Comparative Analysis of Hybrid Feature Optimizaton on Deep Learning and Ensemble MachineLearning in Stock Price Prediction

Abstract

Data mining technologies that have given significant results and is used by researchers in a variety of fields is the neural network in the past. In today's economy, stock market data analysis and prediction play a crucial role. In this proposed study, the Apple Inc. (AAPL) stock data that is listed on NASDAQ stock exchange is considered and the stock’s day wise closing price has been predicted and analysed using two proposed systems. The Proposed System1 integrated key technical indicators with traditional retrieved dataset and uses a combination of SelectKBest-RF fusion method for important feature selection and apply the hyper tunned parameterised-data to train the five ensemble machine learning models- Random Forest, Gradient Boosting, XGBoost, Stacking, and Voting to forecast the stock closing price. In Proposed System2, various volume related, and other essential technical indicators are computed and used as features to improve performance. This work also focuses on reducing the complexity, so a hybrid feature optimization technique (C-R-L) combining Correlation Analysis, Recursive Feature Elimination (RFE) and L1 Regularization is proposed for relevant and optimum feature selection. To anticipate the company's stock price based on existing historical data and computed indicators, three different types of deep learning architectures were employed: Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Convolutional Neural Networks with LSTM (CNN-LSTM). It is observed that CNN-LSTM outperformed the rest of the two models and the five ensemble models. Moreover, comprehensive comparative analysis has been performed for the validation.