Ensemble Deep Architecture Trained on Selective Feature set from Normalized Data for Stock Market Prediction

Stock market prediction refers to forecasting the future price of a company’s shares or other financial instruments traded on an exchange. Accurately predicting these prices offers the potential for significant profits. However, developing a deep learning (DL) model for stock market forecasting is a challenging task. This is primarily due to the inherent volatility and non-linear patterns in market behavior, along with the limitations of historical data, which often fail to capture all the influencing factors. Additionally, the task becomes more complex because external events and investor sentiment can significantly affect market dynamics, further complicating the modeling process. Hence, this research introduces a novel Stock Market Price Prediction (SMP) model that enhances forecasting accuracy through a multi-phase approach comprising preprocessing, feature extraction, feature selection, and prediction stages. Initially, an adaptive Z-score normalization technique is employed for pre-processing the input data, which effectively standardizes the data, mitigating the impact of different data scales and improving the model’s overall performance. Subsequently, a set of technical indicators-based features, including Average True Range (ATR), Rate of Change (RoC), Relative Strength Index (RSI), Average Directional Movement Index (ADM), William’s %R (WILLR), Money Flow Index (MFI), and Triple Exponential Average (TRIX), are extracted from the preprocessed data to capture a comprehensive view of the market’s behavior. Then, the improved wrapper-filter model in the feature selection phase helps address the challenge of high-dimensionality by identifying and retaining only the most relevant features, thus reducing overfitting and enhancing model efficiency. Finally, the prediction phase leverages an ensemble deep learning model composed of an improved Attention-based SqueezeNet (IA-SQN), Deep Convolutional Neural Networks (DCNN), and GhostNet classifiers for producing the final predicted outcomes. The integration of these models allows the method to harness the strengths of each, improving prediction robustness and accuracy. Furthermore, the suggested model is evaluated against the traditional methods. From the analysis, the proposed EC (IA-SQN+GSNet+DCNN) attained less Mean Absolute Error (MAE) of 0.466376, and Mean Squared Logarithmic Error (MSLE) of 0.004, which is exceptionally better than the traditional approaches. Therefore, the proposed method not only improves forecasting accuracy but also enhances decision-making and efficiency in trading operations.