On the Use of Holdout Samples for Model Selection

Researchers often hold out data from the estimation of econometric models to use for external validation. However, the use of holdout samples is suboptimal from a Bayesian perspective, which prescribes using the entire sample to form posterior model weights. This paper examines a possible rationale for the use of holdout samples: data-inspired modifications of structural models are likely to lead to an exaggeration of model fit. The use of holdout samples can, in principle, set an incentive for the modeler not to exaggerate model fit.