A1161
Title: Misspecification, factors, and group sparsity for macro-at-risk models
Authors: Matteo Mogliani - Banque de France (France) [presenting]
Anna Simoni - CREST - ENSAE and CNRS (France)
Abstract: The aim is to consider a framework where the macroeconomist disposes of a large set of variables for estimating macro-at-risk models. In practice, one can consider either dense or (approximate) sparse models with group structure. These two approaches are analyzed in different scenarios that account for: (i) diverse degrees of sparsity, (ii) signal-to-noise ratios in the probabilistic mechanism generating the covariates, and (iii) alternative correlation structures between groups. A Bayesian quantile regression model is used based on the asymmetric Laplace (AL) distribution and different shrinkage priors. This model can be misspecified in two dimensions: The AL likelihood function and the quantile regression model. It is shown that the first type of misspecification does not impact, in general, the asymptotic results. On the other hand, the second type of misspecification is more serious, and the asymptotic performance depends on the degree of sparsity assumed and the prior adopted. It is shown that there are prior distributions that are more robust than others to misspecification of the macro-at-risk model. Monte Carlo simulations and real data exercises provide further illustration of the results.
