Scalable forecast reconciliation through (un)guided sub-hierarchies

Hierarchical forecasting has become important due to its practical value, but existing reconciliation methods struggle with large hierarchies because of their high computational cost. The paper proposes a scalable, general framework based on constructing sub-hierarchies to overcome this limitation. By breaking a large hierarchy into smaller sub-hierarchies, any reconciliation method, whether for point or probabilistic forecasting, can be applied efficiently within each sub-hierarchy. Sub-hierarchies can be defined randomly or through clustering, offering flexibility and allowing the approach to adapt to different data structures. This makes reconciliation feasible even for extremely large hierarchies that would otherwise be computationally prohibitive. We evaluate the method using real-world retail and tourism datasets. Empirical results show that sub-hierarchical forecasting not only enables practical computation but also improves accuracy. We report performance gains over existing reconciliation methods, with improvements exceeding 10% at specific hierarchy levels and up to 4% (retail) and 7.5% (tourism) across the entire hierarchy. Overall, the proposed sub-hierarchical approach provides a powerful, adaptable framework that maintains coherence, scales to any hierarchy size, and delivers consistent forecasting improvements.