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

Title: 解释解释器：我们能否用LLM代理对ECB会议后的波动性进行建模？

Abstract: 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. Show Chinese abstract

Abstract: 本文开发了一种新方法，利用大型语言模型（LLMs）模拟金融市场对欧洲央行（ECB）新闻发布会的反应。 我们创建了一个行为型、基于代理的30个合成交易者模拟，每个交易者具有不同的风险偏好、认知偏差和解释风格。 这些代理解读ECB的沟通内容，并预测3个月、2年和10年期限的欧元利率互换水平，横截面差异作为市场不确定性的衡量指标。 使用涵盖283场ECB新闻发布会（1998-2025）的全面数据集，我们评估了三种提示策略：简单的零样本、少量样本（结合历史背景）以及一种先进的迭代“LLM-as-a-Judge”框架。 我们发现，即使简单的零样本方法也能在合成分歧与实际公告后OIS波动率之间实现显著相关性（约为0.5），特别是在中长期期限上。 “LLM-as-a-Judge”框架在初次应用时进一步提升了性能，相关性接近0.6，并识别出波动性的关键驱动因素。 我们的结果表明，基于LLM的代理可以有效捕捉货币政策信号传播到金融市场的异质解释过程，超越传统的简约事件研究。 这为中央银行提供了一种实用工具，用于事先评估沟通策略并预测市场反应，同时为研究人员提供了一个微观基础框架，以理解对央行沟通的预期形成过程。