“Predictive Analytics for Stock Markets Using Graph Neural Networks (GNNs)”

Financial data is highly complex, with many interdependencies, exhibiting complex temporal correlations and relationship patterns. Therefore, the prediction for stock market movements can indeed be quite complex. In traditional models, stocks are treated in isolation while ignoring the network of relationships of companies and industries. This work introduces the concept of the Non-IID Spatial-Temporal Graph Neural Network (NIST-GNN) model capturing the spatial relationships of the companies as well as patterns in sequences of the stock. The model is built using a tool in the form of Semantic Company Relationship Graph, SCRG, based on cosine similarities from financial news embeddings. The Sharpe Ratio that is gained by the NIST -GNN over the benchmark models is 0.40, and thus its predictive performance and portfolio optimization gain is superior. Furthermore, it is also understandable how information diffuses mainly in between cross-correlated firms of one day lag within the market. These insights challenge traditional market efficiency views and offer new avenues for financial prediction research.