Mike West: Structured Dynamic Graphical Models & Scaling Multivariate Time Series Methodology
Prof. Mike West (Duke University, Dept. of Statistical Science) will give a seminar in room 108, ground floor N.H. Abel's House at 14:15 October 13th.
Title: Structured Dynamic Graphical Models & Scaling Multivariate Time Series Methodology
Abstract: This talk concerns recent developments in time series using concepts of graphical modelling. The aims are efficient scaling of Bayesian dynamic/predictive modelling to increasing dimensions, including flexibility in multivariate volatility (cross-sectional) modelling. Generalizations of multi-regression dynamic linear models - referred to as dynamic dependency network models (DDNMs) - involve underlying, directed acyclical graphical model structures. These form a special subset of the more general class of simultaneous graphical dynamic models (SGDLMs) that are based on underlying undirected graphs. Some aspects of model theory, specification, fitting and computation will be noted, including the use of importance sampling and variational Bayes methods to implement sequential analysis and forecasting. A core conceptual focus is that of sequentially repeated *Decoupling/Recoupling*: decoupling of a large multivariate system into univariate components for (parallel, perhaps GPU-based) efficient analysis, followed by recoupling across series to define coherent Bayesian inferences and predictions of the full multivariate process. Some applied contexts and illustrations come from financial time series analysis, multi-step forecasting, and associated Bayesian decision analysis in portfolio studies. This work draws on current and research with Lutz Gruber (TUM & Duke) and Zoey Zhao (Duke & Citadel).