A machine learning approach for forecasting hierarchical time series

TL;DR

This paper introduces a deep neural network method that directly produces accurate and hierarchically reconciled forecasts for time series data, outperforming existing techniques by integrating hierarchy structure and explanatory variables during training.

Contribution

It presents a novel end-to-end neural network approach that directly generates reconciled forecasts, incorporating hierarchy structure and explanatory variables in training.

Findings

Outperforms state-of-the-art hierarchical forecasting methods on real datasets.

Effectively integrates hierarchy structure into neural network training.

Utilizes explanatory variables to improve forecast accuracy.

Abstract

In this paper, we propose a machine learning approach for forecasting hierarchical time series. When dealing with hierarchical time series, apart from generating accurate forecasts, one needs to select a suitable method for producing reconciled forecasts. Forecast reconciliation is the process of adjusting forecasts to make them coherent across the hierarchy. In literature, coherence is often enforced by using a post-processing technique on the base forecasts produced by suitable time series forecasting methods. On the contrary, our idea is to use a deep neural network to directly produce accurate and reconciled forecasts. We exploit the ability of a deep neural network to extract information capturing the structure of the hierarchy. We impose the reconciliation at training time by minimizing a customized loss function. In many practical applications, besides time series data, hierarchical time series include explanatory variables that are beneficial for increasing the forecasting accuracy. Exploiting this further information, our approach links the relationship between time series features extracted at any level of the hierarchy and the explanatory variables into an end-to-end neural network providing accurate and reconciled point forecasts. The effectiveness of the approach is validated on three real-world datasets, where our method outperforms state-of-the-art competitors in hierarchical forecasting.