ATLAS: Adaptive Trading with LLM AgentS Through Dynamic Prompt Optimization and Multi-Agent Coordination

ACL ARR 2026 January Submission5813 Authors

Keywords: financial/business NLP, LLM agents

Abstract: Large language models (LLMs) offer promising capabilities for financial decision-making, yet their deployment in sequential trading settings faces two key challenges: synthesizing heterogeneous information sources and adapting agent behavior under delayed and noisy reward signals. We address these challenges by introducing ATLAS (Adaptive Trading with LLM AgentS), a unified agentic framework for systematic integration of market data, financial news, and corporate fundamentals, and Adaptive-OPRO, a novel prompt optimization method that dynamically updates agent instructions using real-time stochastic feedback. We evaluate our approach across regime-specific equity trading scenarios and multiple LLM families. Results demonstrate that Adaptive-OPRO consistently outperforms existing methods, particularly in highly volatile regimes. Moreover, our analysis reveals that increased information availability does not necessarily translate to improved performance, highlighting the importance of careful modality integration in noisy market environments.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: financial/business NLP

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 5813