Presentation Information
[2N4-GS-10x-03]Empirical Analysis of the Explanation for LLMs-based financial sentiment scores
〇Kento Tanemura1,3, Kenji Kubo1,3, Kei Nakagawa2,3 (1. The University of Tokyo, 2. Osaka Metropolitan University, 3. Matsuo Institute.)
[[online]]
Keywords:
Finance,Large Language Models,Sentiment Analysis
This paper examines whether large language models (LLMs) extract fundamental information when analyzing managers' MD\&A texts in Japanese.
We estimate regression models where LLM sentiment scores are explained by Fama-French five factors and firm characteristics, and evaluate their incremental explanatory power for forward returns using three approaches: sentiment-only regression, linear residual regression, and nonlinear residual regression.
We analyze Japanese firms (2016--2024) using multiple LLMs and baseline methods.
LLM sentiments exhibit positive correlations with Size and Profitability, indicating partial overlap with existing factors.
After controlling for Fama-French five factors, some LLMs retain incremental explanatory power for forward returns, suggesting they capture text-specific qualitative information beyond fundamentals.
Year-by-year analysis reveals that regression coefficient signs fluctuate across years for all models including LLMs, indicating limited sign stability.
A notable biennial periodicity is observed, with higher $R^2$ values in even years.
Furthermore, LLM dependence on Profitability declines over time, possibly reflecting improvements in MD\&A disclosure quality.
We estimate regression models where LLM sentiment scores are explained by Fama-French five factors and firm characteristics, and evaluate their incremental explanatory power for forward returns using three approaches: sentiment-only regression, linear residual regression, and nonlinear residual regression.
We analyze Japanese firms (2016--2024) using multiple LLMs and baseline methods.
LLM sentiments exhibit positive correlations with Size and Profitability, indicating partial overlap with existing factors.
After controlling for Fama-French five factors, some LLMs retain incremental explanatory power for forward returns, suggesting they capture text-specific qualitative information beyond fundamentals.
Year-by-year analysis reveals that regression coefficient signs fluctuate across years for all models including LLMs, indicating limited sign stability.
A notable biennial periodicity is observed, with higher $R^2$ values in even years.
Furthermore, LLM dependence on Profitability declines over time, possibly reflecting improvements in MD\&A disclosure quality.
Comment
To browse or post comments, you must log in.Log in