New Zealand Statistical Association 2024 Conference

Guoping Hu

University of Auckland

Forecasting Multiple Time Series With Graph Convolutional Networks

This is joint work with Ciprian Giurcaneanu

Forecasting multiple time series at different levels is often required in many situations, which is commonly known as hierarchical time series forecasting. Supply chain management is a typical application that requires demand forecasting at the store, city, or country level for decision-making. In hierarchical forecasting, top-down, bottom-up, and optimal linear combination methods are the most common methods. While top-down and bottom-up methods use only information from the top and bottom levels, respectively, linear combination methods use individual forecasts from all series and levels and combine them linearly, often outperforming traditional top-down and bottom-up methods. Despite this, these approaches do not make use of the explanatory information that may exist at various levels of the hierarchy directly. In addition to producing accurate forecasts, it is necessary to select a suitable method to generate basic and reconciled forecasts simultaneously. Prediction reconciliation involves adjusting predictions to be consistent across different levels. In this talk, we present a neural network model that utilizes graph convolutional neural networks, recurrent neural networks, and fully connected neural networks to generate accurate and reconciled predictions directly. Specifically, we first use graph convolutional neural networks to extract hierarchical information, then recurrent neural networks to extract the temporal dependencies of all time series in the hierarchy, and finally, train a fully connected neural network to minimize the loss function to generate reconciled forecasts.