Description
The paper proposes TMA, a novel Transformer-based multi-agent framework for dynamic portfolio optimization. TMA integrates diverse market perspectives through multiple functional agents and employs an attention mechanism to adaptively fuse heterogeneous information. It features an adaptive decision layer that dynamically adjusts risk preferences based on real-time market volatility. Empirical tests across major indices (STOXX 600, Hang Seng, S&P 500) demonstrate that TMA consistently outperforms traditional models and single-agent deep learning approaches in risk-adjusted returns, robustness, and stability. The framework advances intelligent finance by unifying multi-agent collaboration and attention-based deep learning into an adaptive portfolio system.