Title: Bernoulli vector autoregressive model
Authors: Carolina Euan - King Abdullah University of Science and Technology (Saudi Arabia) [presenting]
Ying Sun - KAUST (Saudi Arabia)
Abstract: Categorical time series appear in many fields such as biology, industry, stocks markets and environmental sciences. Even for univariate binary time series, the analysis is usually more challenging than time series analysis for continuous variables. In a multivariate setting, modeling the dynamics in multiple binary time series is not an easy task. Most existing methods model the joint transition probabilities from marginals pairwisely. However, the resulting cross dependency may not be flexible enough. We propose a vector autoregressive (VAR) model for multivariate binary time series. The model is constructed by latent multivariate Bernoulli random vectors. The Bernoulli VAR model represents the instantaneous dependency between components via latent processes, and the autoregressive structure represents a switching between the hidden vectors depending on the past. Our proposed model provides an intuitive interpretation when analyzing real data sets. We derive the mean and matrix valued autocovariance function for the Bernoulli VAR model analytically and develop a likelihood based inference.