Thesis : Bayesian Model Averaging [library link]

committee : G. Iliopoulos (supervisor), M. Kateri , I. Ntzoufras

Abstract

It is common to discard the uncertainty employed by the choice of a single model. This source of  variation  is  highly  significant  when  the  purpose  of  the  analysis  is  the  prediction  of  a quantity  of  interest  (e.g  a  future observation,  the  cost  of  a decision  etc.). A way  of  dealing with the aforementioned problem is the use of weighted top (or all if possible) models among the class of the models the analyst considers. The bayesian paradigm deals with the problem naturally,  by  simply  regarding  the models  as  parameters  (therefore  assigning  them  a  prior probability) and using the method of Bayesian Model Averaging.

This dissertation aims to present the theoretical background along with practical algorithms in the class of linear models.

Download [pdf/In Greek] & Slides [pdf/In Greek]