It remembers districts
The trees learn to recognise districts they have seen before. On places the model has never seen, its district error rises from about 10 to about 20 points.
CENSUS-26 · LIMITS
What the census cannot see
A model that knows only who lives in a place is blind to some things by design. Here is where that blindness shows, and how much the missing clues would add.
01 The blind spot
A census model cannot see a gerrymander
Nine states vote on new maps this year. Map-drawers work from past election returns, precisely to exploit what demographics miss, such as Hispanic South Texas voting Republican. To the census the new lines barely matter; add 2024 presidential results by district and the redraws appear.
- TX+0.2−2.4
- FL+0.6−1.7
- LA−0.9−1.0
- TN−0.7−0.9
- AL−0.6−0.8
- OH+0.2−0.3
- NC0.0−0.2
- UT+0.9+0.9
- CA+1.9+2.9
Net effect of the new maps: +1.6 expected Democratic seats with the census alone, −3.5 once presidential results are added. How gerrymandering works →
02 Extra clues
What past votes and incumbents would add
The same model, given extra information it deliberately does without. Each clue helps, but less than you might expect: most of what decides a House seat is already written in who lives there.
| Information used | Brier score (lower is better) | Seats called | 2026 Democratic seats | Democratic majority |
|---|---|---|---|---|
| Census only (Census-26) | 0.049 | 93.6% | 239 | 75% |
| + incumbent on the ballot | 0.041 | 95.2% | 233 | 74% |
| + last presidential result | 0.039 | 95.0% | 235 | 73% |
| + both | 0.034 | 95.7% | 229 | 67% |
With every clue the Brier score improves from 0.049 to 0.034. For reference, in 2018 and 2022 the census alone scored 0.061 / 0.057, while MP-26 and FiveThirtyEight, built on polls, scored 0.028 / 0.032 and 0.027 / 0.032.
03 Fine print
Where it is weaker than it looks
Three caveats to keep in mind when you read any Census-26 number.
It does not see candidates
Candidates, money and incumbency are invisible to it. That is why the Senate, where one name on a statewide ballot weighs more, carries a larger error than a House district.
The wave is a guess
The national mood is a historical average, not a measurement. Since 2002 the swing against the president's party has ranged from a small gain to a 17-point loss, and every seat moves with it.
Not to replace the poll-based forecast. It shows how much of American politics is already sorted by geography and demography, and the districts where that stops being true are where campaigns actually matter.
NextSources →