关键词: 借款描述, 文本分析, 临界点, 违约判别

Abstract: Prediction of default is of great significance to financial institutions’ loan and business credit decisions. This paper aims to explore how to use the unstructured loan description to build a default prediction model thtt can help financial institutions identify default customers more efficiently. The innovative and unique points of this study are reflected in two aspects. First, the use of the pca-foword method to extract the information in the loan description can not only avoid the shortcoming of specific character frequency method through which the loan description information extracted is insufficient, but also preclude the possibility that the information extracted from the loan description is irrelevant for the identification of customer default status; second, the use of both loan description data and digital data to establish a default prediction model avoids the shortcoming of single-data-based model that is not highly accurate in default prediction. The research shows that in the comparative analysis, the default prediction model established by using the optimal critical point logistic regression model and two types of data (loan description data and numerical data) has the highest accuracy of default prediction; Regression analysis shows that there are differences in the ability of different borrowing companies to identify the borrower’s default; description of the basic situation shows a significant positive correlation with the default; description in relation to production and operation shows a significant positive correlation with the default; description of repayment commitment has a significant negative correlation with the default; description in relation to the company that the borrower works in or operates has a significant negative correlation with the default. The above relationship remains so after the borrower’s economic characteristics are controlled.

Key words: loan description, text analysis, cut-off point, default prediction