A1317
Title: Model-average impulse responses for nonlinear VARMAs
Authors: Miguel Faria-e-Castro - Federal Reserve Bank of St. Louis (United States)
Neville Francis - University of North Carolina Chapel Hill (United States)
Michael Owyang - Federal Reserve Bank of St Louis (United States)
Daniel Soques - University of North Carolina Wilmington (United States) [presenting]
Abstract: In linear settings, local projections (LPs) can be more robust to misspecification than VAR-based methods when the true data-generating process has a VARMA representation. However, regime-dependent LPs can be biased in nonlinear settings, particularly when shocks affect the future regime. Monte Carlo experiments based on Threshold VARs that vary along two dimensions are used: (i) the degree of moving-average behavior in the data generating process and (ii) the strength of nonlinear regime dependence. These simulations allow comparison of the relative performance of generalized impulse response functions (GIRFs) and regime-dependent LPs across a range of misspecifications. A model-average impulse response that combines GIRFs and LPs using weights based on relative mean squared error is proposed, thereby allowing the estimator to adjust to the underlying source of error rather than relying on a single method. The usefulness of this approach is then assessed in an application based on a regime-dependent DSGE model, showing how model averaging can improve impulse response estimation when both nonlinearities and misspecification arising from moving-average components are present.
