Forecasting irregularly spaced data: An extension of double exponential smoothing

Abstract

Automated forecasts are often required, in practice, using data series from which certain points are missing and from data occurring at completely irregular time intervals. For instance, in computerised inventory control, fast methods of dealing with such data are required. There is an almost complete absence in the literature of computationally efficient methods for such a situation. This paper gives an extension of single and double exponential smoothing adapted to data occurring at irregular time intervals. These extensions are shown to have modest computational requirements and little sensitivity to initial conditions. Results of tests on sample data series are given showing only a minor decrease in accuracy with missing data, and indicating the appropriate method of choosing the smoothing parameter. Application of this method to published government time series is illustrated by two examples, firstly, to river water quality data originating from samples taken at irregular time intervals and, secondly, to divorce rate statistics from which certain points are missing due to summarizing the data. Successive summarizing of these series is found to have a negligible effect on forecast accuracy implying attractive cost saving possibilities in data collection and publication.