Datasets:
Published before the result: US 2026 midterms, market prices next to polls (CC BY 4.0)
Every price and every poll in here was published before the outcome exists. The election is on 3 November 2026.
That is the whole design. The market's odds and the pollsters' numbers are timestamped in public now, while nobody knows the answer, so that after 3 November anyone can check both against the result without taking our word for anything. The result file lands in v2, after the count, and the loop closes.
What is in it, measured on the viewer:
|-------------------------------|-------------------------------------------------------------|
| rows, across 12 configs | 7,002 |
| national generic-ballot polls | 363, from 63 distinct pollsters |
| press headlines | 247 |
| Polymarket contracts | binary control markets and seat / vote-margin distributions |
One race in the bundle is already decided, and it works as a small end-to-end example: the Texas Republican Senate primary of 26 May 2026, with 409 price observations leading up to it and the result verified against four independent outlets.
Two things we do not do. We do not average pollsters into a single line, and we do not subtract market probability from poll vote share. The first destroys the disagreement, which is the signal we are after. The second produces a number with no unit: the market prices who controls the chamber, the poll measures vote share, and in 2012 the Democrats won more votes and fewer seats.
Live panel, updated daily: https://www.afos-analytics.com/en/dashboard/us
Weekly editorial on the same data: https://www.afos-analytics.com/en/weekly/us
The method has been checked against eleven elections across four continents, archived with permanent DOIs:
https://dataverse.harvard.edu/dataverse/afos-analytics
CC BY 4.0. If you find an error, open a discussion here. We publish errata rather than rewriting history.