Abstract:
Stock markets are dynamic, stochastic, and non-linear in nature, and the price prediction of stock markets has long remained a complex problem. In recent years, deep learning has gained a lot of hype and depth, with models like Long Short-Term Memory (LSTM) modules providing tremendous results when applied to understanding temporal relations, and Graph Neural Networks (GNNs) have also proved effective in describing relational structures. Although these advances have made it possible to deal with both temporal dynamics and structural dependencies, there is a tendency for the models to commonly fail to address them simultaneously, and, in most cases, they are black boxes with limited interpretability. This gap contributes towards their weaknesses and impairs their application in high stakes even in areas like in financial applications where the process of explaining and understanding is vital. To address this issue, this research paper proposes a hybrid LSTM-GNN model that uses Explainable AI (XAI). The LSTM unit represents the long-term temporal dynamics, and GNN model represents the patterns of relations between the stock characteristics. The LSTM with GNN combination offers a combined structure of learning that deploys both sequential and structural learning. Moreover, the Local Interpretable Model-Agnostic Explanations (LIME) is incorporated to increase transparency by identifying important features in making model predictions therefore providing enhanced trust and explainability. The empirical experiments involving the stock market data of Tata Motors (2000-2021) revealed that the proposed hybrid model of LSTM-GNN was stable and has performed better as compared to those of independent LSTM, GNN, and the traditional statistical models. The hybrid architecture was found to be a superior reliable architecture with an RMSE of 4.12, MAE of 2.90, and an R 2 of 0.89 that translates to an accuracy of almost 80-90 percent in terms of forecasting the market trends. In addition, the model was found to be resistant to sharp market volatility, where standalone models performed poorly. The integration of LIME also showed that the closing price and trading volume were the most significant features, confirming the feasibility of interpretation and logic on the predictions. Collectively, this study presents an episodic Interpretable deep learning model of financial forecasting, delivering superior decision assistance to investors and clearing a path to the present-time usage of transparent AI in the financial ministries.
