A0172
Title: Bayesian inference for nonlinear and non-Gaussian state-space models with global-local shrinkage process priors
Authors: Mattias Villani - Stockholm University (Sweden) [presenting]
Abstract: The recently proposed global-local shrinkage process priors for time-varying parameter models allow parameters to be essentially constant for longer spells, followed by periods of rapid change or jumps. Most of this literature uses linear and conditionally Gaussian models. The standard posterior samplers based on particle methods, typically used for nonlinear and non-Gaussian state space models, struggle with the parameter constancy implied by the global-local shrinkage processes. Several alternative algorithms are presented, and their performance is illustrated on some commonly used time-varying parameter models in statistics and econometrics.
