Graph Neural Networks for Multivariate Time Series Forecasting

Abstract

Stock market analysis is a hot topic with a very active research community working on it. Stock price prediction is of particular interest owing to the high stakes it offers. It is actually a multivariate time series problem. The fact that the value of a stock is dependent on the performance of other stocks in a hidden way has motivated the research on Graph Neural Networks (GNNs) in a hope to decipher these hidden dependencies. Existing GNNs for stock prediction require pre-defined topology of the graphs to work upon, this is a drawback as it is not always possible to route down complex relationships that exist between different stocks. Moreover, auto-generated graphs are better than the ones based on human knowledge. This work proposes two data driven graph generation methods: one correlation based method and an adaptive graph generation method. These methods in combination with Graph Convolution Neural networks (GCNs) and Graph Attention Networks (GAT) as spatial dependency modelling modules and LSTM as temporal dependency modelling module, three different graph learning models have been developed and tested against baseline machine learning models. It has been shown that the deep learning models could outperform the chosen baselines consistently in all the analysis confirming the superior efficiency of GNNs in modelling inter-series dependencies. An improved efficiency means that the model could learn the sparse short-term trends in the noisy stock time series as and when it occurs. This has reflected in the increase of returns by a minimum of 131.8% in comparison to the baseline models in the market simulation experiment carried out.