An election model that named the winner in 74 of 81 provinces in a backtest of the 2023 presidential runoff.
We model how people will decide: voters, members, customers. Before a model can predict the future it has to know the past. Ours called past elections in backtests.
In a professional chamber election we published our forecast as a preprint one day before the vote, so that it could not be changed afterwards. The race was close to a coin flip, and we wrote down in advance which way it leaned: the challenger's list, with a 55% chance, 32 to 29 in the assembly. The ballot box gave 35 to 26.
Ballot-box tallies; the official announcement is pending. Individual committees were harder to call: 15 of 24. A single election does not show that the method will always hold.
US midterm elections, 3 November 2026: all 435 House seats and 35 Senate seats. The forecast was published before any votes were counted and will be scored when the result is in.
Registered forecast: Democrats win the House, 96% probability; the Senate is close, Democrats 54%.
Antalya Chamber of Commerce and Industry election, 18 October 2026: four lists, 49 professional groups, 124 assembly seats. The forecast was published before the polls opened and will be scored when the result is announced.
Registered forecast: the red list is the largest group in the assembly, 82% probability; outright majority, 55%.
Memorization test: with candidate names hidden, the result stayed within about 2 points. The model isn't recalling the past; it reasons in the voter's place.
A voter simulation built from thousands of personas. As the protocol improved, the Alanya 2024 error fell from 6.7 to 4.1 points. Backtested on seven pairs of national elections.
Sourced setup and operating costs for four Istanbul locations, a three-year model, customer persona simulation and a comparison against deposits and the stock market.
90 days, 2,160 hours, 3,557 headlines. News volume leads BTC volatility by about an hour; sentiment and surprise measures carry no signal.