FABench: Automated Backtesting Environment for Forecasting Agents
November 2, 2026
FABench introduces an automated pipeline for evaluating forecasting agents by discovering past events, generating questions, and resolving outcomes. This scalable environment allows for continuous benchmark renewal to stay ahead of model training cutoffs using 287 initial questions across nine domains.
HOW THIS AFFECTS YOU
●
researcherYou can evaluate agents on time-sensitive forecasting tasks without waiting for future events to resolve.