Multi-Cluster Graph (MCG): A Novel Clustering-Based Multi-Relation Graph Neural Networks for Stock Price Forecasting

Stock price forecasting is one of the indispensable challenges, as financial markets are increasingly becoming volatile due to the complex, noisy, dynamic, and chaotic nature of stock data. Accurate stock price forecasting models play a key role in protecting risks associated with markets and making sound investing decisions. Mostly, existing studies suggested different Machine Learning (ML) and Deep Learning (DL) based methodologies, but they often focus on individual stock data to generate predictions, ignoring the interconnectivity and interconnections that exist between stocks. Recent advances in graph neural networks highlighted the need to model stock-to-stock interactions to improve the reliability of forecasting. Regardless of the potential, modeling these interactions among stocks poses substantial challenges. Existing methods encompass basic heuristics like sector affiliations, shareholdings, or co-occurrence in news headlines, etc., to model relationships among stocks which can disregard the intricate and varied nature of stock interconnection. To address such issues, this study proposed a clustering-based approach to exploit relationships among stocks listed on particular stock exchanges. More precisely, different graph models are created using three proposed strategies including RS (Returns-Sharpe Ratio), VR (Volatility-Return), and YR (Yearly variance and returns) clustering. These clusters are employed to create multi-relation graphs which are then propagated to an LSTM-based graph network to forecast future prices at the node level. The proposed approach can anticipate future stock prices for all stocks featured in an NSADAQ and S&P500 index while utilizing graph nodes from a single-trained standalone model. Moreover, experimental results on NSADAQ-100 and S&P500 stocks, show that the proposed clustering-based stock graph models methodology shows good results in terms of lowest MSE and MAE values of 0.01 and 0.08 respectively.