Predictive modeling of corporate credit ratings using a semi-supervised random forest regression

摘要

A company's ability to fulfill its obligations is evaluated as its credit rating by rating agencies. In recent years, learning models including a neural network and support vector machine have been applied to financial data, to predict company credit ratings with a high degree of accuracy. However, the number of companies that have yet to be rated is overwhelmingly greater than the number of companies that have been rated. To utilize effectively the information on unrated companies, we propose the use of a semi-supervised learning model for their rating prediction. We adopt the use of a random forest, which is a powerful tool in terms of its identification accuracy and generalization capability when applied as a credit rating prediction model. To confirm the effectiveness of the proposed method, through this research, the financial data and rating information for all listed companies in Japan were used to evaluate the prediction accuracy. The prediction residuals of both a normal random forest (RF) and the proposed method were evaluated experimentally based on a cross-validation.