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

This paper develops a novel method to simulate financial market reactions to European Central Bank (ECB) press conferences using Large Language Models (LLMs). We create a behavioral, agent-based simulation of 30 synthetic traders, each with distinct risk preferences, cognitive biases, and interpretive styles. These agents interpret ECB communication and forecast Euro interest rate swap levels at 3-month, 2-year, and 10-year maturities, with cross-sectional variation serving as a measure of market uncertainty. Using a comprehensive dataset of 283 ECB press conferences (1998-2025), we evaluate three prompting strategies: naive zero-shot, few-shot (enriched with historical context), and an advanced iterative 'LLM-as-a-Judge' framework. We find that even the naive approach achieves substantial correlations (roughly 0.5) between synthetic disagreement and actual post-announcement OIS volatility, particularly for medium- and long-term maturities. The 'LLM-as-a-Judge' framework further enhances performance upon initial application, reaching correlations of almost 0.6 and identifying key drivers of volatility. Our results demonstrate that LLM-based agents can effectively capture the heterogeneous interpretive processes through which monetary policy signals propagate to financial markets, moving beyond traditional reduced-form event studies. This provides central banks with a practical tool to ex-ante evaluate communication strategies and anticipate market reactions, while offering researchers a micro-founded framework for understanding expectation formation in response to central bank communication.