B1279
Title: Copula modelling of serially correlated multivariate data with hidden structures
Authors: Robert Zimmerman - University of Toronto (Canada) [presenting]
Vianey Leos Barajas - University of Toronto (Canada)
Radu Craiu - University of Toronto (Canada)
Abstract: In applications where streams of data exhibit variable latent structures, it is natural to model the data-generating process as a finite-state hidden Markov model (HMM). When observing vectorial outcomes, we consider multivariate state-dependent distributions that are fused together by copulas. Such a ``copula-within-HMM'' framework is highly flexible, because it provides the freedom to vary both the marginal distributions of observed outcomes and the copula that determines the dependencies between them. However, inference for this model is not straightforward; while the EM algorithm is the standard technique for parameter estimation within HMMs, a direct application becomes unwieldy in the face of the additional model complexity brought about by the copula. We develop a robust and efficient EM algorithm for the copula-within-HMM model, and show that it performs well in both model estimation and state classification tasks on a variety of simulated and real-world datasets.
