Reproduce “Hierarchical Forecasting at Scale” on OpenResearch

About this paper

Olivier Sprangers, Wander Wadman, Sebastian Schelter, Maarten de Rijke

This research introduces a sparse hierarchical loss function to enable efficient, end-to-end coherent forecasting across millions of time series, specifically designed for large-scale e-commerce operations. The method successfully integrates cross-sectional and temporal hierarchies, demonstrating improved accuracy and dramatic computational efficiency compared to traditional reconciliation methods on both public and proprietary datasets.

Key insights

• Embedding hierarchical coherence directly into the loss function allows for end-to-end coherent forecasting at industrial scale, eliminating the need for separate reconciliation steps.
• Normalizing the loss gradient by the number of elements in an aggregation is essential for stable learning, preventing gradient magnitudes from excessively growing with hierarchy size.
• Leveraging the inherent sparsity of hierarchical aggregation matrices through sparse linear algebra is fundamental to achieving dramatic computational efficiency and practical applicability in large-scale systems.

Summary by alphaXiv.