A Hybrid Deep Learning Model for Dynamic Stock Movement Predictions Based on Supply Chain Networks

An inter-firm relationship of paramount importance is captured by supply chain networks.Embedded in such a network, a firm’s performance is associated with its partners’ and peers’performance. This paper proposes an end-to-end predictive framework named Hybrid andTemporal Graph Neural Network (HT-GNN) to predict the dynamic stock price movement offirms. The model learns time-dependent node embeddings by aggregating network neighbors’features and market trends to provide node classifications over time. Experiments on a real-worldsupply chain network among over 2,700 publicly traded firms show that HT-GNN can improvedynamic stock movement predictions. We define different types of network neighborhoods byidentifying firms that contribute to such predictions in different ways and going even beyondimmediate ties. Our results would naturally help investors understand stock price movement andmanagers identify network neighbors with predictive power over its own stock price movement