Exponential Smoothing with an Adaptive Response Rate

Abstract

A modification is proposed to forecasting systems employing exponential smoothing whereby the response rate is varied and made to depend on the value of a tracking signal. In a simple system, this is equivalent to varying α the smoothing constant according to the extent to which biased forecasts are being obtained. Such a system is shown to react much faster to, for example, step changes whilst still retaining the facility to filter out random noise.