Research On The Models Used For Consumer Credit Scoring In Commercial Banks

Personal credit scoring is the application of financial risk forecasting. Recently, financial institutions have been experiencing serious competition and challenge. Due to the huge growth of the personal loans and credit industry, a wave of default in consumer credit has brought tremendous losses to the commercial banks. As a result, building an effective credit scoring model has been an even important task for banks to decrease lost, save amount cost and make efficient decision.Three research problems were found after an review of previous consumer credit scoring models. First, the newly proposed evaluation models are unable to make a remarkable progress on classification criterions, especially Type II error. Second, most of current evaluation models are based on artifical intelligence method such as neural network and support vector machine, both of which show a poor interpretability. Finally, although the new ensemble method improve the prediction accuracy, its computing cost remains an obvious disadvantage.In order to solve the above issues, in this research, a new hybrid method for credit scoring based on Decision tree and Bayesian network-DTBN, is proposed. The process of building DTBN model includes two steps. First, all of the training data is used to learn a Decision tree model. Then the wrongly-classified samples under each leaf notes of the trees are chosen to build a Bayesian network. The key idea of DTBN method is using Bayesian network to reclassify those incorrectly-classified samples under each leaf note of the Decision tree in order to improve the prediction accuracy.The experiment employed German credit data sets to build DTBN classifier.10times10-fold cross-validation was adopted to estimate the performance of the new classifier. The empirical analysis showed that the average accuracy, Type I error and Type II error of the DTBN method are87.05%,3.01%and34.24%. By comparison, we found that in terms of all three established standard measures, DTBN consistently outperform the Decision tree and Bayesian network single models as well as other methods in previous papers which used the same data set. DTBN method had greatly decreased two types of errors, especially Type II error, which was a great progress. In addition, DTBN has advantages of good interpretability and lower computing cost.The main contribution of this research includes three points. First, a new hybrid method for credit scoring based on Decision tree and Bayesian network was proposed, which successfully solved the three issues and provided a novel idea for research. Second, we employed Greedy stepwise method for feature selection, which was not frequently used in previous research. The results indicated that fewer features produced a better classification performance. Finally, using Bayesian network to reclassify wrongly-classified samples of leaf note successfully decreased Type II error. As a result, this paper offered a new method for decreasing two types of error, especially type II error.