FinHGNN: A conditional heterogeneous graph learning to address relational attributes for stock predictions

Recent financial studies have shown that spillover effects of some market factors play a significant role in stock fluctuations. Previous studies, however, were incapable of capturing the spillovers of these relational market factors because they relied on a homogeneous graph that condenses these factors into firm node attributes and requires their spillover effects to follow firm relationship instead of themselves. This fact brings up a heterogeneous graph learning problem that requires multiple node types to transport different spillover effects. This study proposes a novel conditional heterogeneous graph neural network (FinHGNN) to capture multiple spillover effects in asset pricing with two uniquely designed mechanisms. First, it presents an efficient way to preserve the connectivity of relational attributes in graph learning, which is achieved by converting relational attributes into node variables to form a heterogeneous graph. Second, a conditional message-passing mechanism is proposed to handle multiple spillover effects simultaneously by messaging conditioned on different types of nodes and node attributes. This study paves the way for addressing the relational attributes in graph learning. Experiments on two real-world datasets demonstrate the advantages of the proposed framework over three classic and four state-of-the-art algorithms, including LSTM, GCN, HGNN, eLSTM, TGC, FinGAT, and AD-GAT.