NeuralReconciler for Hierarchical Time Series Forecasting

Abstract

Time series forecasting has wide-ranging applications in business intelligence, including predicting logistics demand and estimating power consumption in a smart grid, which subsequently facilitates decision-making processes. In many real-world scenarios, such as department sales of multiple Walmart stores across different locations, time series data possess hierarchical structures with non-linear and non-Gaussian properties. Thus, the task of leveraging structural information among hierarchical time series while learning from non-linear correlations and non-Gaussian data distributions becomes crucial to enhance prediction accuracy. This paper proposes a novel approach named NeuralReconciler for Hierarchical Time Series (HTS) prediction through trainable attention-based reconciliation and Normalizing Flow (NF). The latter is used to approximate the complex (usually non-Gaussian) data distribution for multivariate time series forecasting. To reconcile the HTS data, a new flexible reconciliation strategy via the attention-based encoder-decoder neural network is proposed, which is distinct from current methods that rely on strong assumptions (e.g., all forecasts being unbiased estimates and the noise distribution being Gaussian). Furthermore, using the reparameterization trick, each independent component (i.e., forecasts via NF and attention-based reconciliation) is integrated into a trainable end-To-end model. Our proposed NeuralReconciler has been extensively experimented on real-world datasets and achieved consistent state-of-The-Art performance compared to well-Acknowledged and advanced baselines, with a 20% relative improvement on five benchmarks.