Cross-Validation for Prediction

Abstract

The authors review and critically evaluate three major approaches to cross-validation in the context of predictive validity: (1) sample-splitting, (2) resampling plans such as bootstrapping, and (3) a method that simultaneously estimates parameters and cross-validates. Because of the information loss involved and the availability of shrinkage formulas, they argue that data-splitting is inferior to the simultaneous approach. An empirical example illustrates how the simultaneous approach can be used in conjunction with bootstrap resampling to construct prediction intervals that are superior to classical prediction intervals.