A Bayesian Technique for Selecting a Linear Forecasting Model

The specification of a forecasting model is considered in the context of linear multiple regression. Several potential predictor variables are available, but some of them convey little information about the dependent variable which is to be predicted. A technique for selecting the “best” set of predictors which takes into account the inherent uncertainty in prediction is detailed. In addition to current data, there is often substantial expert opinion available which is relevant to the forecasting problem. The approach taken here utilizes both data and expert judgment by incorporating them into a Bayesian predictive distribution. Precise forecasting models are constructed by selecting the set of predictors which minimizes a measure of variability in prediction. An empirical demonstration of the technique is provided.