FOMC Statements: LLMs vs Traditional Methods for Inflation Sentiment Extraction

Abstract

This thesis evaluates the effectiveness of different Natural Language Processing (NLP) sentiment models in predicting the inflation expectations of economic agents based on the Federal Open Market Committee (FOMC) communications. We provide two new fine-tuned Large Language Models (LLMs) Flan-T5, and BERT, trained specifically for deriving the inflation sentiment of FOMC minutes and Bank of England (BoE) Monetary Policy Reports. To see how these models compares with the current sentiment methods used in economic research regarding inflation sentiment, we evaluate a total of five sentiment models: the two mentioned fine-tune LLMs, two dictionary based models, and a state-of-the-art chatbot from OpenAI. Our findings indicate inflation sentiment from the fine tuned BERT model, demonstrated slightly better predictive power regarding economic agents’ inflation expectations when controlling for other macroeconomic variables.