Agentic Forecasting with Structured Linguistic Beliefs
The Bayesian Linguistic Forecaster (BLF) is an agentic system designed for binary forecasting that mimics sequential Bayesian updating using a Large Language Model. It uses a structured linguistic belief state to integrate evidence iteratively, combined with multi-trial aggregation and hierarchical calibration. This approach achieved state-of-the-art performance on the ForecastBench benchmark, proving that structured reasoning within an LLM can outperform traditional explicit Bayesian methods.