Model for Bankruptcy Prediction: Naive Bayesian Networks based on MDL Principle

K.W. Deng, S.S. Xia, and H.W. Zhang (PR China)


Naive Bayesian Networks; Bankruptcy Prediction; Mini mum Description Length; Risk Assessment Approaches and Methodologies


To model bankruptcy prediction of listed companies, we first propose a novel machine learning method of building naïve Bayesian networks (NBNs) based on minimum description-length (MDL) principle, and then use the financial ratio data of a-share listed companies from 2002 to 2009 in China to test its feasibility and validation. The results suggest that MDL-NBNs models have several advantages as over some other methods. First, minimum entropy principle (MEP) together with the MDL principle can complete the task of variables selection in a two-stage way. Also, the MDL metric as a fitness criterion balances model simplicity and accuracy, and is effective in minimizing overfitting, which provides a proper guidance in model selection. Second, MDL-NBNs attain a higher level of predictive accuracy with fewer predictor variables than the original NBNs, and have more robust performance at the same time, which results to obtain more interpretable models. Finally, our findings lend strong support to MDL-NBNs as a robust tool for all decision-making contexts other than bankruptcy prediction only.

Important Links:

Go Back