Feature Selection in a Credit Scoring Model

Abstract

This paper proposes different classification algorithms—logistic regression, support vector machine, K-nearest neighbors, and random forest—in order to identify which candidates are likely to default for a credit scoring model. Three different feature selection methods are used in order to mitigate the overfitting in the curse of dimensionality of these classification algorithms: one filter method (Chi-squared test and correlation coefficients) and two wrapper methods (forward stepwise selection and backward stepwise selection). The performances of these three methods are discussed using two measures, the mean absolute error and the number of selected features. The methodology is applied for a valuable database of Taiwan. The results suggest that forward stepwise selection yields superior performance in each one of the classification algorithms used. The conclusions obtained are related to those in the literature, and their managerial implications are analyzed.