Benchmark
Forecast-Dojo
How well does a forecasting agent revise an event probability when new reporting becomes available?
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The authors reconstruct 1,568 resolved Polymarket events and arrange 18.8 million news articles by date. They split events chronologically for training and evaluation.
Methodology
An agent revisits a question at successive dates. Researchers can compare those forecasts with historical market probabilities and measure how research tools or a persistent notebook change performance and cost. These are replays of resolved events. A model’s training cutoff matters when interpreting historical forecasts.