Classifier performances for credit risk analysis

Abstract

Credit Risk Analysis (CRA) is an important and challenging data mining problem in financial analysis domain which is commonly used by many financial organizations. It became one of the most important and hot concept in finance sector since the real market’s credit volume has significant growths while economies have fluctuations which has great impacts on financial organizations. CRA aims to decrease future losses by estimating the potential risk and eliminating the new credit proposal if the risk is higher than a defined tolerance value. In this thesis study, it is aimed to compose a combined classification model to have any little improvement of classification performance when it is compared with existing classifiers. This comparison is based on performance metrics such as accuracy, Receiver Operating Characteristics (ROC) and precision. Any little improvement in accuracy and optimality which seems insignificant, will reduce losses in a large loan portfolio and save very significant amounts which can be defined in terms of billions of dollars. Three different datasets are used in this study. Those are German Dataset, a national bank dataset and a synthetic dataset from a data mining tool. WEKA and GeneXproTools softwares are used to make experiments. Single classification techniques Support Vector Machine (SVM) is applied to German Dataset with different kernel functions. Logistic Regression (LR) is applied on synthetic data with different crossvalidation. Finally, LR, SVM, Neural Networks, Naive Bayes and Dynamic Bayesian approaches are compared each other on real-life bank dataset. A hybrid approach proposed which combines best-performed single classifiers inside as a unique classifier. Results show that combined classification impact performance in terms of improvement among other single classification techniques.