Linking High-Risk Companies to Special Treatment (ST) Status: A Large Language Model Approach for Corporate Risk Evaluation

Traditional corporate risk analysis has primarily focused on quantitative financial data, often overlooking qualitative factors such as sustainability and corporate governance. This study leverages large language models (LLMs) to assess risk in three non-quantitative sections of A-share companies' annual reports: Management Discussion and Analysis, Corporate Governance, and Environmental and Social Responsibility. We analyzed annual reports from 200 companies using the ChatGLM-4 model, which was pre-trained on a multilingual dataset and optimized for long-context financial reports. The model's performance was enhanced through techniques like Rotary Positional Encoding and Group Query Attention, allowing it to accurately assess qualitative data and identify high-risk firms. By analyzing these sections, the GLM-4 model evaluates the overall risk profile of each company. The findings demonstrate a correlation between companies identified as high-risk by the LLM and those classified as Special Treatment (ST), indicating the model's potential in providing risk assessment from non-quantitative sentences.