違約預測模型中變數選擇方法之應用

Abstract

For discovering the useful variables in predicting the occurrence of firm’s default event, we examine 8 variable selection methods, including best subset selection, stepwise regression, stagewise regression, ridge regression (RR), the lasso, least angle regressions (LAR), principle component regression (PCR), and partial least squares regression (PLS). According to our simulation results, the stagewise regression, the lasso, and LAR, among all the variable selection methods, have stable in-sample fitting ability and robust performance in terms of root mean square error (RMSE). As a representative of the other 5 methods, PCR has similar result as the best 3 methods in our empirical analysis. Nevertheless, this thesis recommends the best 3 methods since the computation is less time-consuming and the results are more intuitive to interpret. Moreover, we find the combination of selected variables is time-dependent. Therefore, the incorporation of the frailty factor is inevitable to construct the default prediction model in the future studies.