Paper: arXiv 2510.15949

Authors: Charidimos Papadakis, Angeliki Dimitriou, Giorgos Filandrianos, Maria Lymperaiou, Konstantinos Thomas, Giorgos Stamou

Abstract

Large language models show promise for financial decision-making, yet deploying them as autonomous trading agents raises fundamental challenges: how to adapt instructions when rewards arrive late and obscured by market noise, how to synthesize heterogeneous information streams into coherent decisions, and how to bridge the gap between model outputs and executable market actions. We present ATLAS (Adaptive Trading with LLM AgentS), a unified multi-agent framework that integrates structured information from markets, news, and corporate fundamentals to support robust trading decisions. Within ATLAS, the central trading agent operates in an order-aware action space, ensuring that outputs correspond to executable market orders rather than abstract signals. The agent can incorporate feedback while trading using Adaptive-OPRO, a novel prompt-optimization technique that dynamically adapts the prompt by incorporating real-time, stochastic feedback, leading to increasing performance over time. Across regime-specific equity studies and multiple LLM families, Adaptive-OPRO consistently outperforms fixed prompts, while reflection-based feedback fails to provide systematic gains.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; reported Sharpe ratio of 6.63 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.0)

Why this score: The paper presents a novel multi-agent framework (ATLAS) with Adaptive-OPRO, relying on conceptual integration and prompt engineering rather than advanced mathematical derivations. It demonstrates high empirical rigor through regime-specific equity studies, backtests across multiple LLM families, and reported performance comparisons, though it lacks explicit code or dataset disclosures.

Research Flowchart

  flowchart TD
  A["Research Goal: Deploying LLMs as Autonomous Trading Agents"] --> B["Key Methodology: ATLAS Framework"]
  B --> C{"Core Components"}
  C --> D["Adaptive-OPRO<br>Dynamic Prompt Optimization"]
  C --> E["Order-Aware Action Space<br>Executable Market Orders"]
  C --> F["Multi-Agent Coordination"]
  
  G["Data Inputs: Market Data, News, Corporate Fundamentals"] --> B
  
  subgraph H ["Computational Processes"]
      D --> I["Feedback Loop: Real-time, Stochastic Rewards"]
      F --> J["Heterogeneous Information Synthesis"]
  end
  
  I --> K{"Key Findings"}
  J --> K
  E --> K
  
  K --> L["Adaptive-OPRO outperforms fixed prompts"]
  K --> M["Order-aware actions bridge model outputs to execution"]
  K --> N["Reflection-based feedback fails to provide systematic gains"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes ATLAS, a multi‑agent framework for LLM‑based stock market decision‑making, and introduces Adaptive‑OPRO, a prompt‑optimization mechanism for sequential trading.

Source sentence
we propose ATLAS, a multi‑agent framework that provides a foundational structure for experimentation in LLM‑based stock market decision‑making; (ii) we introduce Adaptive‑OPRO, a prompt optimization mechanism for sequential settings that supports behavioral adaptation through ongoing market interaction

Assumes 5

  • Agents make daily end‑of‑day decisions and receive delayed, noisy feedback. model
    Source sentence
    End‑of‑day decisions provide stable feedback for optimization under delayed, noisy outcomes, but prevent agents from reacting to intraday moves
  • The order‑level simulator abstracts away market microstructure such as slippage, partial fills, latency, and intraday dynamics. execution
    Source sentence
    Agents operate in an order‑level simulator that enforces trading semantics while abstracting market microstructure: slippage, partial fills, latency, and intraday dynamics are not modeled
  • Experiments are limited to three liquid equities over two‑month, regime‑specific windows. market
    Source sentence
    we focus on three liquid equities over two‑month, regime‑specific windows with daily decisions to reduce confounding from asset heterogeneity and shifting market structure
  • Each configuration is run three times due to resource constraints, limiting statistical power. statistical
    Source sentence
    Each configuration runs three times due to resource constraints, capturing stochastic variance but limiting statistical power
Show 1 more
  • Results are interpreted as behavioral evidence about Adaptive‑OPRO rather than market‑wide performance claims. other
    Source sentence
    Results should be read as behavioral evidence about Adaptive‑OPRO, not as market‑wide performance claims

Data

Backtest sample: equities · equities · daily

  • Massive API text-news unknown
    Source sentence
    as batches of timestamped items containing title, URL, summary, and keywords

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
Agents operate in an order‑level simulator that enforces trading semantics while abstracting market microstructure: slippage, partial fills, latency, and intraday dynamics are not modeled

Finds 5

Reported performance (ATLAS Adaptive-OPRO): Sharpe 6.63 · annual return -6.2% · max drawdown 14.1%

  • Adaptive‑OPRO consistently outperforms fixed prompts across multiple market regimes and LLM families. performance
    Source sentence
    Across regime‑specific equity studies and multiple LLM families, Adaptive‑OPRO consistently outperforms fixed prompts
  • Reflection‑based feedback fails to provide systematic gains and can be inconsistent across models. negative
    Source sentence
    while standard reflection proves inconsistent
  • Additional input modalities (e.g., news, fundamentals) are not uniformly beneficial and depend critically on market conditions. performance
    Source sentence
    additional input modalities are not uniformly beneficial and depend critically on market conditions
  • Adaptive‑OPRO improves risk‑adjusted metrics such as Sharpe Ratio, Sortino Ratio, ROIC, and profit per trade compared to baselines and reflection. performance
    Source sentence
    Adaptive‑OPRO reduces these gaps, yet reflection often increases variance without clear risk‑adjusted gains
  • Reflection can degrade performance, as shown in a case where it caused re‑entry into a weak market leading to a loss. negative
    Source sentence
    free‑form reflection can degrade performance by injecting prescriptive guidance that the agent follows even when market conditions do not justify action

Checked by others

Contradicted by

  • An Adaptive Multi Agent Bitcoin Trading System (2025)
    On the question: Does providing feedback to adaptive multi‑agent trading systems improve performance?
    Compare the findings

    This paper: Reflection‑based feedback fails to provide systematic gains and can be inconsistent across models.

