SLOTH: Structured Learning and Task-Based Optimization for Time Series Forecasting on Hierarchies

Multivariate time series forecasting with hierarchical structure is widely used in real-world applications, e.g., sales predictions for the geographical hierarchy formed by cities, states, and countries. The (HTS) includes two sub-tasks, i.e., reconciliation. In previous works, information only integrated reconciliation step to maintain coherency, but not accuracy improvement. this paper, we propose novel tree-based feature integration mechanisms, top-down convolution bottom-up attention leverage of improve performance. Moreover, unlike most methods which either rely on strong assumptions or focus coherent constraints only, utilize deep neural optimization networks, achieve coherency without any assumptions, also allow more flexible realistic task-based targets, lower under-estimation penalty meaningful decision-making loss facilitate subsequent downstream tasks. Experiments datasets demonstrate that our mechanism achieves superior performances tasks compared state-of-the-art methods, networks can be applied effectively additional effort under coherence constraints.