MULTISCALE DEEP LEARNING APPROACH FOR STOCK MARKET INDEX FORECASTING

Predicting stock price indices remains one of the most complex challenges in financial research due to its profound implications for investment decision-making.Despite numerous studies, many existing models fail to adequately capture the inherent nonlinearity, long-term dependencies, and high noise levels in stock index time series data, thereby limiting their predictive accuracy.To address these limitations, this Study introduces a novel multiscale deep learning-based prediction framework, ICEEMDAN-RF-SSA-LSTM.The proposed model employs Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) to denoise and decompose stock price index data into distinct high-and low-frequency components.Subsequently, Random Forest (RF) is used to identify key influencing factors from a comprehensive feature set.To optimize the model's predictive capability, the Sparrow Search Algorithm (SSA) is applied for hyperparameter tuning of the Long Short-Term Memory (LSTM) network, which serves as the final prediction engine.The model is empirically evaluated using daily closing data of the Shanghai Stock Exchange Index (SSE) and the CSI 300 Index spanning from July 1, 2013, to June 30, 2022.Comparative analysis with individual models and alternative hybrid approaches demonstrates that the ICEEMDAN-RF-SSA-LSTM framework significantly outperforms existing methods in prediction accuracy and robustness.This Study offers a valuable methodological reference for developing more accurate and holistic predictive models for stock market indices and other financial instruments.