Unveiling Economic Sentiment: Using a Large Language Model for Economic Sentiment Analysis from Monetary Policy Reports

Abstract

This thesis explores the transformative potential of large language models (LLMs) in economic sentiment analysis, focusing on the Swedish central bank (Riksbank) as a case study. Utilizing the advanced LLaMa-2-7b-chat-hf model, we extract and analyze sentiment from the Riksbank’s monetary policy reports. Our methodology involves fine-tuning the LLM with minimal annotated data, demonstrating that contemporary LLMs can effectively overcome traditional challenges of the sentometrics field such as dependency on extensive datasets. Our fine-tuned model achieves high accuracy and F1 scores in classifying both economic and monetary policy sentiments. We further validate the performance of our extracted economic and monetary policy sentiments, called Net Score Index (NSI) and Net Monetary Policy Index (NMPI), through their strong correlation with GDP fluctuations and CPIF changes in Sweden during the 2014-2023 period. This paper therefore bridges the existing gap between the traditional lexicon-based approaches dominantly used by researchers in this field and modern Natural Language Processing (NLP) techniques. We then explore the application of our extracted sentiments by answering key monetary economics questions, investigating how sentiments extracted from the Riksbank’s communication impact the financial markets, the future monetary policies, and consumers. Our empirical findings reveal that positive current monetary policy sentiments significantly increase Swedish Treasury bill yields, particularly with longer maturities, while negative sentiments have a muted effect. Additionally, we identify significant impacts of negative sentiment on the SEK/USD exchange rate, highlighting the nuanced effects of central bank communication on financial markets. Furthermore, our analysis shows that the forward-looking NMPI can effectively predict future policy rate changes, demonstrating strong positive correlations with the Repo rate, and therefore validating the effectiveness of the Riksbank’s forward guidance. This research therefore complements existing literature by providing empirical evidence of the impact of central bank communication, within the Swedish context. It underscores the importance of central bank communication in influencing market behavior and expectations. It also contributes to existing literature in the field of sentometrics by validating the efficacy of LLMs in economic sentiment analysis.