A0464
Title: Fast forecast reconciliation using sub-hierarchies
Authors: Fotios Petropoulos - University of Bath (United Kingdom) [presenting]
Abstract: Hierarchical forecasting has been a prominent research topic for the last 15 years, mainly due to its practical relevance. Various reconciliation methods have been proposed, and these methods offer coherent point and probabilistic forecasts across the various aggregation levels coupled with improved forecasting performance across the hierarchy. However, one main issue of such reconciliation methods is their limited applicability on large hierarchies due to the computational complexity related to the matrix calculations involved. This issue is addressed by proposing an overarching approach to forecast reconciliation methods that is based on the construction of sub-hierarchies. The approach can be applied in conjunction with any known forecast reconciliation method, either for a point or for probabilistic forecasts. Sub-hierarchical forecasting not only renders the reconciliation calculations possible for hierarchies of any size but also results in robust improvements over existing reconciliation methods. The proposed approach is applied to a large hierarchical dataset in the retail context, and its value is showcased in practice.
