HierNBeats: Hierarchical Neural Basis Expansion Analysis for Hierarchical Time Series Forecasting

Hierarchical time series forecasting has become increasingly prominent in numerous practical applications. While various deep neural networks have been proposed for forecasting non-hierarchical time series, there has been a lack of extensive research focused on hierarchical time series forecasting using deep learning techniques. Moreover, existing studies using deep neural network algorithms for hierarchical time series forecasting have lacked attention to the transparency and interpretability of the models. This paper proposes a novel method called HierNBeats (Hierarchical Neural Basis Expansion Analysis for Interpretable Hierarchical Time Series) by improving upon the renowned algorithm NBeats (Neural Hierarchical Basis Expansion Analysis for Interpretable Time Series), which provides interpretable predictions for non-hierarchical time series using deep learning. HierNBeats provides a novel multi-branch structure to enable the network to generate neural basis corresponding to information from time-series at different levels, which allows HierNBeats to exploit the information available across all levels to produce coherent forecasts. Moreover, HierNBeats has a certain interpretability of the network outputs to improve transparency in the modeled processes. We assess the value of HierNBeats by comparing it to state-of-the-art hierarchical time-series forecasting approaches across three public datasets and observe a relative improvement over the state-of-the-art.