DEVELOPMENT OF CREDIT RISK PREDICTION MODEL USING SUPPORT VECTOR MACHINE TECHNIQUE

Abstract

Financial institutions are always saddled with the responsibility of measuring credit worthiness of their clients in order to manage the risk of default for loans advanced to client. Obtaining credit rating involves special credit rating agencies investing large amount of resources and time to manually carry out in depth risk assessment of clients’ risk. An accurate, automated credit rating prediction tool can prevent liquidation of loans advanced to customers resulting in increased profitability to the bank. Research showed that that inferior credit risk assessment tool is the primary reason of enterprise bankruptcy and that heuristic tools have better accuracy than the traditional statistical tools in predicting credit scoring. This study aims to determine support vector machine kernel type and the kernel parameters to achieve optimal credit risk prediction accuracy on different datasets – the German and Taiwan datasets. The datasets were sampled with stratified sampling technique, optimal features were selected from the dataset by ranking their Pearson correlation coefficient of the feature relative to the target label. After the selection, SVM classifier was modelled to predict the credit risk on the datasets. Model generated for different kernel/kernel parameters were analysed on sensitivity, specificity and accuracy of models. The performance of the SVM classification is greatly influenced by the tuning of the two parameters kernel, kernel coefficient (k) and the regularization constant (C). Radial basis function (RBF) kernel yielded the best total accuracy. Optimal classification performance of 76% and 71% were recorded for radial basis function (RBF) kernel at k = 65, C = 205 and k = 60, C= 160 for the German credit dataset 70:30 and 60:40 distributions respectively while 80.0% and 61.0% were recorded at k = 10 and C = 200 for Taiwan credit data on both distributions. SVM model performs best for the 70:30 data samples