|
Abstract:
|
I discuss sparse, graphically structured dynamic models for multivariate time series, emphasizing scaling to higher-dimensions. Our class of simultaneous graphical dynamic linear models (SGDLMs) provides for (a) flexible "decoupled" modelling of dynamics in individual univariate series, while enabling (b) "recoupling" across series in an overall sparse multivariate dynamic structure. I describe the underlying direct and undirected graphical model basis, and discuss aspects of modelling and computation for sequential filtering and forecasting. Efficient computational strategies exploit GPU-parallelization. An example in financial time series is drawn from recent applications in policy-oriented economics and forecasting for portfolio decisions in finance. A key to these studies is the use of sparse SGDLMs to improve the characterization of patterns of multivariate stochastic volatility structures. The example highlights improved forecast accuracy, portfolio risk and return characteristics, as well as generating novel metrics to monitor and predict (short-term) global and local market risk.
|
