Investigating explainability of deep learning models for sequential data on stock price prediction

Accurate stock price prediction remains a significant challenge within the financial sector, where deep learning models have shown considerable predictive power but suffer from limited interpretability due to their inherent ”black box” nature. This study addresses this issue by integrating Explainable artificial Intelligence (XAI) techniques with Conv1D and LSTM deep learning models to enhance transparency in stock price forecasting. We propose a novel XAI algorithm tailored for deep learning, utilizing post-hoc explanation methods such as LIME and SHAP to elucidate the decision-making processes of the models. In addition, we evaluate the predictive performance of Conv1D and LSTM models against state-of-the-art techniques, including NBEATS and LTCN, revealing that NBEATS consistently achieves superior performance across multiple datasets. Our findings highlight the predictive significance of stock price of Adani power equity from the previous two days of stock price, whereas data from two days prior shows reduced influence on predictive accuracy. The study demonstrates that XAI-augmented Conv1D and LSTM models strike a balance between transparency and predictive performance, offering investors more interpretable insights into market trends. This research underscores the importance of explainable stock price prediction and calls for continued exploration of this emerging domain.