Is dimensionality reduction a curse? Bayesian analysis of the mean-volatility dynamic factor model

Abstract

The factor stochastic volatility (FSV) model is a powerful dimensionality reduction device and has gained much popularity due to its parsimonious structure in modelling both time-varying mean and covariance matrix for multivariate time series. We document the failure of FSV models by observing a strong common volatility component left in the residuals, irrespective of the chosen number of factors. To adequately model co-movement in mean and volatility, we introduce the mean-volatility dynamic factor model which assumes separate factor structures for the first and the second moment of a high-dimensional vector time series. We identify and extract the mean factors and volatility factors via a Bayesian variable selection technique that pins down zeros in associated loading matrices and thus the factor space. We also propose a computationally efficient multi-move sampler that samples all volatility series in parallel to speed up estimation. In the empirical study, we fit the model to the Fred-MD data of 128 monthly macroeconomic variables and find 18 mean factors and 8 volatility factors.

About the presenter's visit

Dr Li is willing to have (30-min) 1-to-1 meeting with our colleagues and graduate students during 1:00 – 4:00pm on the seminar day. Anyone who wants to meet with him on June 2nd can contact Dr Fu Ouyang (f.ouyang@uq.edu.au) to book a time slot.