Cross‐estimation for decision selection

Abstract

We propose a data-driven procedure, cross-estimation for decision selection (CrEDS), to choose from an abundance of off-the-shelf statistical models or computer algorithms at a decision-maker's disposal. CrEDS combines the ideas of cross-validation (CV) and local smoothing, a nonparametric statistical technique. We demonstrate the power of CrEDS with five numerical experiments in inventory and revenue management problems, ranging from low to high dimensional and from exogenous to endogenous. We also conduct a case study using an auto-lending data. CrEDS performs favorably compared to other existing selection criteria and provides a practical framework for a broad range of optimal decision selection problems.