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

Battle Count: 69

Directly addresses LLM-based stock trading with order-level execution, regime-specific evaluation across bearish/sideways/bullish markets, risk-adjusted performance metrics, and practical prompt optimization for sequential trading decisions. The framework provides actionable insights for deploying LLM agents in trading environments, though it remains in simulation without live-trading validation.

Requires multi-agent architecture (Market Analyst, News Analyst, Fundamental Analyst, Central Trading Agent, Optimizer LLM), integration with StockSim execution environment, Adaptive-OPRO with rolling window evaluation and template separation, daily decision cadence management, and coordination across multiple LLM API calls. The prompt optimization loop with history tracking and template preservation adds significant engineering complexity.

StockSim environment and configuration files are referenced (configs/o4-mini-adaptive-opro-config.yaml), experiments run on MacBook Pro M3 Pro with 18GB RAM, models accessed via Amazon Bedrock and OpenAI APIs. However, only 3 runs per configuration are reported, and the full codebase availability is not explicitly confirmed. Market data sourced from Massive API.