Résumé

This paper introduces RaT-HyperGAT, a novel neural architecture designed to enhance Economic Intelligence and support data-driven decision-making in complex financial
environments. Modern financial markets, as dynamic information systems, exhibit intricate inter-asset dependencies and evolving market regimes that traditional forecasting methods fail to capture, limiting their usefulness for strategic EI applications.
To address these challenges, RaT-HyperGAT integrates three key innovations: (1) a multi-scale regime encoder that leverages macroeconomic indicators via LSTM networks with attention to extract relevant economic signals, (2) a temporal encoder based on bidirectional GRU networks to model sequential dynamics, and (3) a hypergraph attention mechanism that dynamically captures inter-asset relationships while adapting to changing marketregimes. Experimental results demonstrate that RaT-HyperGAT significantly outperforms benchmark models, achieving high predictive accuracy across bull, bear, and volatile market conditions (e.g., MSE of 0.000023 and MAE of 0.003828 in bull markets). By intelligently integrating multi-scale and relational information, RaT-HyperGAT strengthens Economic Intelligence capabilities, providing actionable insights for forecasting, strategic planning, and risk-aware decision-making in today’s uncertain and rapidly evolving global markets.