Assessing the Alignment of FOMC Statements with Minutes using Large Language Models

This paper pioneers a quantitative and AI-driven framework for evaluating central bank communication.The Federal Open Market Committee (FOMC) releases an official statement when meetings conclude and monetary policy is being announced, and minutes of the meeting three weeks after.The statement acts as a summary of the meeting whereas minutes provide an overview of meeting discussions.The former is contained within one to two pages while the latter covers approximately ten pages.This study leverages Amazon Nova, a large language model (LLM) on Amazon Bedrock in AWS, to evaluate the accuracy and completeness of the statements relative to the minutes.Abstractive summaries are created from the full minutes and semantically compared to the official FOMC statements using SBERT-based embeddings.By computing and analyzing these similarity scores, this study assesses if FOMC statements sufficiently capture the key themes and policy discussions reflected in the minutes.Given the significant financial market movements immediately following FOMC statements due to the rise of automated trading, ensuring transparency in central bank communication is critical for market efficiency.Findings indicate that FOMC documents consistently display general alignment, showcasing stability over time, with no discernible heterogeneity by year or season.Only two observations (n=57) qualify as statistical outliers.These findings are robust to multiple iterations of abstractive summarization.