Multiple time series forecasting with Graph Neural Networks

Time series forecasting aims to predict future values to support organizations making strategic decisions. This problem has been studied for decades due to its relevance in almost all industries and areas, ranging from financial data to product demand. Recently, modern solutions based on deep learning have gained popularity among academia and industry, mainly due to the necessity to automatize the forecasting of multiple time series and exploit external explanatory variables. Considering the recent successes of Graph Neural Networks (GNNs) in modelling graph data, this study extends previous works based on time series forecasting from visibility graphs. In particular, in the first direction, the link prediction task, targeted by local random walks in the related work, is resolved by custom GNNs. In the second direction, a new strategy based on graph regression using GNNs is proposed to learn graph representations able to combine hidden historical patterns and external features. The M5 competition dataset is used to compare the proposed models to the related work and other traditional and machine learning benchmarks. Final results show promising performances on various higher levels of the M5 competition and delineate multiple limitations from the related work.