Enhanced Stock Market Forecasting: A Hybrid Multivariate Approach Incorporating CEEMDAN and Deep Recursive Neural Networks

Abstract Forecasting the future of the stock market is a crucial challenge for companies, financial analysts, policymakers, and other stakeholders. The primary objective is to develop a stable model that can effectively predict stock prices in the face of a volatile financial environment. However, financial time-series forecasting is inherently difficult due to its nonstationary, non-linear, and noisy nature. While various statistical and deep learning models are commonly used for stock price prediction, there is a noticeable lack of research on the effectiveness of different feature pre-processing methods and their combinations for diverse time series. To address this challenge, we propose a hybrid method that incorporates multiple perspectives to achieve optimal results. Our study focuses on financial time-series forecasting in the context of crude oil price fluctuations. The method involves several stages, including feature pre-processing, signal processing, and deep recursive neural networks. Anomalies, representing financial noises, are detected using an unsupervised Autoencoder. Additionally, a data enrichment phase refines the time series using CEEMDAN decomposition and Fourier Transform. Finally, a bi-directional LSTM model is utilized to extract latent patterns and forecast future trends in the time series. This approach aids in the progression of stock market forecasting, imparting valuable perspectives for decision-making among businesses, financial analysts, policymakers, and other individuals with a vested interest.