Abstract

I discuss some of our recent R&D with dynamic statistical models for multivariate time series forecasting that represents a shift in modelling approaches in response to the coupled challenges of scalability and model complexity. Building “simple” and computationally tractable models of univariate time series is a starting point. Decouple/Recouple is an overlaid strategy for coherent Bayesian analysis: That is, “decouple” a high- dimensional system into the lowest-level components for simple/fast analysis; and then, “recouple”– on a sound theoretical basis– to rebuild the larger multivariate process for full/formal/coherent inferences and predictions. I discuss Bayesian dynamic dependency networks (DDNs) and the broader class of simultaneous graphical dynamic linear models (SGDLMs) that define a framework to address these goals. Aspects of model specification, fitting and computation include importance sampling and variational Bayes methods to implement sequential analysis and forecasting. Studies in financial time series forecasting and portfolio decisions highlight the utility of the models. The advances in Bayesian dynamic modelling– and in thinking about coherent and implementable strategies for scalability to higher-dimensions (i.e. to “big, dynamic data”)– are nicely exemplified in these contexts.

Aspects of this talk represent recent joint work with: Zoey Zhao, 2013 PhD at Duke University, now at Citadel llc, Chicago; Lutz Gruber, 2015 PhD at the Technical University of Munich, now at Quantco, Cologne; and Meng Amy Xie, 2012 BS at Duke University, and current PhD student in Statistical Science at Duke.