Support Vector Machines Based Methodology for Credit Risk Analysis

Abstract

Credit risk analysis is a classical and crucial problem which has attracted great attention from both academic researchers and financial institutions. Through the accurate classification of borrowers, it enables financial institutions to develop lending strategies to obtain optimal profit and avoid potential risk. Actually, in recent decades, several different kinds of classification methods have been widely used to solve this problem. Owing to the specific attributes of the credit data, such as its small sample size and nonlinear characteristics, support vector machines (SVMs) show their advantages and have been widely used for scores of years. SVM adopts the principle of structural risk minimization (SRM), which could avoid the “dimension disaster” and has great generalization ability. In this study, we systematically review and analyze SVM based methodology in the field of credit risk analysis, which is composed of feature extraction methods, kernel function selection of SVM and hyper-parameter optimization methods, respectively. For verification purpose, two UCI credit datasets and a real-life credit dataset are used to compare the effectiveness of SVM-based methods and other frequently used classification methods. The experiment results show that the adaptive Lq SVM model with Gauss kernel and ES hyper-parameter optimization approach (ES-ALqG-SVM) outperforms all the other models listed in this study, and its average classification accuracy in the two UCI datasets could achieve 90.77% and 75.21%, respectively. Moreover, the classification accuracy of SVM-based methods is generally better or equal than other kinds of methods, such as See5, DT, MCCQP and other popular algorithms. Besides, Gauss kernel based SVM models show better classification accuracy than models with linear and polynomial kernel functions when choosing the same penalty form of the model, and the classification accuracy of Lq-based methods is generally better or equal than L1- and L2-based methods. In addition, for a certain SVM model, hyper-parameter optimization utilizing evolution strategy (ES) could effectively reduce the computing time in the premise of guaranteeing a higher accuracy, compared with the grid search (GS), particle swarm optimization (PSO) and simulated annealing (SA).