    That paper: Adding weekly verbal feedback improved total performance by 31% and reduced bearish losses by 10%.

    A shows +31% total performance and -10% bearish losses with weekly verbal feedback; B finds reflection‑based feedback gives no systematic gains and is inconsistent across models.

Cited by 3 papers in the archive (3 by independent authors): QuantaAlpha: An Evolutionary Framework for LLM-Driven Alpha … (uses its method or data); Signal or Noise in Multi-Agent LLM-based Stock … (uses its method or data)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (50)

  1. Blackboard Systems (1986)
  2. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education (1983)
  3. Classification Problem Solving (1984)
  4. New Ways to Make Microcircuits Smaller (1980) — link
  5. New Ways to Make Microcircuits Smaller—Duplicate Entry (1980)
  6. Strategic explanations for a diagnostic consultation system (1984) — link
  7. Strategic Explanations in Consultation—Duplicate (1983)
  8. Poligon: A System for Parallel Problem Solving (1986)
and 42 more
  1. Transfer of Rule-Based Expertise through a Tutorial Dialogue (1979)
  2. The Engineering of Qualitative Models (2021)
  3. Attention Is All You Need (2017) — link
  4. Pluto: The ‘Other’ Red Planet (2015)
  5. INVESTORBENCH : A Benchmark for Financial Decision-Making Tasks with LLM -based Agent — → 2412.18174
  6. Walking a Tightrope – Evaluating Large Language Models in High-Risk Domains (2023) — link
  7. Breaking the Reasoning Barrier A Survey on LLM Complex Reasoning through the Lens of Self-Evolution (2025) — link
  8. Large Language Models as Optimizers (2024)
  9. Autoprompt: Eliciting knowledge from language models with automatically generated prompts (2020) — link
  10. Query-Dependent Prompt Evaluation and Optimization with Offline Inverse RL
  11. Connecting large language models with evolutionary algorithms yields powerful prompt optimizers (2023) — link
  12. C rypto T rade: A Reflective LLM -based Agent to Guide Zero-shot Cryptocurrency Trading (2024) — link
  13. StockSim: A Dual-Mode Order-Level Simulator for Evaluating Multi-Agent LLMs in Financial Markets (2025) — link
  14. TradingAgents: Multi-Agents LLM Financial Trading Framework — → 2412.20138
  15. FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design — → 2311.13743
  16. TradExpert: Revolutionizing Trading with Mixture of Expert LLMs — → 2411.00782
  17. MarketSenseAI 2.0: Enhancing Stock Analysis through LLM Agents — → 2502.00415
  18. FLAG-Trader: Fusion LLM-Agent with Gradient-based Reinforcement Learning for Financial Trading — → 2502.11433
  19. Large Language Models as Optimizers (2024) — link
  20. EvoPrompt: Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers (2025) — link
  21. GRAD-SUM: Leveraging Gradient Summarization for Optimal Prompt Engineering (2024) — link
  22. Large Language Models Prompting With Episodic Memory (2024) — link
  23. Enhanced Financial Sentiment Analysis and Trading Strategy Development Using Large Language Models (2024) — link
  24. SocioDojo: Building Lifelong Analytical Agents with Real-world Text and Time Series (2024) — link
  25. The Good, The Bad, and The Greedy: Evaluation of LLM s Should Not Ignore Non-Determinism (2025) — link
  26. Non-Determinism of “Deterministic” LLM Settings (2025) — link
  27. Non-linear prediction of security returns with moving average rules (1996)
  28. Predicting stock price trend using macd optimized by historical volatility (2018)
  29. The profitability of bollinger bands trading bitcoin futures (2023)
  30. Multi-agent design: Optimizing agents with better prompts and topologies (2025) — link
  31. Agentnet: Decentralized evolutionary coordination for llm-based multi-agent systems (2025) — link
  32. Dynamic llm-agent network: An llm-agent collaboration framework with agent team optimization (2023) — link
  33. FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making (2024) — link
  34. New Concepts in Technical Trading Systems (1978)
  35. Technical Analysis of the Financial Markets: A Comprehensive Guide to Trading Methods and Applications (1999)
  36. Trading Systems and Methods (2013)
  37. Technical Analysis from A to Z (2000)
  38. Financial Statement Analysis and Security Valuation (2012)
  39. Investment Valuation: Tools and Techniques for Determining the Value of any Asset (2012)
  40. Principles of Corporate Finance (2019)
  41. Business Analysis and Valuation: Using Financial Statements (2019)
  42. Further Evidence on Managerial Motives for Stock Splits (1993)

6 of 50 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.