A0902
Title: On the difference of the existing hierarchical forecasting approaches
Authors: Nikolaos Kourentzes - University of Skovde (Sweden) [presenting]
George Athanasopoulos - Monash University (Australia)
Abstract: Hierarchical forecasting refers to a class of forecasting problems where different time series are interconnected through aggregation constraints. For instance, cross-sectional hierarchies appear in supply chains, where sales at the store/product level aggregate to sales across various demarcations, such as product categories. Nowadays, the standard approach is to produce independent forecasts for each separate series and reconcile them across the hierarchy to ensure that aggregation constraints are met. The literature is rich with different reconciliation estimators, which, at their core, combine the base forecasts into the reconciled ones. Their quality is typically considered in terms of forecast accuracy. Leveraging the geometric interpretation of forecast reconciliation, it is demonstrated that existing solutions explore a very limited range of the possible solution space. It is first characterized by that space, and then it investigates how these unexplored solutions can be obtained, providing a new class of estimators. These solutions are benchmarked in terms of efficiency and forecasting performance.
