ForecastBench, created by the Forecasting Research Institute, is a dynamic, continuously updated benchmark of AI forecasting ability. Because it asks models to assign probabilities to events whose outcomes are not yet known, there is no risk of training contamination, and it directly compares AI forecasts against human baselines, including professional superforecasters and the general public.

We source results from the public ForecastBench leaderboard.

ForecastBench draws on two question types: “market” questions sourced from prediction markets (such as Manifold, Metaculus, and Polymarket) and “dataset” questions auto-generated from databases (such as ACLED, FRED, and Yahoo! Finance). The system runs nightly to ingest questions and resolutions, and samples new rounds of questions on a rolling basis. Forecasts are scored with a difficulty-adjusted Brier Index on an interpretable 0–100% scale, where higher is better (100% is perfect, 50% corresponds to an uninformed guess, and 0% is maximally wrong). Our chart defaults to the overall score and also exposes the separate dataset and market scores.

For full details, see the ForecastBench paper and the methodology documentation.