Abstract:
|
The dynamic latent threshold concept, which allows for adaptive, dynamic sparsity modeling in multi-variate time series, has involved intensive MCMC methods in its analysis and model fitting. Here, in contrast, we address sequential Bayesian learning in dynamic latent threshold models (LTMs). The methodological advances consist of: (1) customization of existing methods of sequential Monte Carlo (SMC), including the auxiliary particle filter and particle learning, for LTMs; and (2) a novel approach to model emulation, using analytically tractable dynamic linear model-based filtering to define emulators of the inherently complex, nonlinear threshold models. This latter approach allows us to integrate more theoretical aspects of the model structure of LTMs into approximate sequential simulation-based analysis; this defines a new SMC strategy that yields improved performance in filtering and forecasting. This is exemplified in an applied study in macroeconomics, where the efficacy of our emulation approach is clearly evident. This is quantified using several metrics including marginal likelihood, effective sample size, and traditional metrics for point forecasting accuracy.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.