Forecasting 1 to h steps ahead using partial least squares.

This paper proposes a methodology to jointly generate optimal forecasts from an autoregression of order p for 1 to h steps ahead. The relevant model is a Partial Least Squares Autoregression, which is positioned in between a single AR(p) model for all forecast horizons and different AR models for different horizons. Representation, estimation and forecasting using the new model are discussed. An illustration for US industrial production shows the merits of the methodology.