Financial Sentiment and Stock Price Dynamics: An Exploratory LLM-Based Analysis of the High-Attention Media Event of President Macron at Davos

This paper provides an exploratory investigation of financial sentiment extracted by large language models (LLMs) and its association with stock-price reactions during a short-lived, high-attention media event. Focusing on 39 financial news articles surrounding President Emmanuel Macron’s appearance at the World Economic Forum in Davos, we compare three generative LLMs, GPT-5.3, Mistral Large 3, and Llama 3.3, with FinBERT. The results show that sentiment scores generated by GPT-5.3 and Mistral Large 3 are positively associated with cumulative abnormal returns (CARs), whereas Llama 3.3 shows no significant association. These findings provide exploratory evidence of heterogeneous LLM sentiment signals during an extreme media event. Moreover, leave-one-day-out (LODO) analyses indicate that the estimated relationships for GPT-5.3 and Mistral Large 3 remain broadly stable, whereas the Llama 3.3 coefficient is sensitive to excluding specific trading-day clusters. However, given the small sample and limited number of trading-day clusters, the results should not be interpreted as evidence of generalizable return predictability, causal effects, or broader market-efficiency mechanisms.