Monolith Analytics

How the Monolith 2026 Nebraska forecast did

The scorecard, with the misses first

That's Not How This Works · November 9, 2026

The election-day post here said the forecasts wouldn't be off by much if they were off at all. The votes are mostly counted now (unofficial, as of November 7; the canvass won't change any of this), so here's how they did.

Two numbers on every card

Anyone who looked at the forecast page saw two numbers for each of the big races, and some people asked why the polls were in the small print. That was deliberate.

The big number used no polls. It started from the voter file: each registered voter's odds of turning out this year, based on their age, their party registration and which recent elections they'd voted in, multiplied by how voters like them in their own precinct voted the last time that race was on the ballot. Early ballots that had already come back in Lancaster County were counted as votes. It also built in the usual midterm drag on the president's party, using the average from the 1998 through 2014 midterms.

It had no way of knowing how big this particular year would be. Polls are the only thing that measures that.

The smaller number, marked "for comparison," was the same model with this year's polls added. Polls were adjusted based on how polls from the same kind of sponsor have missed in the past (Republican-sponsored polls got bumped about a point toward Democrats, independent ones were left alone).

The reason to lead with the no-poll number was to see how far it could get. Nebraska doesn't get many public polls, and the ones it does get aren't great: the last ones in 2024 had Osborn about three points better than the actual result against Deb Fischer. If who votes and how they've voted before gets you most of the way to the result, polls are a correction. If it doesn't, polls are doing the real work. The difference between the two numbers measures that.

No polls

That's four of five winners with no polling at all, and the Senate number was within a point.

Other than the Senate, the misses were all in the same direction, with Democrats doing 3 to 9 points better than forecast. It was a better year for Democrats than an average midterm, and an average midterm is what the model assumed.

With polls

The four races missed by 3.0 points on average, compared with 5.2 for the no-poll numbers in those same races, and every winner was right. Nobody polled the Secretary of State's race, so there's no polled number for it.

Adding the polls cut the misses in the House seats roughly in half, nearly erased the governor's, and fixed the NE-2 call. They made the Senate race worse, though. The polls had Osborn closer than the result, and the no-poll number was almost dead on.

Legislature

There were 17 seats where a Democratic-backed candidate ran against a Republican-backed one in November, and the forecast got 15 of them. Again, no polls.

The two it missed were District 14, where the Democratic side had a 75% chance, and District 36, at 61%. The Democratic side lost both.

For those who like a single number, the Brier score was 0.09. On that scale 0 is perfect, and calling every race a coin flip would score 0.25.

Where that leaves things

Without a single poll, the forecast got 4 of the 5 big statewide and House races and 15 of the 17 contested Legislature seats. Where it was off, it was off in one direction, and the reason is pretty clear: it assumed an ordinary midterm, and Democrats had a better one than that.

The polls were worth a few points and one House seat this year. Most of what decided these races, though, was who turned out and how those same people had voted before, and that was knowable without them.

George Box's line was that all models are wrong but some are useful. This one was wrong on the size of the Democratic vote and close on most of the rest.


Results are the state's unofficial counts as of November 7. Shares are of the two leading candidates' votes, except Secretary of State, which was forecast as a share of all votes in a four-way race. Forecast numbers are from the version locked on the morning of November 3.

First published on That's Not How This Works. Subscribe there to get each post by email.