Exponential smoothing methods have been popular and effective methods of forecasting for many years. But until recently, they have been considered "ad hoc" and without any underlying stochastic modelling framework. I will discuss a new state space modelling framework which includes all the usual exponential smoothing methods, and several other related methods. This framework allows (1) easy calculation of the likelihood, the AIC and other model selection criteria; (2) computation of prediction intervals for each method; and (3) random simulation from the underlying state space model. It also provides the basis for a new automatic forecasting algorithm. I will demonstrate the algorithm by applying it to the data from the M-competition and the M3-competition. The resulting forecasts are as accurate as the best methods in the competitions; and beat all other methods for short forecast horizons with seasonal data.

I will also discuss some of the properties of the state space models underlying the various exponential smoothing methods and show (1) that some well-known methods are inherently unstable and have infinite forecast variances; and (2) that the usual parameter space can lead to non-invertible models.
 

Convenors:Aidan Sudbury