A new method of hierarchical time series prediction based on adjustable error

Abstract

In the area of big data, hierarchical time series prediction (HTSP) has attracted extensive attention. For the traditional HTSPs, the uncertainty of model selection cannot be alleviated due to the shortcomings of data information loss. This article proposes a new HTSP method based on the adjustable error to deal with the above issue. Detailly speaking, in the essential prediction stage, the adjustable prediction error, including all data information, is introduced to improve the current prediction. By comparison of four cases of optimal single prediction, optimal combined prediction, optimal single prediction under the new method, and optimal combined prediction under the new method, the two assertions are obtained: (1) the prediction accuracy of this new method is significantly improved compared with the widely used methods such as bottom-up, generalized least squares; (2) the uncertainty of model selection is greatly reduced. The method’s validity is verified by the examples of American clothing retail and the Germany national electricity consumption.