Scholarlink

<p>Risk factor disclosures in annual reports, which discuss the potential uncertainties faced by companies in textual form, are rarely utilized in financial distress prediction. This paper innovatively employs the large language model (LLM) to extract systematic textual predictors for financial distress from massive textual risk factor disclosures. Based on risk factor disclosures from 27,059 financial reports of 4,694 U.S. publicly listed companies between 2014 and 2023, a total of 24 LLM-based textual predictors are extracted, including two new predictors not included in prior studies. Compared to commonly used textual features, LLM-based textual predictors improve financial distress prediction performance by achieving a significantly higher AUC (Area Under the Curve). Furthermore, following instructions in specially designed prompts, the LLM generates detailed explanations on how to utilize extracted textual predictors to distinguish distressed companies from non-distressed ones, making LLM-based textual predictors highly interpretable. This study highlights the critical role of risk factor disclosures in assessing financial distress risks and demonstrates the powerful textual information extraction capabilities of LLMs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Uncovering Financial Distress with Textual Risk Disclosures in Annual Reports: Insights from Large Language Models

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Xiaoqian Zhu,
  • Hanlin Jin,
  • Hao Sun,
  • Jianping Li

摘要

Risk factor disclosures in annual reports, which discuss the potential uncertainties faced by companies in textual form, are rarely utilized in financial distress prediction. This paper innovatively employs the large language model (LLM) to extract systematic textual predictors for financial distress from massive textual risk factor disclosures. Based on risk factor disclosures from 27,059 financial reports of 4,694 U.S. publicly listed companies between 2014 and 2023, a total of 24 LLM-based textual predictors are extracted, including two new predictors not included in prior studies. Compared to commonly used textual features, LLM-based textual predictors improve financial distress prediction performance by achieving a significantly higher AUC (Area Under the Curve). Furthermore, following instructions in specially designed prompts, the LLM generates detailed explanations on how to utilize extracted textual predictors to distinguish distressed companies from non-distressed ones, making LLM-based textual predictors highly interpretable. This study highlights the critical role of risk factor disclosures in assessing financial distress risks and demonstrates the powerful textual information extraction capabilities of LLMs.