A0464
Title: Variational Bayes for dynamic sparsity in time varying parameter regression with many predictors
Authors: Nicolas Bianco - University of Padova (Italy) [presenting]
Mauro Bernardi - University of Padova (Italy)
Abstract: Time-varying parameter models are powerful statistical tools for the analysis of dynamical systems. However, in high-dimensional problems the risk of over-parametrization is high, thus dynamic sparsity is desired. The latter is defined in two directions: vertical, where we look at the parameter vector at a fixed time, and horizontal, where we focus on a given variable and observe its behaviour across the timeline. We propose an extension of the Bernoulli-Gaussian model for variable selection to deal with time-varying sparsity by assuming a time dependence in the inclusion probabilities. We tackle the inference within a variational Bayes framework and we provide a global flexible approximation of the latent states exploiting a non-stationary Gaussian Markov random field representation. The properties of Bernoulli-Gaussian model together with the computational efficiency of variational methods enable a fast estimation and signal extraction also in regressions with many predictors.
