Using All Observations when Forecasting under Structural Breaks

We extend the idea of the trade-off window approach by Pesaran and Timmermann (2007) of using observations preceding the last structural break to estimate model parameters for the purpose of forecasting. Our weighted least squares method utilizes information in all observations but with weights varying from one to another interval between breaks. This leads to a smaller mean squared prediction error which is illustrated by simulations. The proposed procedure is computationally simple having a convenient associated optimization program. We also describe and evaluate a cross-validation analog of the proposed method.