Abstract:
Credit risk estimation is a key determinant for the success of financial institutions. The aim of this paper is presenting a new hybrid model for estimating the probability of default of corporate customers in a commercial bank. This hybrid model is developed as a combination of Logit model and Neural Network to benefit from the advantages of both linear and non-linear models. For model verification, this study uses an experimental dataset collected from the companies listed in Tehran Stock Exchange for the period of 2008-2014. The estimation sample included 175 companies, 50 of which were considered for model testing. Stepwise and Swapwise least square methods were used for variable selection. Experimental results demonstrate that the proposed hybrid model for credit rating classification outperform the Logit model and Neural Network. Considering the available literature review, the significant variables were gross profit to sale, retained earnings to total asset, fixed asset to total asset and interest to total debt, gross profit to asset, operational profit to sale, and EBIT to sale.
Machine summary:
Due to the significance of credit risk, a number of studies have proposed embracing statistical modeling in banks to improve their risk assessment models and hence increase the prediction accuracy of existing models (Akkoc, 2012; Al-Kassar & Soileau, 2014; Jones & Hensher, 2004; Permachandra, Bhabra & Sueyoshi, 2009; Yalsin, Bayrakdaroglu & Kahraman, 2012; Vuran, 2009; McKee & Lesenberg, 2002).
Artificial Neural Networks, genetic algorithms, genetic programming, support vector machines, and some hybrid models have been used to evaluate credit risk with promising results in terms of performance accuracy.
Using data mining techniques in application evaluation would improve credit decision effectiveness and control loan officer tasks, as well as save analysis time and cost (Bekhet & Eletter, 2014).
Also, numerous models have been proposed to solve credit rating problems but they have the following drawbacks: (1) lack of explanatory power; (2) reliance on the restrictive assumptions of statistical techniques; and (3) numerous variables, which result in multiple dimensions and complex data (Chen & Cheng, 2013).
Credit scoring is a group of decision making models and their underlying techniques which provide support for lenders while providing credit for customers (Heiat, 2012; Thomas, Edelman & Crook, 2002).
Normally, a credit scoring model is built using statistical techniques such as linear discriminant analysis (LDA) and logistic regression (LR) or artificial intelligence (AI) techniques such as support vector machines (SVMs) and Neural Networks (NN).
Table 1 categorizes some results of previous studies in credit risk, which have used Logit model or Neural Network.