Models, Computational Experiments and Reality

DSGE models are designed to mimic only certain aspects of reality, usually speci…ed moments of observable data. They typically have other implications that are clearly false and lead to their immediate rejection if taken literally. Widely used calibration exercises compare the implications of DSGE models for the distribution of speci…ed sample moments with the corresponding data. This paper shows that this procedure takes DSGE models literally, and therefore retains the implications that lead to their immediate rejection. If, instead, the DSGE model is interpreted only to imply particular population moments, and not the distributions of the corresponding sample moments, this logical dif…culty does not emerge but the model then has no falsi…able implications. The constructive contribution of the paper is to merge the DSGE model with an atheoretical econometric model in a logically consistent way that has refutable implications for observable data. This leads to practical procedures that compare the prior distribution of the DSGE model and the posterior distribution of the atheoretical model for the population moments the DSGE model is intended to describe. The concepts are illustrated using four competing DSGE models of the risk-free rate and the equity premium. The synthesis advanced in the paper resolves the equity premium puzzle in this context. The U.S. National Science Foundation provided partial …nancial support through grants SBR9819444 and SBR-0720547. Comments from seminar participants at CEPREMAP, Johns Hopkins University, New York University, Stanford University, University of Iowa and the Federal Reserve Bank of Atlanta were helpful in writing and revising the paper. Responsibility for errors rests with the author.