Abstract: The textual risk disclosures in corporate financial reports, which can more directly and proactively describe corporate potentially important risk factors than other types of texts, were seldom considered by most existing studies on financial distress prediction. This paper innovatively introduces the textual risk disclosures for financial distress prediction, constructs a novel textual feature which can reflect the likelihood that the risks would impact the corporate—risk likelihood, and then builds prediction models based on machine learning methods. Based on the 35706 firm-year observations in the United States from 2006 to 2020, the empirical study demonstrates that, based on the common quantitative indicators (financial and market variables), incorporating the textual risk disclosures can significantly improve the performance of financial distress prediction. Compared with commonly used textual features, the risk likelihood feature proposed has the highest importance in financial distress prediction. The financial distress prediction ability of quantitative indicators decreases as the time horizon of financial distress prediction increases, while the predictive ability of textual risk disclosures has not declined, but has, instead, also shown a more significant improvement. This paper can help market investors and regulators understand the interpretation of textual risk disclosures in financial reports, and provides a theoretical guidance for integrating textual disclosure into financial distress prediction.

Key words: financial distress, financial reports, textual risk disclosure, text mining, machine learning

摘要： 现有的财务困境预测研究大多忽略了公司在年度财务报告中以文本形式披露的风险信息，而相比于公司披露的其他类型的文本，这些风险信息能够更加直接和前瞻地反映公司经营中的潜在重要风险。创新性地引入财务报告中的文本风险信息进行财务困境预测，并构建了能够反映所披露的风险对公司产生影响的可能性的文本特征指标——风险可能性，而后采用机器学习方法构建预测模型。基于2006~2020年美国35706个上市公司年度样本的实证研究发现：在常用的财务及市场各类定量指标的基础上，融合财务报告中的文本风险信息能够显著提升公司财务困境的预测效果；相比于其他常用文本特征指标，本研究提出的风险可能性指标在财务困境预测中表现的重要度最高；定量指标的财务困境预测能力随着预测时间窗口的提前而明显下降，而文本风险信息的预测能力不仅没有下降，还呈现出了更为显著的提升效果。本研究可以帮助市场投资者、监管机构理解如何解读公司在财务报告中披露的文本风险信息，并为实现财务困境预测中融合文本信息提供了理论指导。

关键词: 财务困境, 财务报告, 文本风险信息, 文本挖掘, 机器学习