Evaluating DSGE models for monetary and fiscal policy analysis.

摘要

This dissertation evaluates dynamic stochastic general equilibrium (DSGE) models that are widely used for policy analysis at central banks and other policy institutions. DSGE models, like all economic models, abstract in many ways from reality and are misspecified along certain known and unknown dimensions. This dissertation adapts the posterior predictive analysis, well-known in the statistics literature, to highlight the strengths and weaknesses of DSGE models as they relate to the intended task of policy analysis.;The first chapter provides a motivation and introduction to the thesis. In the second chapter we adapt the tools of prior and posterior predictive analysis to the DSGE context and illustrate their usefulness in highlighting the discrepancies between the model and the realized sample. We argue that standard criticisms of prior and posterior predictive analysis, whatever their merits in other contexts, miss the point in the DSGE context. We illustrate that posterior predictive analysis, in particular, can be useful for DSGE model evaluation and it can be viewed as a natural pragmatic Bayesian response to a murky modelling problem. In the third chapter, we apply this framework to evaluate a DSGE model for the task of monetary policy analysis. We argue that policymaking at central banks can be characterized as interpreting the structural sources of unexpected outcomes in the observed data and accordingly acting upon it. In the DSGE context this amounts to checking whether the model implied structure (first and second moments) of the one-step ahead forecast errors is consistent with the structure observed on the realized sample. We show that in practice, in order to reconcile the U.S. marco dataset with the iconic Smets-Wouters model, we need that the observed sample must involve a highly unlikely sample correlation of structural shocks that are assumed to be uncorrelated in the model.;The fourth chapter is an empirical exercise to shed light on fiscal policy effectiveness at the zero bound for interest rates. We estimate the fiscal policy multipliers for taxes and spending in Japan both before and after the economy hit the zero lower bound in the mid nineties using a tax-code based structural VAR identification methodology. The exercise is an attempt to see if there is enough information in the data to resolve whether the fiscal policy operates differently at the zero lower bound and finds some useful results, but mainly concludes that the data are not sufficiently informative to resolve the issue using these methods.;Keywords: Bayesian Analysis; DSGE Model Evaluation; Forecast Errors; Monetary Policy; Fiscal Policy; Zero Lower Bound; Structural VAR