Research indexJournal article
Optimizing Hierarchical Time Series Forecasting using Machine Learning Algorithms
Mahalik, Panigrahi
21 Aug 2025
NobleIDark:/48914/w1/20260525/B49C160BDOI10.1109/indiscon66021.2025.11252070
Optimizing Hierarchical Time Series Forecasting using Machine Learning Algorithms
Mahalik, Panigrahi · 2025 · Subscription
NobleIDark:/48914/w1/20260525/B49C160BDOI10.1109/indiscon66021.2025.11252070
The full text is at the publisher
You can read the abstract here, or open it at the publisher.
Abstract
Hierarchical forecasting (HF) methods are widely employed to support accurate decision-making by ensuring coherence across multiple aggregation levels. Traditionally, statistical models have been utilized for prediction in hierarchical forecasting (HF). Inconsistencies from their restricted expressiveness, which frequently overlooks the dynamic, nonlinear, and aperiodic features of time series during aggregation and disaggregation. This study addresses these limitations by employing machine learning (ML) techniques to model such complex behaviors in HF. After post-forecast reconciliation, methods consisting top-down (TD), bottom-up (BU), middle-out (MO), minimum trace (MT), and optimal combination (OC) over one to multi-step horizons on the M5 competition dataset, a performance comparison of 17 machine learning models is carried out at each hierarchical level. Model performance is ranked for interpretability utilizing nonparametric statistical metrics. The study methodically answers ten HF-related research concerns. Simulation results suggest that the k-nearest neighbours regression (KNNR) model over the BU reconciliation approach provides statistically superior performance among all the ML models, considering eight distinct horizon forecast simulation results. It is also observed that applying ML algorithms at specific hierarchical nodes, followed by reconciliation, significantly enhances forecast accuracy across the hierarchy.