Enhancing Stock Market Prediction Using Attention-Based LSTM Models and Classical Models

Stock price prediction remains an elusive problem due to the non-linearity and dynamic nature of the financial markets. Statistical models such as the Auto-Regressive Integrated Moving Average (ARIMA) have been established to be widely applicable but tend to overlook short-term fluctuations and complex interdependencies between stock price movements. In this paper, we propose an attention-based Long Short-Term Memory (Attention-LSTM) model and contrast its prediction accuracy with ARIMA, optimised ARIMA, and simple LSTM across multiple trading datasets. Our findings depict that Attention-LSTM strictly performs the best by achieving the minimum Root Mean Squared Error (RMSE) against all other models in forecasting stock prices. Hyperparameter-tuned ARIMA is more precise than normal ARIMA but not as precise as deep learning models. LSTM can effectively capture temporal relationships and reduce errors in predictions, further enhancing forecast precision by dynamically weighing historical data through attention mechanisms by Attention-LSTM. Results corroborate the current literature on AI-based stock forecasting, confirming the superiority of deep learning models over statistical models. Future research will explore hybrid AI models, sentiment-aware prediction, and reinforcement learning-based trading regulations to further enhance stock price forecasting accuracy. The study confirms that attention-based deep learning models have a robust architecture for financial forecasting, resulting in AI-driven decision-making in stock market analysis.