Model Selection and Uncertainty in Online Prediction
Professor Bertrand Salem Clarke, Department of Statistical Science, University College - on sabbatical from Department of Statistics, University of British Columbia

Abstract: In this presentation we discuss a technique for combining model selection principles (MSPs) in a prequential setting. This is a work in progress so helpful comments and questions will be very much appreciated. Combining MSPs can be done in several ways to improve prediction. We fix one simple way and try to study its properties. This is different and possibly disociable from Bayesian model averaging. The central idea is that different MSPs represent physically different assumptions which we cannot a priori rule out. The MSPs therefore correspond to catchment areas of models on which one MSP or another tends to perform best. It is novel to combine MSPs. A second source of novelty is that in the present version of the procedure we leave open the possibility of omitting some data. This adaptability permits us to consider nonstationary settings, however we conjecture that in cases of substantial model misspecification (especially involving dependence) neglecting early data will be optimal. Under moderately strong assumptions we establish that this approach does no worse than a constant MSP strategy.

This seminar was held at the Department of Computer Science, Royal Holloway, University of London on 3 February 1999.

back