A Study of Stock Price Forecasting Using Improved LSTM-Based Models

Abstract

With the advancement of big data technology, machine learning has found widespread application in financial market forecasting due to its exceptional data processing and predictive capabilities. Upon reviewing recent research on stock prediction based on machine learning, it is found that under different market environments, Long Short-Term Memory significantly reduces prediction errors compared to other single-model methods. Therefore, to further explore the potential of LSTM models, this paper reviews recent research on stock price prediction based on improved LSTM models, focusing on the basic principles and characteristics of five models: BiGRU-LSTM、 VMD-CSSA-LSTM、 DMD-LSTM、 SF-GET-LSTM、 Doc-W-LSTM. Comparative analysis is conducted with other advanced models. Meanwhile, through comparative analysis of the literature, it is found that current research faces issues such as insufficient interpretability, limited research markets, and inadequate verification methods. It is proposed that introducing methods such as SHapley Additive exPlanations, causal inference, and time-series cross-validation may address these shortcomings. Finally, future research may endeavour to integrate the five models, constructing a multimodal adaptive and explainable intelligent financial forecasting system. This would provide theoretical and technical support for intelligent financial decision-making, driving further optimisation and innovation in deep learning within the forecasting domain.