Deep Learning-Based Stock Market Prediction: A Review of Existing Models, Performance, and Research Gaps

Abstract Prediction of the stock market is still a difficult problem in research due to the large dependence of financial time-series on the rapidly changing economic and behavioural factors, the high volatility and non-stationarity, the structural changes, the nonlinear behaviour and the presence of noise. However, (DL) has become an important computational technique for modelling these complex characteristics because of its ability to learn the hierarchical structure and temporal dynamics from massive financial data. This review critically analyzes the key deep learning methods applied to stock market prediction, such as recurrent neural networks, long short-term memory networks, gated recurrent units, convolutional neural networks, attention-based networks, Transformer models, generative models and hybrid models. Special focus is given to the change in the price-based approach to prediction to multimodal approaches that include technical indicators, fundamental information, financial news, investor sentiment, macroeconomic variables, and alternative data. Review additional evaluation methods, data, prediction time horizons and performance metrics. The advantages of hybrid and attention-enhanced models over traditional standalone models have been reported in the literature on multiple datasets, with reported advantages that are very sensitive to the datasets, data preprocessing, validation procedures, and market conditions. Standardized benchmarking, explainability, economic relevance, regime shifts, real-time deployment, reproducibility and integration of heterogeneous data are all major gaps in the field of research. Thus, adaptive, interpretable, multimodal, computationally efficient, and economically validated, deep-learning frameworks are needed in future studies.