Non-English textual analysis with large language models: Analysts’ use of MD&A sentiment in earnings forecasting

Abstract

After a decade of research on the information content of firms’ financial disclosures, do analysts actually use this information? This study investigates whether analysts incorporate information from the Management Discussion and Analysis (MD&A) sections of annual reports into their earnings forecasts. We also explore the effectiveness of Large Language Models (LLMs) for non-English text sentiment analysis, particularly where a traditional dictionary approach is impossible. We demonstrate that LLM-extracted sentiment is robust and mitigates concerns about look-ahead bias inherent in their training data. Analyzing MD&A sections from South Korean firms, we find that positive sentiment is negatively associated with analyst earnings forecast errors, indicating that analysts do incorporate it into their forecasts. However, negative sentiment is positively associated with forecast errors, suggesting incomplete analyst response. Leveraging LLMs for textual sentiment analysis, our findings suggest that analysts partially, but not fully, utilize the textual sentiment of MD&A sections in their earnings forecasts.