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

Abstract

Central banks cannot observe market reactions to their communications before release. We propose a framework in which Large Language Models simulate 30 heterogeneous traders interpreting European Central Bank press conference transcripts, yielding a measure of cross-sectional disagreement among synthetic agents. Across 293 Governing Council events from 1998 to 2026, this measure correlates at approximately 0.5 with realized Overnight Index Swap volatility, outperforming standard text-based alternatives in explaining market reactions. LLM-implied disagreement adds information beyond volatility clustering and remains robust in out-of-sample validation on genuinely unseen conferences from January 2025 onwards. We further show that providing historical examples of pre and post-conference volatility improves the calibration of model responses. The framework offers a practical tool for assessing, prior to release, how central bank communication is likely to be interpreted by financial markets