Scorecard · Results as of · Final forecast

How our forecast did

Results and calls come from civicAPI. Counts are still being finalized, and a few races remain uncalled, so these numbers can still move a little. Races with candidates from only one party are left out.

The chambers

Senate totals include the seats that weren't on the ballot this year. The colored marker is the number of Democratic seats called so far; the line is our forecast's most likely number. For the national House vote, our final forecast expected Democrats to win by 7.5 points (80% range: 5.0 to 10.0); the Cook Political Report's vote tracker had Democrats up 12.8 as of November 7.

When we said 90%, did it happen 90% of the time?

Forecast vs. result, race by race

How the other forecasters did

"Score" is the Brier score: the average squared gap between the chance given and what happened (1 for a win, 0 for a loss), so a confident miss costs much more than a Toss-up. To score categories as chances, every forecaster's ratings, ours included, are converted the same way: Safe = 97.5%, Likely = 87.5%, Lean = 70%, Toss-up = 50%. That conversion is our convention, not something the outlets publish. PredictIt's prices are chances already, so it's compared with our actual odds. "Their Toss-ups" counts races they rated a Toss-up and we didn't.

Ratings are each outlet's final published ratings, collected before Election Day for our Compare page; PredictIt prices are from its free data feed on . Margins aren't compared here because most of these forecasters publish categories, not margins.

Every miss

Every called race

"Expected" is our final forecast's most likely margin; "vs. forecast" is how far the result landed from it, shown only for races with a near-final count and one candidate from each major party. Click a race for its full page.

What we're taking from this

The forecast saw the direction right: it had Democrats favored for the House () and the Senate (), and those calls held. Where it missed was size. It expected roughly a normal midterm electorate, and far more people voted than that, more of them in Democratic-leaning places. The model allowed for races to miss together, since every simulation shares one national environment, but not by this much, so it was too sure of Republicans in seats that moved toward Democrats.

The model will be rebuilt around those lessons once the voter files, which show who actually voted, come out over the coming months. Until then, the final forecast stays up exactly as it was on November 3, so anyone can check it against the results. How the model works · The final forecast · The Turnout, our newsletter.