Paper: arXiv 2508.00554

Authors: Li Zhao, Rui Sun, Zuoyou Jiang, Bo Yang, Yuxiao Bai, Mengting Chen, Xinyang Wang, Jing Li, Zuo Bai

Abstract

In financial trading, large language model (LLM)-based agents demonstrate significant potential. However, the high sensitivity to market noise undermines the performance of LLM-based trading systems. To address this limitation, we propose a novel multi-agent system featuring an internal competitive mechanism inspired by modern corporate management structures. The system consists of two specialized teams: (1) Data Team - responsible for processing and condensing massive market data into diversified text factors, ensuring they fit the model’s constrained context. (2) Research Team - tasked with making parallelized multipath trading decisions based on deep research methods. The core innovation lies in implementing a real-time evaluation and ranking mechanism within each team, driven by authentic market feedback. Each agent’s performance undergoes continuous scoring and ranking, with only outputs from top-performing agents being adopted. The design enables the system to adaptively adjust to dynamic environment, enhances robustness against market noise and ultimately delivers superior trading performance. Experimental results demonstrate that our proposed system significantly outperforms prevailing multi-agent systems and traditional quantitative investment methods across diverse evaluation metrics. ContestTrade is open-sourced on GitHub at https://github.com/FinStep-AI/ContestTrade.

Complexity vs Empirical Score

  • Math Complexity: 3.0/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs only partially modelled; reported Sharpe ratio of 3.12 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 8.0)

Why this score: The paper focuses on a novel architectural design (multi-agent system with contest mechanism) and empirical validation, with the math involved being at a systems-engineering level rather than advanced mathematical finance. The empirical rigor is high due to the reported backtests, open-source code availability on GitHub, and focus on real market data processing and performance metrics.

Research Flowchart

  flowchart TD
  A["Research Goal: LLM Agents in Trading"] --> B{"Problem: High Sensitivity to Market Noise"};
  B --> C["Proposed System: ContestTrade Multi-Agent"];
  C --> D["Key Innovation: Internal Contest Mechanism"];
  D --> E["Data Team: Process Market Data -> Text Factors"];
  D --> F["Research Team: Parallel Trading Decisions"];
  E --> G["Real-time Scoring & Ranking via Market Feedback"];
  F --> G;
  G --> H["Outcome: Top Agent Outputs Adopted"];
  H --> I["Final Results: Outperforms SOTA & Traditional Methods"];

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 ContestTrade, a novel multi‑agent trading system with internal competitive mechanisms and a Deep Research methodology.

Source sentence
To address these limitations, we propose ContestTrade (A Multi-Agent Trading System Based on Internal Contest Mechanism).

Assumes 5

  • Factor performance exhibits short‑term momentum. market
    Source sentence
    We hypothesize that factor performance exhibits short‑term momentum.
  • Optimal short‑term windows are m = 5 days and n = 3 days for factor prediction. statistical
    Source sentence
    with optimal values found at m = 5, n = 3
  • Testing period (January‑June 2025) is after the LLM knowledge cutoff, eliminating data leakage. data
    Source sentence
    The testing period (January-June 2025) is chosen to be entirely after the knowledge cutoff of our LLMs, eliminating potential leakage from their pre‑training data.
  • Agents can be evaluated using a hybrid score combining realized Sharpe Ratio over a trailing m‑day window and qualitative judgments from an LLM Judger Panel. model
    Source sentence
    Realized Performance: Standard quantitative metrics calculated over a trailing m‑day window (e.g., realized Sharpe Ratio).
Show 1 more
  • LightGBM is an appropriate model for mapping historical hybrid scores to future utility. model
    Source sentence
    we employ LightGBM for this task, reusing the same simple and robust setup.

Data

Backtest sample: 2025-01 – 2025-06 · multi-asset

  • real‑world financial dataset (news, corporate financials, market data) text-news unknown
    training: July‑December 2024; testing: J
    Source sentence
    Our experiments utilize a real‑world financial dataset encompassing news, corporate financials, and market data.

Costs

Partial Costs appear only as a robustness check, a flat fee without spread or impact, or outside the headline numbers.

