. Before simulating, it adjusts polls for the gap between registered-voter and likely-voter samples, applies house effects to correct for pollsters that lean one way, and weights more heavily the pollsters with better records. Its adjusted generic congressional ballot had Democrats up 8 points nationally.

A prediction market does none of this. Polymarket's number is the price of a contract that pays out if an outcome happens. If Democratic Senate control trades at 61%, a $100 bet returns about $165. The price is whatever buyers and sellers settle on, which can incorporate polls, models, gossip, and someone's hunch about turnout, all at once and with no published method.

Three things are going on in that grid, and only one of them is a real disagreement.

The first is that the 59% and the 80% are about different elections. One is the Senate, the other the House. Republicans are defending a 53 to 47 Senate majority with genuinely close races in Maine, Texas, Michigan and Ohio, while the House is a different map with a different history: the president's party usually loses seats at a midterm. Nothing requires the two chambers to carry the same probability.

The second is timing. Cornell's data was collected at least 100 days before November 3, so it predates the September polling that moved everything else. Newsweek reported Silver Bulletin's Democratic Senate probability rising from 50.6% on September 6 to 58.6% on September 14, following Democratic leads of 2 to 5 points in Texas polls, 4 points in Ohio and 2.9 points in Michigan. Polymarket's House price flipped from Republican-favoured to Democrat-favoured over roughly the same week. A number from September 3 and a number from September 18 are answers to different questions.

The third is that markets do not speak with one voice even about themselves. Checked on September 19, Polymarket's balance-of-power market priced a Democratic sweep of both chambers at 60%, a Republican Senate with a Democratic House at 31%, a Republican sweep at 9% and a Democratic Senate with a Republican House at about 1%. Add up the ones with a Democratic House and you get roughly 61%. A standalone House market on the same platform had Democrats at 88%. Same site, same day, a 27-point gap, across $12.6 million of trading volume. That is a fact about how thinly separate contracts trade, not a fact about the election.

The misreading worth naming is treating 80% as "Democrats win the House." It is not. It is closer to a weather forecast: a 20% chance of rain is not a promise of a dry afternoon, and nobody calls the meteorologist a fraud when it rains. A forecaster who says 80% on a hundred separate races and is well calibrated should see the favourite lose about twenty of them. If the underdog never won, the forecasts were too timid.

This is why the argument that erupts after every election where the favourite loses, that the model "failed," does not follow from a single result. The only honest test is the track record. Cornell reports 96% accuracy across more than 6,500 House races since 1996, that outcomes fell inside its simulated range 95 times out of 100, and that it called the winning party in the previous 14 congressional elections. It also reports getting only about two-thirds of expert-rated toss-ups right, which is what being uncertain looks like when you write it down honestly.

There is a second confusion worth clearing. A win probability is not a margin. Cornell's 80% sits on top of a projected 226 to 209 seat split, which is a close election. A comfortable probability and a narrow expected result are entirely compatible, because the probability answers only "which side of the line," not "by how much."

That is the correct way to state a forecast, and it is also the reason the gap between these numbers is not a scandal. What is not yet known, and will not be settled by any of this, is what happens on November 3. A single election result will not tell you which of these three sources was the better forecaster. It will only add one data point to each of their records.