Before voting began in New York's and Maryland's June 23 congressional primaries, we published a forecast for every contested race, including the expected winner, margin, and confidence level. We used AI to simulate each district's electorate from voter-file profiles.
The forecasts correctly called 9 of the 10 winners, including a close race, a crowded Manhattan primary, and an incumbent's defeat. The results and the missed call are listed below.
How the forecast combines three sources
We sample a district's registered voters and use AI to simulate their responses, based on age, location, and other voter-file information. We combine those estimates with fundraising data and available public polls:
- The voter simulation. AI estimates candidate support from the sampled voters' profiles.
- Fundraising. We give fundraising a small weight as an indicator of campaign organization and reach.
- Public polling. Where a credible independent poll is available, we include it as another measure of candidate support and name recognition.
Each source has limits. A simulation may underrate a well-known candidate, fundraising may overrate a wealthy one, and a poll may miss a late shift in support. Combining them reduces our dependence on any single estimate.
Results from 23 past primaries
Before forecasting the 2026 races, we tested the method against 23 past New York and Maryland primaries. We withheld those results from the forecasting process and compared the predictions with the outcomes.
In New York, the method called 14 of 17 winners, about 82%, with a typical vote-share error of roughly 6 points. It identified Ocasio-Cortez's 2018 win over Crowley and Bowman's 2020 win over Engel, both upsets that contemporaneous polls missed.
The one we missed
Our miss was NY-13, spanning Upper Manhattan and the Bronx. We gave a narrow edge to the incumbent, Adriano Espaillat. The winner was his energized, DSA-backed challenger, Darializa Avila Chevalier.
NY-13 was our lowest-confidence New York call. The simulation put the candidates nearly even, but Espaillat's fundraising lead and outside support tipped the published forecast toward him. We had excluded the only public survey, which came from a progressive-aligned firm and showed a dead heat, because we could not establish its neutrality. That survey came closer to the result than our final call. In this race, adding the fundraising adjustment made the forecast worse.
The scorecard includes all ten calls, including NY-13.
NY-21 was the only Republican primary in the forecast. The method had been tuned on Democratic primaries.
Polling races with little public data
Most races in this forecast had no public poll. Where polls existed, they were often campaign internals. The cost of conventional polling leaves many down-ballot campaigns with limited information about voter preferences.
Synthetic polling can produce estimates in days at a lower cost, including in races without public polling. Campaigns can also use the simulated electorate to test your message and find the voters most likely to move. Where a field poll is available, it provides another source to compare with the simulation.