Finds 3

Reported performance (ContestTrade): Sharpe 3.12 · total return 52.8% · max drawdown 12.4%

  • ContestTrade achieves a cumulative return of 52.80%, Sharpe ratio of 3.12, and maximum drawdown of 12.41%, outperforming all baselines. performance
    Source sentence
    It achieves a Cumulative Return (CR) of 52.80%, a Sharpe Ratio (SR) of 3.12, and a Maximum Drawdown (MDD) of only 12.41%.
  • Ablation shows removing the LLM Judge reduces CR to 50.55%, SR to 2.57, and MDD to 13.48%; removing the contest in the Research team drops CR to 32.83%, SR to 1.78, MDD to 16.70%. performance
    Source sentence
    w/o LLM Judge 50.55 2.57 13.48 ... w/o Contest - Researcher 32.83 1.78 16.70
  • Internal contest mechanisms yield high Rank IC and ICIR, indicating strong predictive quality of selected factors and signals. predictive-accuracy
    Source sentence
    The consistently high scores in these metrics demonstrate that our mechanism effectively identifies high‑quality inputs, which in turn is the primary driver of the final portfolio’s outperformance.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 44 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (3 by independent authors)

Code

https://github.com/FinStep-AI/ContestTrade — the authors' own repository

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

Cited References (37)

  1. FinRL: deep reinforcement learning framework to automate trading in quantitative finance (2022) — link
  2. FinGPT: Open-Source Financial Large Language Models — → 2306.06031
  3. TradingGPT: Multi-Agent System with Layered Memory and Distinct Characters for Enhanced Financial Trading Performance — → 2309.03736
  4. Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models — → 2402.03659
  5. QuantAgent: Seeking Holy Grail in Trading by Self-Improving Large Language Model — → 2402.03755
  6. Alpha-GPT: Human-AI Interactive Alpha Mining for Quantitative Investment — → 2308.00016
  7. Alpha-GPT 2.0: Human-in-the-Loop AI for Quantitative Investment — → 2402.09746
  8. FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design — → 2311.13743
and 29 more
  1. A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist — → 2402.18485
  2. TradingAgents: Multi-Agents LLM Financial Trading Framework — → 2412.20138
  3. FINCON: a synthesized LLM multi-agent system with conceptual verbal reinforcement for enhanced financial decision making (2025)
  4. MASS: Multi-Agent Simulation Scaling for Portfolio Construction (2025) — link
  5. Generative Agents: Interactive Simulacra of Human Behavior (2023) — link
  6. React: Synergizing reasoning and acting in language models (2023)
  7. Toolformer: Language models can teach themselves to use tools (2024)
  8. DeepSeek-V3 Technical Report (2024) — link
  9. BloombergGPT: A Large Language Model for Finance (2023) — link
  10. Large Language Model Agent in Financial Trading: A Survey — → 2408.06361
  11. QWEN TECHNICAL REPORT (2023) — link
  12. FinBERT: a pre-trained financial language representation model for financial text mining (2021)
  13. Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models — → 2304.07619
  14. Can Large Language Models Beat Wall Street? Unveiling the Potential of AI in Stock Selection — → 2401.03737
  15. Unveiling the Potential of Sentiment: Can Large Language Models Predict Chinese Stock Price Movements? — → 2306.14222
  16. LLMF actor: Extracting Profitable Factors through Prompts for Explainable Stock Movement Prediction (2024) — link
  17. HAD: Heterogeneous multi-Agent framework for Financial sentiment analysis (2024) — link
  18. FinRobot: An Open-Source AI Agent Platform for Financial Applications using Large Language Models — → 2405.14767
  19. A Random Walk Down Wall Street (1973)
  20. Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of UK Inflation (1982) — link
  21. Agent-Based Computational Finance (2006)
  22. ABIDES: Towards High-Fidelity Multi-Agent Market Simulation (2020) — link
  23. NoLiMa: Long-Context Evaluation Beyond Literal Matching (2025) — link
  24. Context Rot: How LLMs Degrade with Context (2024)
  25. GSM-Infinite: How Do your LLMs Behave over Infinitely Increasing Reasoning Complexity and Context Length? (2025) — link
  26. Lost in the Middle: How Language Models Use Long Contexts (2024) — link
  27. Allocative Efficiency of Markets with Zero-Intelligence Traders: Market as a Partial Substitute for Individual Rationality (1993) — link
  28. Submodular Function Maximization (2014) — link
  29. Accelerated Greedy Algorithms for Maximizing Submodular Set Functions (1978) — link

14 of 37 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.