关键词:

文本挖掘, 管理层讨论与分析, 违约预测, 代价敏感GBDT, 信息不对称

Abstract:

This paper, by employing text mining techniques, analyzes the text of management discussion and analysis (MD&A) content in annual reports of listed companies and constructs an MD&A evaluation system from the perspectives of text similarity, text readability, text tone, and management expectations. By constructing a cost-sensitive gradient boosting decision tree (csGBDT) model, it examines the impact of multidimensional MD&A indicators on corporate default prediction and further analyzes the MD&A indicators that have a significant impact on corporate default status and their marginal effects on the role of default status. It is found that MD&A indicators can be used as an alternative data source to accurately predict the default status of listed companies. MD&A indicators are less effective predictors compared to traditional default prediction variables. MD&A indicators provide additional information content on top of traditional default discriminators. The csGBDT model significantly improves the discriminatory ability of firms (especially for defaulted firms) in the large scale of default prediction data methods, which has obvious advantages. Among the many MD&A indicators that have a significant impact on corporate default are, in order, text similarity compared to the previous year, total vocabulary, sentiment tone 2, total vocabulary/number of sentences, sentiment tone 1 and whether management has issued a performance forecast. This paper extends the research boundary of corporate default prediction from structured data to unstructured textual data, which helps to curb the risk of corporate default due to information asymmetry.

Key words:

 , text mining, management discussion and analysis, default prediction, cost-sensitive gradient boosting decision tree （GBDT）, information asymmetry