Corporate fuzzy disclosure, investors’ limited attention, and capital market overreaction: An LLM-based index construction and empirical study

This paper studies fuzzy language in corporate annual reports as a knowledge-processing friction under limited investor attention. We define fuzziness as an overall hard-to-understand barrier that is distinct from readability or textual complexity and from dictionary-based uncertainty. We develop a Large language model-based pipeline to construct a fuzzy disclosure index ( 𝐹 ⁡ 𝐼 ⁢ 𝑋 ) from Management’s Discussion and Analysis texts of China’s A-share firms. Annual reports from 2002–2007 are used only for out-of-sample calibration, while the empirical analyses focus on reports released during 2008–2025. After merging text-based measures with market outcomes and controls, the baseline sample contains 56,565 firm–year observations. Using double machine learning with firm and year fixed effects, we find that higher 𝐹 ⁡ 𝐼 ⁢ 𝑋 is associated with higher trading volume and turnover but lower abnormal returns in the announcement quarter. This pattern remains after controlling for textual complexity and uncertainty. In the following quarter, the trading effect weakens and abnormal returns partially reverse. The trading effects are stronger when non-institutional ownership is higher, and prior-year exchange inquiries predict higher subsequent fuzziness