Data-Based Priors for Bayesian Model Averaging

The uncertainty of models is now becoming one of the most important issues in the process of dealing with practical applications. In order to improve reliability and accuracy of inference, one usually adopts the model averaging method instead of selecting a single final model through a model selection procedure. Under the Bayesian framework, two upper bounds of the risk are derived and the posteriors are obtained by minimizing the bounds with a fixed prior. Then we propose two databased algorithms to get proper priors for Bayesian model averaging in this paper. Simulations show that by using these priors, smaller mean squared prediction errors can be gotten both in synthetic data and real data studies, especially for the data of poor quality.