A modular approach for SMEs credit risk analysis

Abstract

Credit risk analysis is a challenging problem in financial analysis domain. It aims to estimate the risk occurred when a customer is granted. The risk estimation depends on both customer behavior and economical condition. The challenge is how the credit expert will determine which information should be collected from applicants, under which condition a customer will be classified as good and how much risk will be taken if the credit is granted to the customer. Consequently, credit experts need intelligent customer-specific risk analysis modules to support them when they make these decisions. In this thesis, we present a cascaded multilayer perceptron (MLP) rule extractor and a logistic regression (LR) model a for real-life Small and Medium Enterprises (SMEs). In the preprocessing phase, the features of Turkish SME database are selected by decision tree (DT), recursive feature extraction (RFE), factor analysis (FA) and principal component analysis (PCA) methods. The best feature set is obtained by RFE. In the first module, the classifier is selected among MLP, k-nearest neighbor (KNN) and support vector machine (SVM). The optimal classifier is obtained as MLP and the following modules are built on MLP. For classification purpose, MLP is followed by neural rule extractor (NRE) in the second module. NRE reveals how the decision is made for customers as being “good”. For the probability of default estimation (PD), we propose a cascaded MLP which is followed by a LR model in the third module. MLP-LR model is followed by clustering method in the last module for scorecard development purpose. In experiments, confidential Turkish SME database is used. The cascaded MLP-LR model provides high accuracy rate and outperforms commonly used classical LR.