High-Dimensional Dynamic Models and Time-Adaptive Sparsity in Networks

• Activity Number: 143
• Type: Invited
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Mike West*+
• Companies: Duke University
• Keywords: Bayesian forecasting ; Bayesian sparsity modelling ; Dynamic networks ; GPU computing ; Multivariate time series ; Parallel computation

Abstract

We discuss developments in Bayesian modeling and computation for multivariate time series and dynamic networks. The need for coherent analysis of high-dimensional time series raises challenges to concepts, theory, and computation. These are addressed in recent developments in Bayesian sparsity modeling for dynamic models. Topics presented will be drawn from (i) latent threshold modeling, a critical concept in enabling dynamic sparsity and adaptation in time series and network studies that has been demonstrably valuable in a number of recent studies; (ii) restructuring multivariate dynamic modeling strategies to enable analysis decoupling and massively parallel processing; and (iii) the concept of "lean computing" that drives model reformulations for efficient sequential processing in increasingly high-dimensional problems. Applications---both motivating and currently topical---will be drawn from routine analysis applications in areas including macro-economics, finance portfolio studies, very large-scale commercial demand and sales forecasting, and large-scale networks.