A Hybrid Deep Learning Architecture for Multi-Task Financial Forecasting: Integrating Bi-LSTM and Hierarchical Attention

Abstract

This paper introduces and comprehensively evaluates a hybrid ensemble model for forecasting the Vietnamese VN-Index, achieving a significant R2 of 0.0980 and a directional accuracy (DA) of 67.03%. More than just a performance benchmark, this study dissects the model's architecture to understand the drivers of its success. We demonstrate its superiority against a wide range of baselines, from classical ARIMA to standard deep learning models. Through rigorous ablation studies, we provide definitive evidence that both the hierarchical attention mechanism and the multi-timeframe input structure are indispensable; their removal leads to a complete collapse in predictive performance, with F1-scores dropping to zero. Furthermore, an in-depth interpretability analysis reveals the distinct forecasting 'strategies' learned by specialized sub-models by visualizing their internal attention patterns. Our findings not only present a highperforming model but, more importantly, contribute to a deeper understanding of why and how such complex architectures succeed in volatile emerging markets.