A quantile regression approach to generating prediction intervals

Exponential smoothing methods do not involve a formal procedure for identifying the underlying data generating process. The issue is then whether prediction intervals should be estimated by a theoretical approach, with the assumption that the method is optimal in some sense, or by an empirical procedure. In this paper we present an alternative hybrid approach which applies quantile regression to the empirical fit errors to produce forecast error quantile models. These models are functions of the lead time, as suggested by the theoretical variance expressions. In addition to avoiding the optimality assumption, the method is nonparametric, so there is no need for the common normality assumption. Application of the new approach to simple, Holt's, and damped Holt's exponential smoothing, using simulated and real data sets, gave encouraging results.