Estimation of Credit Risk by Artificial Neural Networks Models

Abstract

Credits mostly form a considerable part of banks assets and is one of the most risky types of them. Credits for banks are not only the source of income but also they can be the main reason of loss. The main risk that banks meet with lending money is credit risk. It is risk that debtor will not be able to repay his obligations due to certain reasons. Seeking to reduce potential loss due to crediting not reliable clients banks must be able to measure credit risk of each client properly. One of possible methods is using of internal credit risk estimation models.Due to the importance of credit risk analysis, many methods were widely applied to credit risk measurement tasks: linear discriminant analysis, logit analysis, probit analysis, linear programming, integer programming, k-nearest neighbour, classification tree, artificial neural networks (ANN), genetic algorithm, support vector machine, some hybrid models ant other. An increasing field of research in artificial neural networks is the one mainly concerned with interactions between economics and computer science, studying their potential applications to economics. Artificial neural networks represent an easily customizable tool for modelling learning behaviour of agents and for studying a lot of problems very difficult to analyze with standard economic models (Gallo, 2006).ANN have many advantages over conventional methods of analysis. According to Shachmurove (2002), they have the ability to analyze complex patterns quickly and with a high degree of accuracy. Artificial neural networks make no assumptions about the nature of the distribution of the data. Since time-series data are dynamic in nature, it is necessary to have non-linear tools in order to discern relationships among time-series. ANN are best at discovering these types of relationships. Also neural networks perform well with missing or incomplete data. Compared with an econometric model, it is easier to use ANN where a forecast needs to be obtained in a shorter period of time.The purpose of this research is to define the rates of classification accuracy and to measure the classification accuracy of artificial neural networks credit risk estimation models. The methods of the research are analysis of scientific publications about credit risk estimation models and analysis of artificial neural networks credit risk estimation models classification accuracy.One important component needed to accomplish credit risk evaluation is to seek an accurate classifier in order to categorize new applicants or existing customers as good or bad (Lai, Yu, Wang, Zhou, 2006). In this paper rates of credit risk estimation models accuracy and their calculation were described: correct classification and misclassification rates, false negative and false positive rates, model sensitivity and specificity, positive and negative predictive values, F-measure, ROC analysis. Analysis of scientific publications about credit risk estimation models has shown that the most efficient of the most commonly used methods are logistic regression and artificial neural networks. Less reliable methods are decision trees and discriminant analysis. Also artificial neural networks models (multilayer perceptrons) were constructed for the analysis of Lithuanian enterprises credit risk. Calculation of models accuracy rates has shown that the most efficient model analyses data about clients is of 3 years. Experiments have shown that analyzing data of 4 and 5 years classification accuracy decreases due to high quantity of input information and neural network‘s overlearning. Models accuracy rates allowed to estimate risk of client misclassification and other characteristics. Also they helped to make the decision which model is to be used in practice in order to measure credit risk successfully.