The Research into the Application of Artificial Neural Networks Method in the Analysis of Borrowers' Creditworthiness on Small Samples

Currently the importance of enhancing the efficiency of credit risk management in commercial banks is increasing substantially. The research conducted in this paper concerns potential opportunities of applying credit models built on the basis of artificial neural networks (ANN) to the modelling and predicting private borrowers’ creditworthiness. Neural networks are powerful forecasting instruments that have non-linear structure which fits the structure of the problem of credit risk modelling, unlike binary choice models – logit and probit – traditionally used in credit scoring. Due to the lack of interpretability of coefficients in neural-network models, it is necessary to construct linear models on the basis of neural networks with the use of so called weight-interpretation techniques for connection weights. The aim of this research is to conduct an analysis of predictive abilities of the aforementioned models and to compare their ability to classify credit risks on small samples with that of logit and probit models; it may appear relevant in the case of very limited suitable data. The empirical analysis was conducted on 5 small and 5 large samples of data on loans which had been provided to private borrowers through the online P2P-lending platform «Lending Club» from 2007 to 2011. According to the results obtained, the use of interpretation techniques may lead to increasing the power of original neural networks to classify credit risks; linear models built on the basis of ANNs appear to be considerably more accurate than binary choice models traditional for credit scoring on small samples; when it comes to predicting «bad» credit risks, models built with the use of interpretation techniques show much higher levels of classificatory precision than logit- and probit-models on all samples. Consequently, credit models that are constructed on the basis on neural networks can be a worthy alternative to the models traditionally applied in credit scoring on small samples.