Approach to corporate credit grading prediction using Bayesian Network model

Abstract

In the field of corporate credit grading, researchers have recently been interested in machine learning techniques rather than typical statistical ones in order to find better prediction performance. However, little research has been done to find a model for both describing causal relations and better prediction performance at the same time. This research intends to propose an enhanced prediction model for corporate credit grading to find the relationship among the influencing factors. For the sake of this purpose, we applied the Bayesian Network which is a brand new approach in this area. To validate our research, we analyzed 1,019 records collected by a Korean corporate credit rating agency based on the Bayesian Network, and other famous machine learning techniques such as the Decision Tree and Neural Network are also applied for benchmarking tests. In addition, we applied several algorithms of Bayesian Networks, which have been frequently adapted in prior research. The results show that the TAN algorithm of the Bayesian network and the J48 algorithm in the Decision Tree are the best prediction models, performing slightly better than other benchmarking techniques. However, the K2 algorithm could be a recommended model for its richness of interpretation capability and practical application: Firstly, it can explain relationships among factors clearly, which is critical to the stakeholders of a corporation. Secondly the performance is not much behind that of the best prediction model. Lastly, its simulation capability, such as in the what-if analysis, provides rich information for the researcher as well as industry experts. © 2011 International Information Institute.