ENHANCING STOCK PRICE PREDICTION USING COMPLETE ENSEMBLE EMPIRICAL MODE DECOMPOSITION AND PRINCIPAL COMPONENT ANALYSIS ALGORITHMS-BASED FEATURE ENHANCEMENT

Abstract

Predicting stock prices is a complex task due to the nonlinear, nonstationary, and noisy characteristics of financial time series data. Traditional statistical and economic models often fail to capture the intricate and dynamic behavior of stock markets. To address these challenges, this study proposes three hybrid deep learning architectures that integrate advanced preprocessing and sequence modeling techniques. First, Complete Ensemble Empirical Mode Decomposition (CEEMD) is employed to denoise and decompose the financial time series into Intrinsic Mode Functions (IMFs). Then, Principal Component Analysis (PCA) is applied to extract the most significant features from the IMFs and reduce dimensionality. The transformed data is processed by a Convolutional Neural Network (CNN) to capture local patterns, and subsequently by either a Long Short-Term Memory (LSTM) network or a Bidirectional LSTM (BiLSTM) network to model temporal dependencies. The proposed architectures, including CEEMD-PCA-CNN-LSTM, CEEMD-PCA-CNN-BiLSTM, and CEEMD-PCA-BiLSTM are evaluated using four major stock indices: S&P 500, Dow Jones, DAX, and Nikkei 225. Results show that combining CEEMD and PCA significantly enhances predictive performance. The CEEMD-PCA-CNN-BiLSTM model outperforms others for the Dow Jones and Nikkei 225 datasets, achieving reductions in Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) by up to 93.7%, 92.7%, and 90.2%, respectively. The CEEMD-PCA-BiLSTM model yields the best results for the S&P 500 and DAX indices, with reductions in RMSE, MAE, and MAPE reaching up to 96.8%. These findings demonstrate the effectiveness of combining decomposition, feature selection, and deep learning for robust stock price prediction.