Deep learning for financial forecasting: A review of recent trends

Deep Learning (DL) has revolutionized financial forecasting, yet most reviews remain purely descriptive and lack actionable insight. This paper presents a comprehensive review of 187 Scopus-indexed studies (2020–2024) on DL applications for financial forecasting. We examine key DL architectures—Deep Multilayer Perceptrons, Recurrent Neural Network variants (Long Short-Term Memory (LSTM), Bidirectional LSTM, Gated Recurrent Units, and Reservoir Computing), Convolutional Neural Networks, Temporal Convolutional Networks, and Autoencoders. The studies are organized by forecasting task—stock, index, forex, commodity, bond, cryptocurrency, and volatility—and by model type (standalone vs. hybrid), preprocessing strategies, multi-modal feature integration (technical, fundamental, and sentiment signals), and novel methodological contributions.Our findings confirm the dominance of RNN-based models, particularly LSTMs, while hybrid architectures combining convolutional and recurrent components (i.e., CNN-LSTM) are increasingly used to capture complex spatial–temporal dependencies. Stock and index forecasting remain prevalent, though cryptocurrency prediction is emerging due to continuous trading and high volatility. Integrating diverse signals (technical, fundamental, and sentiment) improves robustness but remains inconsistently applied.To move beyond brand-name-focused literature, we propose a design-principles taxonomy encompassing sequential-memory frameworks, spatial–temporal hybrids, attention-augmented ensembles, preprocessing-augmented pipelines (e.g., EMD, VMD, wavelets), chaos/quantum-inspired hybrids, and application-oriented architectures supporting algorithmic trading, robo-advisors, and portfolio optimization. This taxonomy highlights design rationale, under-explored intersections, and practical guidance for model selection.Critical gaps persist, including robustness under extreme market conditions, lack of standardized evaluation protocols, finance-specific interpretability, and computational constraints for real-time deployment. We outline a forward-looking research agenda emphasizing resilience, interpretability, scalability, lightweight deployment, modular hybrid models, behavioral insights, meta-learning, causal reasoning, emerging paradigms, such as chaos-inspired, and quantum-enhanced models, and real-time and continual learning. By framing DL in financial forecasting as a design-centered challenge rather than a black-box prediction task, this review aims to guide the development of adaptive, transparent, and high-performing financial systems.