Title: Identification of SVAR models via independent component analysis: A comparative study
Authors: Alessio Moneta - Scuola Superiore Sant'Anna (Italy) [presenting]
Gianluca Pallante - Scuola Superiore Sant Anna (Italy)
Abstract: Independent Component Analysis (ICA) is a statistical method that transforms a set of random variables in least dependent linear combinations. Under the assumption that the observed data are mixtures of non-Gaussian and independent processes, ICA is able to recover the underlying components that generated the data, up to a scale and order indeterminacy. Its application to structural vector autoregressive (SVAR) models is straightforward because it allows to recover the impact of independent structural shocks on the observed series from the estimated residuals. We compare the performances of three different ICA techniques: fastICA algorithm, minimization of Cramer-von-Mises (CvM) distance, and minimization of distance covariance (DCov). We investigate through Monte Carlo experiments the ability of these procedures of recovering structural impulse response functions from a VAR model. Using a $p$-generalized normal distribution, we let the underlying processes approach or diverge from a Gaussian distribution. Our results suggest a relatively better performance of DCov, on average, when normality is approached. Despite the relatively larger bias of the estimates, fastICA algorithm is relatively more robust to distributional assumptions. We also present an empirical illustration using Japanese data to study the real effects of unconventional monetary policies.