Enhancing Prediction Accuracy of Default of Credit Using Ensemble Techniques

Abstract

Credit rating of an institution or individual provides a suggestive financial picture and strength of the individual or the institution. It gives the lender the ability to visualize the potentiality to the extent to which credit could be availed by the institution or the individual. Default prediction on the sum of all attributes such as payment history is a common instrument used for generation of credit rating. This research is aimed at comparing the predictive accuracy of ensemble of base classifiers using techniques of bagging, boosting, and random forest in the prediction of default of credit card clients and suggesting the technique with the highest accuracy. Customers’ default payment in Taiwan dataset is used to build the model. ML classification algorithms such as K-nearest neighbor, Naive Bayesian, decision tree, and support vector machines are applied to create the base model on the dataset. Bagging, boosting, and random forest are applied on the dataset to generate model for prediction. The accuracy of each of the models for various degrees is tabulated. Information gain feature filter method is used to identify features with maximum entropy. The features with high entropy suggested by information gain together with ensemble techniques are used to build the new model. The accuracy of the new model is then tabulated. Boosting ensemble technique is found to have the best accuracy of prediction.