Jonathan Stroud
Department of Statistics
George Washington University
Dynamic state-space models are widely used in many fields, including engineering, finance and environmental science. In many applications, the models need to be estimated sequentially in time, as new data arrive. This presents many challenges, especially when the models are nonlinear, non-Gaussian or involve unknown parameters. In this talk, we will describe the existing approaches for on-line Bayesian state and parameter estimation in general state-space models: (1) state augmentation; (2) sufficient statistics; and (3) rolling-window MCMC. We will discuss the strengths and weaknesses of each of these approaches, and demonstrate them using a benchmark autoregressive plus noise model, a stochastic volatility model and a high-dimensional spatio-temporal model. We will then introduce a new method based on Gaussian mixtures, and show that it outperforms existing approaches, particularly when the state vector is high-dimensional. We demonstrate the advantage of the new method on simulated data and an application to spatio-temporal modeling of cloud motion data
All statistics students are expected to attend.