MAKING MACRO MODELS BEHAVE REASONABLY

ABSTRACT. Using the idea of generalized dummy observations, we extend the methods of Del Negro and Schorfheide, who have proposed a way to use a dy-namic stochastic general equilibrium (DSGE) model to generate a prior distribu-tion for a structural vector autoregression (SVAR). The method proposed here is more explicit and systematic about the prior’s assertions about the SVAR identifi-cation, and it provides a mechanism for varying the tightness of the prior across frequencies, so that for example the long run properties of the DSGE can be as-serted more confidently than its short-run behavior. In every large scale macro modeling project we make compromises. Models whose properties we understand and can interpret behaviorally are generally not rich enough, either in number of variables or in dynamics, to fit the data well. Models that are rich enough to fit well become complex and can end up hav-ing implications that we believe are implausible. Practical macro modelers who face real-time demands for forecasts and policy projections have struggled contin-uously with these tradeoffs