Does News Tone from Large Language Models (LLMs) Predict Stock Returns? Evidence from Korea

This paper investigates whether textual tone derived from large language models (LLMs) can predict future stock returns. Using Korean news articles, we employ five LLMs to extract textual tones: BERT (KrFinBERT), DistilBERT, RoBERTa, ELECTRA, and Llama3. The results show that LLM-derived news tones significantly forecast next-day stock returns. Portfolio analyses further indicate that long–short portfolios constructed on the basis of LLM news tone produce statistically significant profits. In particular, the BERT-based high–low portfolio achieves the highest daily return of 0.242% (5.324% per month). These profits are especially pronounced among firms with lower size, analyst coverage, and foreign ownership, indicating that LLM-based signals are particularly effective under high information asymmetry, and for growth stocks, where fundamental valuation is inherently more difficult. These findings suggest that LLM-based textual analysis provides valuable investment signals and serves as a powerful tool for financial decision-making.