Finite-Sample Identification-Robust Inference for Nonlinear DSGE Models

We develop identification-robust likelihood-free simultaneous confidence sets for dynamic stochastic general equilibrium models, without relying on linear approximations. Our methodology integrates simulation estimation methods using auxiliary statistics with Monte Carlo test principles. Results cover deep parameters and impulse responses. Auxiliary statistics include coefficients of linear and nonlinear vector autoregressions and local projections. Proposed procedures are illustrated through laboratory experiments and an empirical application on a nonlinear real business cycle model. In simulations, we study size, power, and the trade-off between robustness and insensitivity to misspecification. Empirically, results underscore the information content of asymmetric shocks and the identification gains on impulse responses.