Integrating Deep Learning Techniques for Enhanced Stock Price Prediction

The volatility and complicated structure of financial markets make it difficult to precisely predict stock prices. The multifaceted patterns and non-linear correlations seen in stock market data are sometimes challenging to capture using conventional methods. Therefore, there is an increasing need for investigation into innovative approaches that can overcome these limitations and improve accuracy in predictions. Previous studies predominantly focused on the utilization of single neural network-based Deep Learning (DL) models. While these models possess the capability to recognize immediate and prolonged correlations within time-series data, they often prove insufficient when addressing the complex intricacies of stock market dynamics. The field of hybrid DL models offers a significant opportunity to address the challenges that individual models face when predicting stock values. Merging models provides the ability to leverage the advantages of each architecture to create a more reliable and comprehensive prediction framework. The integrated networks in the proposed architecture provide a powerful ensemble of models, enabling a comprehensive analysis of the time-series dataset that captures the non-linear patterns embedded in stock market data. This approach offers dependable predictions amidst market noise, aiding better decision-making for analysts and investors. In this research, the assessment of various DL models for stock price prediction underscores a substantial enhancement in predictive accuracy and data fitting through the adoption of hybrid models over their single DL baseline counterparts. The Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) model is remarkable due to its significantly low Root Mean Squared Error (RMSE) of 0.012354 and an impressive coefficient of determination ( \(R^2\) ) score of 0.974957. This signifies a substantial performance enhancement in comparison to the baseline CNN and LSTM models. Similarly, the CNN-Bidirectional Long Short-Term Memory (BiLSTM) model surpasses the baseline BiLSTM, reducing the RMSE to 0.008575 and elevating the \(R^2\) score to 0.911205, indicating significant progress in predictive accuracy. The CNN-Gated Recurrent Unit (GRU) model also exhibits notable enhancements over the baseline GRU model, with an RMSE of 0.012663 and an \(R^2\) score of 0.973688. Furthermore, the hybrid models, including Bidirectional Gated Recurrent Unit (BiGRU) and BiLSTM, showcase outstanding predictive capabilities. The BiGRU-BiLSTM model, in particular, achieves an RMSE of 0.009001 and an impressively high \(R^2\) score of 0.9989122.