INTERNATIONAL ECONOMIC REVIEW
Vol. 36, No. 2, May 1995
SENSITIVITY ANALYSIS AND MODEL EVALUATION IN SIMULATED DYNAMIC GENERAL EQUILIBRIUM ECONOMIES*
BY FABIO CANOVA 1
This paper describes a Monte Carlo procedure to evaluate dynamic nonlinear general equilibrium macro models. The procedure makes the choice of parameters and the evaluation of the model less subjective than standard calibration techniques, it provides more general restrictions than estimation by simulation approaches and provides a way to conduct global sensitivity analysis for reasonable perturbations of the parameters. As an illustration the technique is applied to three examples involving different models and statistics.
A growing body of research in the applied macroeconomic literature uses simulation techniques to derive the time series properties of nonlinear stochastic general equilibrium models, to compare them to real world data and to evaluate policy options (see e.g. King, Plosser, and Rebelo 1988, or Cooley and Hansen 1990). In implementing numerical analyses of general equilibrium models, one has to overcome four hurdles. First, an economy must be specified and functional forms for its primitives selected. Second, a decision rule for the endogenous variables in terms of the exogenous (and predetermined) variables and of the parameters must be computed. Third, given the probability structure of the economy, values for the parameters must be chosen. Fourth, the closeness of functions of simulated and the actual data must be assessed in a metric which is relevant to the problem and policy conclusions, if any, should be drawn.
While models are often specified with an eye to analytical tractability and there has been progress in developing techniques to numerically approximate unknown decision rules for the endogenous variables (see e.g. Sims 1984, Coleman 1989, Novales 1990, Baxter 1991, Tauchen and Hussey 1991, Judd 1992, Marcet 1992 and the January 1990 issue of the Journal of Business and Economic Statistics), surprisingly little attention has been paid to the problems connected with the other two steps of the simulations. In particular, the selection of the parameters and the evaluation of the simulation results have been undertaken using procedures which
* Manuscript received October 1991; revised November 1994.
1 I would like to thank David Backus, Javier Diaz, Frank Diebold, John Geweke, Eric Ghysels, Gary Hansen, Bruce E. Hansen, Jane Marrinan, Yaw Nyarko, Adrian Pagan, Franco Peracchi, Victor Rios-Rull, Gregor Smith, Herman van Dijk, Randall Wright, two anonymous referees, and the participants of seminars at Brown University, European University Institute, New York University, University of Montreal, University of Rochester, University of Pennsylvania, University of Rome and Carlos III Madrid for comments and suggestions.