Causal partitioning and sales forecasting accuracy

Abstract

This paper examines the benefits of partitioning a data set into components and then forecasting the components and adding (combining) the component forecasts as a method of forecasting a time series. The methodology is a modification of the combining forecasts methodologies that have proven in past studies to be a superior forecasting method. Three different partitioning methodologies are explored in this paper. For all three, an example is given where partitioning and then recombining produces more accurate results than the best nonpartitioned forecast. The paper also explores some reasons why combining methodologies produce more accurate results than a single best model forecast.