The List / Agents and Multi-Agent Systems / paper
Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks
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.
In one sentence
Summary
The paper addresses the performance degradation and lack of interpretability in existing multi-agent LLM trading systems, which often rely on abstract, coarse-grained instructions that fail to mimic real-world analyst workflows. The authors propose a hierarchical framework that explicitly decomposes investment analysis into fine-grained tasks, assigning specific, concrete analytical protocols to specialized agents rather than vague high-level objectives. This approach aims to enhance both the operational performance and the transparency of decision-making processes in automated trading environments. The methodology employs a seven-agent system evaluated on the TOPIX 100 from September 2023 to November 2025, using a leakage-controlled backtesting setting with GPT-4o. The architecture includes four Level-1 analyst agents (Technical, Quantitative, Qualitative, and News) that generate scores and rationales, a Level-2 Sector Agent for relative valuation, a Macro Agent for economic regime assessment, and a Level-3 Portfolio Manager for final long-short construction. The study compares fine-grained prompts, which provide pre-calculated indicators and structured metrics, against coarse-grained prompts that feed raw data, while also conducting ablation studies to isolate agent contributions. Results indicate that fine-grained task decomposition significantly improves Sharpe ratios across most portfolio sizes, with the Technical Agent playing a critical role in driving this advantage. Text analysis reveals that fine-grained prompts produce more specific, domain-relevant vocabulary and better semantic alignment between intermediate outputs and final decisions. Furthermore, combining the agent-generated signals with the market index via portfolio optimization yields superior risk-adjusted returns. However, the study notes that removing certain agents can sometimes improve performance, suggesting that fine-grained coordination is essential to prevent noise introduction from redundant or conflicting signals.
Research problem
Existing multi-agent LLM trading systems often suffer from degraded inference performance and poor interpretability due to the use of abstract, coarse-grained instructions that do not reflect the intricate workflows of professional investment analysts. The paper seeks to determine if explicitly decomposing investment analysis into fine-grained, expert-defined tasks can improve both the risk-adjusted returns and the transparency of autonomous trading agents.
Core contributions
- Proposes a hierarchical multi-agent LLM framework that decomposes investment analysis into fine-grained, expert-defined tasks rather than using coarse-grained instructions.
- Demonstrates through controlled backtesting that fine-grained task decomposition significantly improves risk-adjusted returns compared to conventional coarse-grained designs.
- Provides an ablation study revealing that the Technical Agent is the primary driver of performance gains in fine-grained settings, while other agents may introduce noise.
- Analyzes intermediate textual outputs to show that fine-grained prompts enhance interpretability and semantic alignment between analytical outputs and downstream decisions.
- Validates real-world applicability by showing that combining agent signals with the market index via portfolio optimization achieves superior performance.
Method
- Hierarchical manager-analyst architecture with Level-1 analysts, Level-2 sector/macro agents, and a Level-3 portfolio manager.
- Fine-grained prompts provide pre-calculated technical indicators (RSI, MACD, Bollinger Bands) and structured financial metrics (ROE, P/E, etc.).
- Coarse-grained prompts provide raw financial statement data and daily prices for the LLM to process.
- Sector Agent adjusts quantitative scores based on sector averages to identify leaders and laggards.
- Macro Agent evaluates economic indicators (rates, inflation, growth) to assess market regime.
- Portfolio Manager integrates top-down macro views with bottom-up stock scores for final long-short selection.
- Backtesting uses monthly rebalancing with execution at the next month's open to prevent look-ahead bias.
- Textual interpretability is measured via log-odds ratios of word usage and cosine similarity of agent output embeddings.
Results & findings
- Fine-grained task decomposition significantly outperforms coarse-grained designs in 4 out of 5 tested portfolio sizes (20-50 stocks) with high statistical significance.
- The Technical Agent is identified as the most critical component for the success of fine-grained tasks; removing it reverses the performance advantage.
- Ablation studies show that removing Quantitative, Qualitative, or News agents often improves performance in fine-grained settings, suggesting they may introduce noise.
- Text analysis shows fine-grained outputs use more specific financial terminology (e.g., 'Momentum', 'Soundness') compared to coarse-grained outputs.
- Cosine similarity analysis indicates better semantic alignment between intermediate agent outputs and final decisions in the fine-grained framework.
- Combining agent strategies with the market index via portfolio optimization yields a Sharpe ratio of 1.91, outperforming both the index (1.68) and standalone agent strategies (0.95).
Data & experiments
- Backtesting on TOPIX 100 constituents from September 2023 to November 2025 to ensure no data leakage relative to GPT-4o's knowledge cutoff.
- Market-neutral long-short portfolio construction with monthly rebalancing, comparing fine-grained vs. coarse-grained prompt settings.
- Ablation studies removing individual agents (Technical, Quantitative, Qualitative, News, Macro) to assess their specific contributions to system performance.
- Textual analysis using log-odds ratios and cosine similarity of embeddings to evaluate interpretability and information propagation across the agent hierarchy.
- Portfolio optimization combining agent-generated signals with the TOPIX 100 index to demonstrate enhanced risk-adjusted returns.
Limitations & caveats
- The study is limited to the Japanese equity market (TOPIX 100), which may limit generalizability to other markets.
- Removing certain agents (e.g., Quantitative) sometimes improves performance, suggesting that fine-grained coordination is complex and not all agents are beneficial.
- The backtesting period is relatively short (27 months), which may not capture long-term market dynamics.
- The system relies on GPT-4o, and results may vary with different LLM architectures or versions.
- The study does not account for transaction costs in the primary comparison, though they are included in the portfolio optimization analysis.
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