A Transformer Architecture for Learning Trading Strategies

Agent-based financial market simulations have advanced significantly through the integration of deep reinforcement learning (dRL), enabling adaptive trading agents such as the previously introduced trained response order network (TRON). Although TRON agents demonstrated improved market performance compared to traditional heuristic methods, their reliance on recurrent neural networks limits their ability to capture complex, long-range dependencies inherent in sequential financial data. In this paper, we introduce TRONformer, a novel trading agent architecture that combines transformer-based self-attention mechanisms with policy-gradient dRL. Our approach leverages the superior representational power of transformers to effectively model intricate market dynamics and dependencies. Empirical evaluations across multiple simulated financial market scenarios demonstrate that TRONformer agents consistently outperform previous TRON agents in profitability.