Dynamic Model Averaging for Online Prediction of Continuous and Discrete Processes

• Activity Number: 149
• Type: Invited
• Date/Time: Monday, August 2, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Memorial
• Author(s): Adrian E. Raftery*+
• Companies: University of Washington
• Keywords: Bayesian model averaging ; state space model ; Markov chain ; forgetting ; online prediction ; rolling mill

Abstract

I. J. Good pioneered the ideas that led to Bayesian model averaging (BMA). We extend BMA to the dynamic situation where predictions are updated as observations arrive and there is model uncertainty. In Dynamic Model Averaging (DMA) a state space model for the parameters of each model is combined with a Markov chain model for the correct model, which changes over time. The state space and Markov chain models are both specified in terms of forgetting, giving parsimony. When the model and parameters do not change over time, DMA is a recursive version of BMA, called Recursive Model Averaging (RMA). Versions for continuous and binary outcomes are proposed. The method is applied to prediction for a cold rolling mill, laparascopic surgery in children, and simulated data. This is joint work with Miroslav Karny, Pavel Ettler, Tyler McCormick, David Madigan and Randall Burd.