Narrative Factor Extraction Using Large Language Models and Its Implications for Asset Pricing

Drawing on Shiller's theory of narrative economics, this article puts forward the "narrative-sentiment-pricing" (NSP) transmission mechanism and divides the transmission process from narrative to asset pricing asset pricing into three stages, namely narrative germination, sentiment diffusion and expectation solidification. Then, through non-diversifiability and heterogeneity relative to classical factors, it is argued that the narrative constitutes an independent source of systematic risk. At the factor construction level, this paper proposes a narrative factor (NARR) extraction method based on large language models (LLMs). At the level of asset pricing testing, this paper designs three testing paths. Incorporate NARR into the Fama-French multi-factor model for incremental information testing; place NARR under Merton's ICAPM framework to distinguish between rational hedging demand and behavioral deviations; and finally, introduce NARR into the stochastic discount factor (SDF) framework for a no-arbitrage test. The above paths jointly build a complete evaluation system to verify the independent pricing power of narrative factors, and also provide practical reference for subsequent empirical research on narrative factors.