Science
We called it: Brazil and the 2026 US midterms
Before polls opened in Brazil and the US, synthetic users made public calls on every race. Here is how they did, wins and misses.
Two elections, one question: can synthetic users see what the polls miss?
Most research claims are hard to check.
Elections are not. There is a date, a count, and a winner.
So we put synthetic users on the record.
Before polls opened in Brazil and the US, our Decision Desk published a number for every race it covered.
Not a poll. Not an average of polls.
A population built to match each electorate, given what voters would have known in the final days.
Here is how it went, wins and misses.
Brazil: the call nobody else made
On 20 October, five days before the runoff, we published Lula by 4.3.
That came from 600 synthetic users built to match Brazil's electorate.
The polls disagreed. Seven of eight had Flávio ahead. The eighth had it level.
| Margin | |
|---|---|
| Poll average | Flávio +2.2 |
| Our call | Lula +4.3 |
| Official count | Lula +3.5 |
We were within 0.3 points of each candidate's share.
The poll average was 5.7 points off and named the wrong winner.
It was the first time since 1989 that the round-one runner-up won Brazil's presidency.
The polls were not careless. They leaned the same way, so nobody was placed to say otherwise.
The misses: we called 3 of 6 governor runoffs. Tocantins and Rio Grande do Norte went the other way.
The same Sunday, our three calls on Bulgaria's first round all held. Our margins there were too wide.
Read the Brazil and Bulgaria breakdown
The US midterms: we saw the wave coming
On 2 November, we published a number for every Senate and governor race.
Each came from 200 synthetic users built to match their state.
We called Senate control for the Democrats at 68%. They won it.
We named the winner in 60 of 70 races decided so far.
In the six battlegrounds, the poll average missed every one, always in the same direction.
Our calls landed within about two points.
| Senate race | Poll average | Our call | Count so far |
|---|---|---|---|
| Georgia | D+9.7 | D+16.4 | D+17.8 |
| North Carolina | D+11.7 | D+22.1 | D+20.0 |
| New Hampshire | D+7.1 | D+16.0 | D+14.5 |
| Minnesota | D+5.8 | D+11.9 | D+10.5 |
| Maine | D+2.8 | D+6.3 | D+7.9 |
| Ohio | D+3.0 | D+3.4 | D+4.4 |
We also called three races against the polls. Georgia and Kansas governor, and the Alaska Senate. All three are going our way, with Alaska still counting.
The misses: ten wrong winners, including Iowa, Kansas and South Carolina in the Senate.
Safe seats came out too lopsided. Red states that swung hard were undercounted.
Popular cross-party incumbents, like Phil Scott in Vermont, beat our calls.
In Mississippi, we under-read Black voters. It is the largest known gap in the method.
In short: it knew the Democrats would win. It could not always see by how much.
Why this matters for research teams
Elections are the hardest public test we know.
The question is simple. The answer arrives on a fixed date. Everyone can check it.
Synthetic users will not replace organic research. But they can do something useful alongside it.
They can hold a view when every other instrument leans the same way.
That is the same job our customers ask of them:
- Explore a hypothesis in interviews.
- Test it across a larger audience.
- Go back into the motivations behind the numbers.
One race is a data point, not a track record. So we keep score in public, misses included.
Depth and scale, in the same place.
Every voter in these studies is synthetic. Figures describe simulated electorates, published before polls opened as testable estimates. US results are the count as of the morning of 4 November 2026.