On Lasso and adaptive Lasso for non-random sample in credit scoring

Prediction models in credit scoring are often formulated using available data on accepted applicants at the loan application stage. The use of this data to estimate probability of default (PD) may lead to bias due to non-random selection from the population of applicants. That is, the PD in the general population of applicants may not be the same with the PD in the subpopulation of the accepted applicants. A prominent model for the reduction of bias in this framework is the sample selection model, but there is no consensus on its utility yet. It is unclear if the bias-variance trade- off of regularization techniques can improve the predictions of PD in non-random sample selection setting. To address this, we propose the use of Lasso and adaptive Lasso for variable selection and optimal predictive accuracy. By appealing to the least square approximation of the likelihood function of sample selection model, we optimize the resulting function subject to L1 and adaptively weighted L1 penalties using an efficient algorithm. We evaluate the performance of the proposed approach and competing alternatives in a simulation study and applied it to the well-known American Express credit card dataset