Keqin Li

This paper studies the role of linguistic instability, which we term the semantic shift, in explaining the cross section of expected stock returns. Using 3 million news articles from over 500 media outlets spanning 2010 to 2024, we construct a monthly semantic shift measure for each firm via a pre‐trained large language model in China. Empirically, we find a robust and significant return spread for low‐minus‐high semantic shift portfolio, with a monthly excess return as high as 1.39% after controlling for common risk factors. We further construct a semantic shift factor and a tradable mimicking portfolio and document consistent results. Exposure to negative innovations in the SemShift‐sorted portfolio return carries a larger estimated price than exposure to positive innovations. Our findings highlight media narrative dynamics as a key dimension in asset pricing beyond traditional fundamentals.