Paper: arXiv 2511.13614

Authors: Jerick Shi, Burton Hollifield

Abstract

Multi-strategy hedge funds face a fundamental organizational choice: should analysts generating trading strategies communicate, and if so, how? We investigate this using 5-agent LLM-based trading systems across 450 experiments spanning 21 months, comparing five organizational structures from isolated baseline to collaborative and competitive conversation. We show that communication improves performance, but optimal communication design depends on market characteristics. Competitive conversation excels in volatile technology stocks, while collaborative conversation dominates stable general stocks. Finance stocks resist all communication interventions. Surprisingly, all structures, including isolated agents, converge to similar strategy alignments, challenging assumptions that transparency causes harmful diversity loss. Performance differences stem from behavioral mechanisms: competitive agents focus on stock-level allocation while collaborative agents develop technical frameworks. Conversation quality scores show zero correlation with returns. These findings demonstrate that optimal communication design must match market volatility characteristics, and sophisticated discussions don’t guarantee better performance.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 8.0)

Why this score: The paper’s mathematical content is limited to alpha expression formulas and basic statistical measures, with no complex derivations or advanced theory, placing math complexity at the lower end. Empirical rigor is very high due to 450 experiments over 21 months, three market universes, multiple organizational structures, and detailed performance metrics (returns, Sharpe ratios, strategy correlations), making it highly backtest-ready.

Research Flowchart

  flowchart TD
  A["Research Goal: Optimal Communication<br>in Multi-Agent Trading"] --> B["Methodology: 5 LLM Agents<br>5 Org Structures × 21 Months"]
  B --> C["Data: 450 Experiments<br>Across 3 Stock Types"]
  C --> D["Computation: Simulate Trading<br>with Varying Conversation Modes"]
  D --> E{"Findings: Market-Dependent Outcomes"}
  E --> F["Vol Tech Stocks:<br>Competitive Wins"]
  E --> G["Stable General Stocks:<br>Collaborative Wins"]
  E --> H["Finance Stocks:<br>No Communication Benefit"]

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 empirically demonstrates that optimal communication structures for multi‑agent LLM trading depend on market volatility, and that all structures converge to similar strategy correlations.

Source sentence
Our contributions are: • Market-dependent communication effectiveness: Large-scale empirical demonstration that optimal communication structure varies systematically with market volatility and correlation characteristics, with competitive conversation excelling in volatile markets and collaborative conversation dominat

Assumes 1

  • Daily OHLCV data from Yahoo Finance are accurate and sufficient for alpha evaluation. data
    Source sentence
    We retrieve daily OHLCV data from Yahoo Finance for January 1, 2024 to September 30, 2025 (21 months) with monthly rebalancing.

Data

Backtest sample: 2024-01 – 2025-09 · US stocks (technology, general, finance universes) · equities · monthly

  • Yahoo Finance daily OHLCV · Yahoo Finance market-prices public
    2024-01-01 to 2025-09-30 · General Universe: SPY, JNJ, JPM, WMT, XOM, PG, UNH, HD, VZ, KO
    Source sentence
    We retrieve daily OHLCV data from Yahoo Finance for January 1, 2024 to September 30, 2025 (21 months) with monthly rebalancing.

Recognised sources: Yahoo Finance

Costs

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

Finds 5

Reported performance (LLM-based multi-agent trading with communication structures): Sharpe 2.47

  • Communication improves performance, with competitive conversation excelling in volatile technology stocks and collaborative conversation dominating stable general stocks; finance stocks show no benefit. performance
    Source sentence
    Communication improves performance in multi-agent trading systems, but optimal design depends on market characteristics. Competitive conversation excels in volatile technology stocks, while collaborative conversation dominates stable general stocks. Finance stocks resist all interventions.
  • All organizational structures converge to similar strategy correlations regardless of information sharing, eliminating diversity loss. theoretical
    Source sentence
    All organizational structures converge to similar strategy correlations regardless of information sharing, eliminating "diversity loss" as communication’s primary failure mode.
  • Conversation quality scores have zero correlation with returns; finance shows highest quality yet minimal performance, while tech shows strong returns with lowest scores. statistical
    Source sentence
    Conversation quality scores show zero correlation with returns: finance exhibits highest quality yet minimal performance, while tech achieves strong returns with lowest scores.
  • Competitive agents focus on tactical positioning while collaborative agents develop methodological sophistication. other
    Source sentence
    competitive agents focus on tactical positioning while collaborative agents develop methodological sophistication.
  • CORE (conversation quality) change is negatively correlated with performance improvement (r = -0.54). statistical
    Source sentence
    Correlation between CORE change and performance improvement yields r = -0.54.

Checked by others

Contradicted by

  • Social and individual learning in the Minority Game (2023)
    On the question: Does information sharing affect strategy diversity in multi‑agent market games?
    Compare the findings

    This paper: All organizational structures converge to similar strategy correlations regardless of information sharing, eliminating diversity loss.

    That paper: Social learning causes a crash in strategy‑table diversity, potentially locking the population into a maximally inefficient state.

    A reports a crash in diversity due to social learning, while B reports no diversity loss as all structures converge, contradicting A.

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

Cited References (14)

  1. TradingAgents : Multi-Agents LLM Financial Trading Framework — → 2412.20138
  2. MSPM : A modularized and scalable multi-agent reinforcement learning-based system for financial portfolio management (2021) — link
  3. Alpha-GPT 2.0: Human-in-the-Loop AI for quantitative investment — → 2402.09746
  4. Sentiment trading with large language models — → 2412.19245
  5. MAPS : Multi-agent reinforcement learning-based portfolio management system (2020) — link
  6. Multi-agent reinforcement learning in sequential social dilemmas (2017) — link
  7. QuantAgent : Seeking holy grail in trading by self-improving large language model — → 2402.03755
  8. Can LLM-based financial investing strategies outperform the market in Long Run? — → 2505.07078
and 6 more
  1. ContestTrade : A multi-agent trading system based on internal contest mechanism — → 2508.00554
  2. AlphaAgent : LLM-driven alpha mining with regularized exploration to counteract alpha decay (2025) — link
  3. A multimodal foundation agent for financial trading: Tool-augmented, diversified, and generalist — → 2402.18485
  4. FinMe : A performance-enhanced large language model trading agent with layered memory and character design — → 2311.13743
  5. MarketSenseAI 2.0: Enhancing stock analysis through LLM agents — → 2502.00415
  6. Information sharing in financial markets (2025) — link

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