Credit risk analysis using Hidden Markov model

Abstract

The purpose of this study is to investigate the performance of Hidden Markov Model (HMM) for credit risk analysis in terms of classification and probability of default (PD) modeling where PD modeling assigns default bankruptcy probabilities to credit customers instead of strictly classifying them as good (solvent) borrowers and bad (insolvent) borrowers. PD modeling process makes use of some data belonging to credit applicants and helps banks or credit companies compute a probability that the customer should not pay his/her debts in a timely manner, prior to the decision of granting credit. In the first part of this study, classification ability of HMM is compared to that of Logistic Regression (LR) and k-Nearest Neighbors (k-NN) which are two conventional methods for classification. In the second part, PD modeling performance of HMM is analyzed and compared to that of LR which is known to be one of the most popular PD modeling methods so far. Australian Credit Database and German Credit Database are two public datasets utilized in this thesis study. Classification performances of the aforementioned methods are judged according to accuracy, error cost and Receiver Operating Characteristics (ROC) analysis with supporting experiments using six-fold cross validation. PD modeling performances of HMM and LR are also compared by directly examining the average PD values for solvent and insolvent borrowers. Matlab’s HMM toolbox by Kevin Murphy is used for HMM computations whereas a web based tool is utilized for LR analysis. Matlab is also used for k-NN analysis in classification experiments. The aim of this study is to build appropriate algorithms for HMM to make it an effective way of credit risk analysis as well as conventional methods. The results of the experiments show that HMM is a powerful and effective method for credit risk analysis and can be utilized by financial institutions.