Electa Called 96% of 472 Declared Midterms Races Right, With Every Chamber Inside Its Range

The call, written before the count

Electa froze its midterm projection on 4 November at 15:22 UTC, while the first polls were still open. Frozen means exactly that: the run was written once, stored, and left untouched from that moment on. Everything published here is measured against that file, not against a version revised after the results were known.

The freeze put Democrats at 51 Senate seats against 47 for Republicans, with a 95% range of 48 to 55. In the House it projected 255 Democratic seats against 180, with a range of 247 to 262. In the governor races it projected 23 Democratic and 13 Republican.

Two days later the count reports 54 to 44 in the Senate, 250 to 161 in the House with 24 districts still unprojected, and 24 to 12 among governors. All three chambers landed inside the range published before election day. The full projection, the live count and the accuracy panels sit side by side in the midterms module.

Point estimate and range are different products

Measured against the frozen point estimates, the count came in three Senate seats more Democratic than projected, five House seats fewer while two dozen districts remain open, and one governor's seat more Democratic.

That is the ordinary behaviour of a distribution, not a correction of it. The Senate interval was seven seats wide and the House interval fifteen, and the count sits inside both. The number a newsroom prints on election night is the point estimate; the number a campaign should plan against is the interval, because the interval is the part of the projection that already contains its own error.

Race by race: 453 of 472

The accountability test here is granular, and it only scores what has actually been decided. A race counts as a hit when the party the model projected is the party the source has declared the winner. Races that are counted but carry no declared winner yet are not graded, in either direction: they are listed separately as still open.

Of 472 declared races, 453 went the way the model had them, 96.0%. The Senate was called correctly in all 31 races declared so far. The House came in at 390 of 407, 95.8%, and the governor races at 32 of 34, 94.1%.

Split by the party that actually won, the model identified the Democratic winner in 272 of 290 declared races, 93.8%, and the Republican winner in 181 of 182, 99.5%. The asymmetry is not a scoring artefact: 2026 was a Democratic year, and a model that had the national climate broadly right will be caught out more often on individual Democratic wins than on Republican holds. A further 32 counted races, two Senate seats, 28 House districts and two governorships, have a leader but no declaration, and none of them is being counted as a win for the model until a caller closes them.

Where it was wrong

Nineteen declared races went against the frozen projection, and they are worth naming rather than smoothing over.

Two were governorships. The model had Democrats at 22% in Ohio and 42% in Texas; both went Democratic, by roughly seven points each.

Seventeen were House districts, and they fall across the confidence spectrum. In California's 1st, 3rd and 41st districts, North Carolina's 6th and Utah's 1st, the model gave Democrats a 3% chance and they won, in California by double digits. Six more, among them Ohio's 10th and 15th, Indiana's 5th and Wisconsin's 1st, sat at 42% and broke Democratic. At the other end, Texas's 32nd district was called for Democrats at 99% and went Republican by three points. A 99% call that loses is not noise to be explained away. It is the clearest signal in this record that a district-level probability built on a national climate can be confidently wrong about a single seat.

Misses of that kind are visible house by house on the states dashboard, and we leave them on the page.

How the number was built

The engine takes two independent inputs. The first is a polling average of 15 generic-ballot surveys assembled with the RealClearPolitics method, which put Democrats 10.4 points ahead nationally. The second is digital sentiment measured from public conversation across platforms, which put Democrats 15.0 points ahead. The model blends them 50/50 into a 12.7-point Democratic national margin.

That climate is then pushed through a district-level engine that covers 435 House districts, 33 Senate races and 36 governor races, converting the national margin and each race's local signal into a win probability and a projected margin. Every seat total ships with a 95% interval. Nothing in the output is a single number pretending to be certain.

Two design choices make this comparison worth anything. The first is versioning: between 27 October and the freeze the model stored 17 dated versions, so the record of what it believed on any given day is auditable rather than reconstructed. The second is the freeze itself, which removes the temptation to re-run the model once the shape of the result is known.

The method, the weighting and the limits of the digital input are documented on the methodology page, and the polling and sentiment inputs behind the projection are tracked on the forecasts and society panels.

What a forecast owes the public

A projection is only worth the risk of publishing it. Any model can be right in private. The value of freezing a call before the count, then grading it race by race and chamber by chamber afterwards, is that the answer cannot be negotiated later.

The count reported here rests on the running tabulation published by NBC News, with the national picture summarised in Al Jazeera's takeaways. Twenty-four House districts remain unprojected, so the House figure in this article is a snapshot and will move. When the last seats are called, the comparison will be re-run against this same frozen file, and the accuracy number will be published as it lands.

The complete projection, the live results and the race-by-race accuracy panels are open in the midterms module.