Customer credit quality assessments using data mining methods for banking industries

Personal credit scoring on credit cards has been a critical issue in the banking industry. The bank with the most accurate estimation of its customer credit quality will be the most profitable. The study aims to compare quality prediction models from data mining methods, and improve traditional models by using boosting and genetic algorithms (GA). The predicting models used are instant-based classifiers (such as k-nearest neighbors), Bayesian networks, decision trees, decision tables, logistic regressions, radial basis function neural networks, and support vector machines. Three boosting (or ensemble) algorithms used for performance enhancement are AdaBoost, LogitBoost, and MultiBoost. The mentioned algorithms are optimized by GA for input features. Empirical results indicated that GA substantially improves the performance of underlying classifiers. Considering robustness and reliability, combining GA with ensemble classifiers is better than traditional models. Especially, integrating GA with LogitBoost (C4.5) is the most effective and compact model for credit quality evaluations.