A0230
Title: Adaptive Bayesian factor spectral analysis of high--dimensional nonstationary time series
Authors: Zeda Li - City University of New York (United States) [presenting]
Ori Rosen - University of Texas at El Paso (United States)
Robert Krafty - University of Pittsburgh (United States)
Abstract: A frequency-domain factor model is proposed to spectral analysis of high-dimensional nonstationary time series. The model provides a general framework for estimating both real- and complex-valued spectra by allowing a time series acts simultaneously and propagates in a lagged fashion. Real and imaginary parts of the factor loading matrix are modeled independently by tensor products of penalized splines and multiplicative gamma process shrinkage priors, which allows infinitely many factors with the loadings increasingly shrunk towards a constant function as the column index increases. Formulated in a fully Bayesian framework, a conditional Whittle likelihood-based Gibbs sampler is developed for efficient model fitting. By using stochastic approximation Monte Carlo (SAMC) and partitioning a time series into an unknown number of approximately stationary segments, the approach automatically and adaptively estimates the power spectrum of both stationary and nonstationary high-dimensional time series.
