Credit Risk Estimation Model Development Process: Main Steps and Model Improvement

Ricardas Mileris, Vytautas Boguslauskas

Abstract


The attribution of credit ratings for clients is a veryimportant issue in the banking sector. Banks must evaluatecredit risk of credit applicants by using standardized(external rating institutions) or internal ratings-based(IRB) methods. Banks which decided to use IRB methodattempt to develop precise internal credit rating models forthe evaluation of creditworthiness of their borrowers.The internal rating method for the estimation of defaultprobability requires to collect the default information fromthe historical data in banks. The major studies about defaultdeterminating factors are based on classification methods(Zhou, Xie, Yuan, 2008). A classification model considersthe default measurement as the pattern recognition whereall borrowers are divided to non-default and default groupsbased on their financial and non financial data. Banksattempt to construct an evaluation model that can be used todiscriminate new sample.This research focuses on a credit rating modeldevelopment which could attribute credit ratings forLithuanian companies. The steps of a model’s developmentand improvement process are described in this paper.The model’s development begins with the selection ofinitial variables (financial ratios) characterizing defaultand non-default companies. 20 financial ratios of 5 yearswere calculated according to annual financial reports.Then statistical and artificial intelligence methods wereselected for the classification of companies into twogroups: default and non-default. A discriminant analysis,logistic regression and artificial neural networks (multilayerperceptron) were applied for this purpose. Often statisticalmethods are not able to operate with a large amount of data,so the analysis of variance, Kolmogorov-Smirnov test andfactor analysis were applied for data reduction. Artificialneural networks often are able to analyze a large amount ofdata so variable selection was accomplished by the networkitself calculating ranks of importance for every initialvariable. There were constructed 15 classification modelsand their classification accuracy was measured bycalculating correct classification rates. The most accuratewas a logistic regression model analyzing data of 3 years(97% of correctly classified companies). Then the sample ofcompanies was supplemented with new data and changes inclassification accuracy were estimated. The significantdecrease of classification accuracy conditioned the need ofmodel update. For this reason the logistic regressioncoefficients were recalculated. In order to classify nondefaultcompanies into 7 classes: profitability, liquidity,financial structure and individual possibility of defaultestimated by a logistic regression model were determinedas rating criterions. Then the rating scale was constructedand credit ratings were attributed for companies in thesample. The calculated probabilities of default indicatedthat some lower ratings have lower probabilities of default.These imperfections were corrected by the modification ofa rating scale. The research has shown that the developedmodel is a valid tool for the estimation of credit risk.


Keywords


bank, credit ratings, credit risk, classification methods.

Full Text: PDF

Print ISSN: 1392-2785
Online ISSN: 2029-5839