Application of Network Embedding to Return Comovement Analysis: Case of Stock Co-attention Networks

This paper proposes a novel application of network embedding in the context of a stock market. We construct and represent stock networks based on investors’ co-attention relationships by learning network embedding using random walk (e.g. DeepWalk, Node2Vec), deep neural networks (e.g., SDNE) and matrix factorization (e.g., HOPE), respectively. Through clustering all the stocks based on their embedding vectors, we separate stocks listed in the market into different clusters. The empirical results suggest that stocks within the same clusters exhibit significant comovement and the observed comovement cannot be fully explained by the similarity in firm fundamentals and the comovement within the same industry, location and trading board.