Incremental Information Mining of Inquiry Letters for Financial Distress Prediction --- Experimental Evidence Based on LLMs

math_shortcuts This study researches the incremental value of textual information from regulatory inquiry letters and their responses for financial distress prediction (FDP). Specifically, this paper dissects incremental information from four perspectives: sentiment, readability, topic, and large language models (LLMs)-enhanced information. Taking China's listed companies from 2014 to 2023 as samples, this study constructs nine models by integrating the classic LightGBM model with incremental information across different dimensions. The findings indicate that using inquiry letter text data yields an AUC value of 0.7817 for predicting financial distress, representing a 6.22% enhancement compared to relying solely on financial indicators. More interestingly, incremental information features processed by LLMs offer greater incremental value compared to machine learning-based methods. Moreover, the topic incremental information in the inquiry letters, particularly that related to production and operations, plays a significant role in identifying financial distress. This study provides novel insights into the relationship between inquiry letters and FDP, while also broadening the application of LLMs in financial risk management. Simultaneously, the interpretability study of the SHAP model offers a more precise reference for investment decisions by information users.