Deep learning architectures for financial time-series forecasting: A comparative methodological review

<ns3:p> This paper presents a comparative review of selected deep learning architectures used in financial time-series forecasting, with particular emphasis on Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNNs), and attention-based hybrid models. The discussion is framed around three representative market environments: cryptocurrency, equity index, and foreign exchange data, in order to highlight how differences in volatility, liquidity, and market structure may affect model suitability. The paper outlines the main challenges of financial forecasting, including nonlinear dynamics, temporal heterogeneity, and sensitivity to changing market conditions. Particular attention is given to hybrid and attention-based approaches, especially the Regularized Self-Attention Regression (RSAR) model, as an example of a more adaptive forecasting framework. The review suggests that no single architecture should be regarded as universally optimal for all financial series. Instead, the suitability of a given model depends on the statistical properties of the analyzed data, the forecasting horizon, and practical constraints such as computational complexity and sensitivity to hyperparameter selection. </ns3:p>