Interpreting the Interpreter: Can We Model post-ECB Conferences Volatility with LLM Agents?

Traditional high-frequency identification of monetary policy communication effects operates ex-post, precluding evaluation of alternative strategies before publication. In this paper, we introduce an ex-ante framework that simulates heterogeneous market reactions to central bank communication before release. Our methodology employs Large Language Models (LLMs) to construct an agent-based simulation of 30 synthetic traders with heterogeneous risk preferences, cognitive biases, and interpretive styles. These agents process European Central Bank (ECB) press conference transcripts and forecast Euro interest rate swap rates across 3-month, 2-year, and 10-year maturities. Cross-sectional forecast dispersion provides a model-based measure of market disagreement, validated against realized overnight index swap (OIS) volatility. Analyzing 283 ECB press conferences (June 1998-April 2025), we document Spearman correlations of approximately 0.5 between simulated and realized disagreement, rising to 0.6 under iterative prompt optimization. Results prove robust across prompting strategies, are temporally stable across training and holdout samples, and fare significantly better than simple language complexity scores. For central banks, the framework provides an operational tool to anticipate communication-induced volatility before release, thus enabling ex-ante language refinement. For researchers, it offers a micro-founded alternative to reduced-form event studies, explicitly modeling the heterogeneous interpretive processes through which policy signals are transmitted to asset prices.