This paper uses a supervised machine learning algorithm to extract relevant (soft) information from annual reports and examines whether such information determines credit risk (as measured by non-performing loans, Ohlson’s O-score, Altman’s Z-score, and credit rating downgrades). The paper also assesses how far both bank- and country-level characteristics influence variations in credit risks both within and between banks across 19 European countries between 2005 and 2017. Based on 1885 firm-year observations, we find that the text-based credit risk (soft) measure explains a substantial portion of the variation in NPLs, O-score, Z-score, and credit rating downgrades. We also find that bank-level characteristics and country-level characteristics are highly important for explaining variations in non-performing loans, O-score, and credit rating downgrades, as compared to Z-score. Overall, our results have implications for firms, regulators, and market participants who are seeking evidence on the credibility of annual reports in conveying relevant information that reflects actual credit risk.

本文使用监督式机器学习算法从年度报告中提取相关（软）信息，并检查此类信息是否确定信用风险（以不良贷款，Ohlson的O评分，Altman的Z评分和信用评级降级来衡量） 。本文还评估了银行层面和国家层面的特征在多大程度上影响了2005年至2017年间19个欧洲国家内部和内部银行之间信用风险的变化。基于1885年的公司年观察，我们发现基于文本的信用风险（软性）衡量标准可以解释不良贷款，O评分，Z评分和信用评级下调的很大一部分变化。我们还发现，银行一级的特征和国家一级的特征对于解释不良贷款，O分数，与Z评分相比，信用评级被下调。总体而言，我们的结果对寻求证据证明年度报告的可信度的公司，监管机构和市场参与者具有影响，这些信息传达了反映实际信用风险的相关信息。