Spatial-Temporal Stock Movement Prediction and Portfolio Selection based on the Semantic Company Relationship Graph

In this paper, we approach stock price movements as a spatial-temporal prediction task, advancing beyond the traditional view of stocks as standalone entities. We first represent companies as vector embeddings, utilizing company name co-occurrence statistics from a large financial news corpus, and then construct a Semantic Company Relationship Graph (SCRG) using cosine similarities between vectors to define the mutual relationships. To tackle the financial prediction task, we introduce a novel Non-Independent and Identically Distributed Spatial-Temporal Graph Neural Network (NIST-GNN). It is specifically designed to propagate features from both neighboring companies and internal historical data while effectively handling the inherent temporal non-IIDness in stock sequences. This innovative aspect of our NIST-GNN allows for a more nuanced understanding and processing of temporal data, setting it apart from traditional spatial-temporal approaches. Our experimental results demonstrate that this methodology significantly outperforms benchmark models, yielding superior profitability and enhancing the Sharpe Ratio by 0.61 compared to the best-performing baseline, with statistical significance. Importantly, our findings provide valuable theoretical insights into the effect of information diffusion within the US market, revealing that public information from cross-correlated companies typically experiences a minimum one-day lag before diffusion, challenging conventional perceptions of market efficiency.