A Comparative Study of Deep Learning Architectures for Multivariate Financial Time Series Forecasting

Accurate forecasting of multivariate financial time series remains a critical challenge due to high volatility, non-stationarity, and complex cross-variable dependencies. Although deep learning models such as LSTM, GRU, TCN, and Transformer have shown notable progress, existing research often evaluates these architectures in isolation, lacks interpretability, and provides limited analysis of robustness across different markets. These limitations impede the deployment of reliable forecasting systems in practical financial settings. This study presents a comprehensive comparative analysis of representative deep learning architectures for financial forecasting and introduces a novel Hybrid Attention-Gated Module (HAGM). HAGM combines convolutional feature extraction, gated fusion, and multi-head self-attention mechanisms to efficiently capture both local and global dependencies. Experiments were conducted on stock indices, foreign exchange, and cryptocurrency datasets, assessing model performance across multiple forecast horizons. The results demonstrate that HAGM consistently outperforms baseline models, achieving lower RMSE and MAPE while exhibiting faster convergence. Ablation studies confirm the complementary contributions of convolution, gating, and attention components, and interpretability analyses identify critical variables such as trading volume and volatility. Robustness evaluations further reveal superior cross-market generalization and resilience under noisy conditions. Overall, this work advances the methodological understanding of deep learning approaches for financial forecasting and provides actionable insights for practitioners aiming to develop accurate, efficient, and interpretable predictive systems.