B1836
Title: Unified methodology for observation- and parameter-driven models for time series
Authors: Takis Besbeas - Athens University of Economics and Business (Greece) [presenting]
Abstract: Discrete time series are frequently encountered in a variety of scientific disciplines and are often subject to covariates in addition to the zero-inflation and overdispersion. Two general classes of time series models have been proposed in the literature to handle such data: the class of observation-driven models (ODMs) and the class of parameter-driven models (PDMs). In the former, autocorrelation is introduced through the dependence of the conditional mean of the current response on its past values, while in the latter this is achieved through an unobserved underlying random process. ODMs and PDMs are formulated in distinct frameworks, namely the partial likelihood and state-space frameworks respectively. A unifying approach is introduced where autocorrelation is introduced through the dependence on both past response values as well as a latent stochastic process, thereby combining OD and PD models into the same framework. A new method is proposed for model-fitting based on hidden Markov model methodology, involving a discretization technique of the underlying state-space and Markov-switching autoregression. The approach is illustrated using real data from the environmental and medical sciences and its performance is evaluated using simulation.
