Flow-Based End-to-End Model for Hierarchical Time Series Forecasting via Trainable Attentive-Reconciliation

Shiyu Wang, Yinbo Sun, Yan Wang, Fan Zhou, Lintao Ma, James Y. Zhang, YangFei Zheng

Abstract: Time Series (TS) is one of the most common data formats in modern world, which often takes hierarchical structures, and is normally complicated with non-Gaussian and non-linear properties. Many businesses rely on accurate TS forecasting, under these complications, to help with operational efficiencies. In this paper, we present a novel approach for Hierarchical Time Series (HTS) prediction via trainable attentive reconciliation and Normalizing Flow (NF), which is used to approximate the complex (normally non-Gaussian) data distribution for multivariate TS forecasting. To reconcile the HTS data, we also propose a new flexible reconciliation strategy via the attention-based encoder-decoder neural network, unlike the existing methods with strong assumptions, such as unbiased forecasts and Gaussian noises. In addition, by using the reparameterization trick, we are able to combine forecasting and reconciliation into a trainable end-to-end model. Extensive empirical evaluations are conducted on real-world hierarchical datasets and the preliminary results demonstrate the efficacy of our proposed method.