Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks

Paper: arXiv 2602.23330

Authors: Kunihiro Miyazaki, Takanobu Kawahara, Stephen Roberts, Stefan Zohren

Abstract

The advancement of large language models (LLMs) has accelerated the development of autonomous financial trading systems. While mainstream approaches deploy multi-agent systems mimicking analyst and manager roles, they often rely on abstract instructions that overlook the intricacies of real-world workflows, which can lead to degraded inference performance and less transparent decision-making. Therefore, we propose a multi-agent LLM trading framework that explicitly decomposes investment analysis into fine-grained tasks, rather than providing coarse-grained instructions. We evaluate the proposed framework using Japanese stock data, including prices, financial statements, news, and macro information, under a leakage-controlled backtesting setting. Experimental results show that fine-grained task decomposition significantly improves risk-adjusted returns compared to conventional coarse-grained designs. Crucially, further analysis of intermediate agent outputs suggests that alignment between analytical outputs and downstream decision preferences is a critical driver of system performance. Moreover, we conduct standard portfolio optimization, exploiting low correlation with the stock index and the variance of each system’s output. This approach achieves superior performance. These findings contribute to the design of agent structure and task configuration when applying LLM agents to trading systems in practical settings.

Complexity vs Empirical Score

• Math Complexity: 2.0/10
• Empirical Rigor: 7.5/10
• Quadrant: Street Traders — practical and empirical, lighter on theory

Why this score: The paper provides a strong empirical framework for evaluating multi-agent LLM systems in trading through fine-grained task decomposition, explicitly accounting for look-ahead bias and backtesting on Japanese equities. While mathematical formulas and theoretical proofs are minimal, the empirical design, agent ablation studies, and qualitative analysis make it a highly practical and solid contribution.

Research Flowchart

  flowchart TD
    A["Research Goal: Improve LLM Trading Performance via Fine-Grained Task Decomposition"] --> B["Data Inputs: Japanese Stock Data<br><i>(Prices, Financials, News, Macro)</i>"]
    B --> C["Methodology: Multi-Agent LLM Framework<br><i>(Coarse vs. Fine-Grained Task Decomposition)</i>"]
    C --> D["Computational Process: Leakage-Controlled Backtesting & Output Variance Analysis"]
    D --> E["Portfolio Optimization: Exploiting Low Index Correlation & Model Variance"]
    E --> F["Key Findings: Superior Risk-Adjusted Returns & Importance of Agent-Decision Alignment"]

Cited References (74)

1. When Wording Steers the Evaluation: Framing Bias in LLM judges (2026) — link
2. Large Language Model Agents for Investment Management: Foundations, Benchmarks, and Research Frontiers (2025) — link
3. LLM Company Policies and Policy Implications in Software Organizations (2025) — link
4. QTMRL: An Agent for Quantitative Trading Decision-Making Based on Multi-Indicator Guided Reinforcement Learning — → 2508.20467
5. TradingGroup: A Multi-Agent Trading System with Self-Reflection and Data-Synthesis (2025) — link
6. AlphaAgents: Large Language Model based Multi-Agents for Equity Portfolio Constructions — → 2508.11152
7. Blueprint First, Model Second: A Framework for Deterministic LLM Workflow (2025) — link
8. Your AI, Not Your View: The Bias of LLMs in Investment Analysis — → 2507.20957

9 of 74 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2518 linked papers.