LLMs Analyzing the Analysts: Do BERT and GPT Extract More Value from Financial Analyst Reports?

Abstract

This paper examines the use of Large Language Models (LLMs), specifically BERT-based models and GPT-3.5, in the sentiment analysis of Korean financial analyst reports. Due to the specialized language in these reports, traditional natural language processing techniques often prove insufficient, making LLMs a better alternative. These models are capable of understanding the complexity and subtlety of the language, allowing for a more nuanced interpretation of the data. We focus our study on the extraction of sentiment scores from these reports, using them to construct and test investment strategies. Given that Korean analyst reports present unique linguistic challenges and a significant ‘buy’ recommendation bias, we employ LLMs fine-tuned for the Korean language and Korean financial texts. The aim of this study is to investigate and compare the effectiveness of LLMs in enhancing the sentiment analysis of financial reports, and subsequently utilize the sentiment scores to construct and test investment strategies, thereby evaluating these models’ potential in extracting valuable insights from the reports. The code is available at https://github.com/msraask3.