Financial Risk Forecasting Using Naїve and Tree Augmented Classifier Based on Bayesian Networks

Abstract

Background . Development and study of characteristics for naive and tree-augmented classifiers in the form of Bayesian networks in the problem of credit risk estimation. Objective . To perform estimation of classification quality for the bank credit borrowers using Bayesian classifiers of two types. Methods . Development of necessary mathematical tools and performing computational experiments aiming towards constructing classifiers in the form of Bayesian networks using actual statistical data characterizing solvency of bank credit borrowers. Results . The following results were achieved: the methodology of constructing and application of the naive and tree-augmented Bayesian classifiers for solving the problem of solvency estimation for bank credit borrowers; an analysis of computational algorithmic complexity was performed; two classification models were constructed in the form of Bayesian networks using actual statistical data from banking system; a comparative analysis was performed for the models developed. Conclusions . It was established that the tree-augmented classifier exhibits higher computational complexity than the naive Bayesian one, but it showed higher classification results while solving the problem of bank clients classification into two groups: those who return the credits and those who don’t.