Modeling Financial Market Dynamics with Temporal and Relational Learning: An LSTM-GNN Approach

Abstract

This study proposes a stock market prediction method based on long short-term memory network (LSTM) and graph neural network (GNN) to simultaneously capture the time series characteristics of individual stocks and the market network relationship. Traditional stock prediction methods mainly rely on time series models such as LSTM and GRU, but these methods often ignore the complex interactive relationships between stocks. To solve this problem, this study constructs a graph structure of the stock market and uses GNN for relationship modeling, so that the prediction model can comprehensively consider the historical trading data of individual stocks and the overall market dynamics. The experiment is verified based on the trading data of the S&P 500 index in 2018, and MLP, 1D-CNN, GRU and Transformer are used as comparison models. The experimental results show that LSTM+GNN achieves the best performance in terms of mean square error (MSE), root mean square error (RMSE) and coefficient of determination (R²), verifying the effectiveness of integrating time series and market structure information. In addition, ablation experiments and optimizer sensitivity experiments further illustrate the contribution of GNN structure and AdamW optimizer in improving prediction accuracy. Future research can combine multimodal data and reinforcement learning to further optimize the model and explore more efficient financial market prediction methods to enhance the decision-making ability of intelligent trading systems.