Simultaneous model construction and noise reduction for hierarchical time series via Support Vector Regression

In several applications, there are hierarchically-organized time series that can be aggregated at various levels. In this paper, a novel Support Vector Regression approach is proposed for dealing with hierarchical time series forecasting. The main idea is to pool information across levels of hierarchy, preventing bottom-level series from deviate much with respect to the series at the upper levels. The reasoning behind this approach is to estimate robust bottom-level models that can deal with the intrinsic noise present at this level due to the lack of information. Two variants are presented: First, we solve a single optimization problem that constructs all the related regression functions together, relating the bottom level series with the root node, while the second variant pools relates the leaf nodes with their respective parent nodes. The proposed approach showed best performance when compared with the state of the art on hierarchical time series forecasting using well-known benchmark datasets.