Dynamic Model Averaging in the Context of Prequential Prediction

• Activity Number: 10
• Type: Invited
• Date/Time: Sunday, July 31, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Council of Chapters
• Author(s): Jennifer Clarke*+ and Chi Wai Yu and Bertrand Salem Clarke
• Companies: University of Miami and Hong Kong University of Science and Technology and University of Miami
• Keywords: model averaging ; prediction ; prequential ; median loss ; regression

Abstract

Many regression problems involving complex or high dimensional data are not amenable to conventional modeling because the uncertainty associated with the use of any single model is too high to permit reliable inferences. We investigate the effect of using the median of the loss function as an optimality criterion in the context of model averaging. We consider median loss in a prequential context and offer an extension of the median prediction technique of Berger and Barbieri (2004) that emphasizes the role of the model list used to form predictions. We will present examples of our method and comparisons to a variety of model averaging techniques via their predictive performance (broadly defined) to identify settings where one technique outperforms another.