Forecasting Applications of an Adaptive Multiple Exponential Smoothing Model

Abstract

This paper introduces a class of multiple exponential smoothing models useful in automated or minimal intervention industrial forecasting systems. These models are an alternative to simple univariate exponential smoothing and Trigg and Leach type adaptive models, which treat time series as unrelated and so cannot explicitly accommodate interrelationships that may exist between two or more time series. Moreover, the multiple models are adaptive in that the smoothing matrix, which is a generalization of the smoothing constant of univariate models, changes from period to period. Maximum likelihood estimates of the model parameters, including the full variance-covariance structure as well as the smoothing matrix, are provided, thus freeing the model user from the need for making ad hoc estimates of parameter values, a feature of simple univariate exponential smoothing. The forecast performance of this multiple time series model is compared with that of other univariate models using automobile sales data and some promising results are obtained.