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

In one sentence

Summary

The paper addresses the challenge of deploying LLMs as autonomous trading agents in sequential, noisy environments where rewards are delayed and information is heterogeneous. It proposes ATLAS, a unified multi-agent framework that integrates market data, news, and fundamentals through specialized analyst agents, feeding a Central Trading Agent that emits executable orders. The core innovation is Adaptive-OPRO, a prompt optimization technique that dynamically updates the agent's static instructions using real-time, stochastic feedback over rolling windows, while keeping the execution interface stable to ensure edit locality and prevent overfitting to transient noise. Experiments evaluate ATLAS across three market regimes (bearish-volatile, sideways, bullish) using seven LLM families (GPT, Claude, Llama, Qwen) and a custom StockSim environment. The study compares Adaptive-OPRO against fixed expert-engineered baselines and reflection-based feedback mechanisms. Metrics include ROI, Sharpe Ratio, Maximum Drawdown, Win Rate, and trade frequency. The design isolates the adaptation mechanism by running each configuration three times to account for LLM stochasticity, ensuring that performance gains are systematic rather than random. The evaluation covers both the decision policy and the prompt optimizer, testing cross-family transfer and model capacity effects. Results show that Adaptive-OPRO consistently improves performance over static prompts, particularly in volatile regimes, with GPT-o3 and GPT-o4-mini achieving significant positive returns where baselines failed. In contrast, reflection-based feedback often deteriorates performance, suggesting it amplifies stochasticity in high-noise environments. The study finds that increased information availability does not uniformly improve performance, highlighting the importance of careful modality integration. Limitations include the abstraction of market microstructure (no slippage or latency) and the potential for over-optimization if the evaluation window is too short. The work demonstrates that systematic prompt evolution can yield interpretable, actionable trading constraints.

Research problem

The paper tackles the difficulty of deploying LLMs as autonomous trading agents in sequential decision-making settings. Key challenges include synthesizing heterogeneous information streams (market data, news, fundamentals) into coherent actions and adapting agent behavior under delayed, noisy reward signals. Static prompts fail to adapt, while existing reflection methods may disrupt effective policies in noisy environments.

Core contributions

• Proposes ATLAS, a multi-agent framework integrating specialized analysts for market, news, and fundamental data into a Central Trading Agent.
• Introduces Adaptive-OPRO, a novel sequential prompt optimization method that updates static instructions using rolling window feedback while preserving execution interfaces.
• Demonstrates that Adaptive-OPRO consistently outperforms fixed prompts and reflection-based feedback across multiple LLM families and market regimes.
• Provides empirical evidence that reflection-based feedback can actively degrade performance in high-noise environments, a phenomenon termed the 'reflection paradox'.
• Reveals that additional information modalities do not uniformly improve performance, emphasizing the need for careful integration in noisy markets.

Method

• Market Intelligence Pipeline with specialized agents for technical, news, and fundamental analysis.
• Central Trading Agent emitting executable orders (type, side, size, price) to StockSim.
• Adaptive-OPRO maintains a static instruction prompt and dynamic runtime content, updating only the static block.
• Rolling evaluation windows (K=5 days) compute scalar performance scores to reduce noise and address delayed feedback.
• Optimizer LLM generates prompt updates based on history, meta-prompt, and recent outcomes, constrained by template preservation.
• Reflection baseline uses weekly trajectory summaries to update agent behavior, serving as a comparison for sequential feedback.

Results & findings

• Adaptive-OPRO consistently outperforms fixed prompts and reflection across all tested models and regimes.
• GPT-o3 and GPT-o4-mini shifted from negative baseline returns to strong positive performance under Adaptive-OPRO in bearish regimes.
• Reflection-based feedback showed a negative correlation with baseline performance, often deteriorating results in volatile markets.
• GPT-o3 with Adaptive-OPRO achieved a 72.81% win rate and 0.146 Sharpe Ratio in the bearish regime.
• Qwen3-235B moved from losses to gains under Adaptive-OPRO, demonstrating cross-family robustness.
• Increased information availability did not guarantee better performance, highlighting the risk of noise amplification.
• Adaptive-OPRO improvements were largely independent of initial baseline performance, suggesting genuine adaptation rather than amplification.

Data & experiments

• Evaluated across three distinct market regimes: bearish-volatile, sideways, and bullish, spanning two-month windows.
• Tested on seven LLM families including GPT-o3, GPT-o4-mini, Claude Sonnet 4, Llama 3.3-70B, and Qwen3 variants.
• Used StockSim, a custom trading simulator enforcing order-level execution semantics (type, size, timing, price).
• Compared against non-LLM baselines (Buy & Hold, MACD, SMA, SLMA, Bollinger Bands) and LLM baselines (Fixed Prompt, Reflection).
• Metrics included ROI, Sharpe Ratio, Maximum Drawdown, Win Rate, and Number of Trades.
• Each configuration was run three times to account for LLM stochasticity and report mean ± standard deviation.

Limitations & caveats

• Market microstructure aspects like latency, slippage, and market impact are abstracted away, potentially overestimating real-world performance.
• The evaluation horizon is short (two months), which may not capture long-term regime shifts or structural breaks.
• Adaptive-OPRO's effectiveness depends on the quality of the scalar scoring function, which may not fully capture complex risk preferences.
• The study assumes deterministic execution within the simulator, ignoring partial fills and order book dynamics.
• Reflection-based feedback's negative impact may be specific to the tested regimes and model families, requiring broader validation.
• Prompt optimization is limited to static instruction edits, potentially restricting the agent's ability to adapt its information processing pipeline.

Curator notes

Discovery status: trading_agent_focus_promoted. Source links: FinMem: A Performance-Enhanced Large Language Model Trading Agent with Layered Memory and Character Design