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

Abstract: The advancement of 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.

• The paper demonstrates that fine-grained prompt engineering in a multi-agent LLM system significantly improves out-of-sample Sharpe ratios across various portfolio sizes.
• The paper details a hierarchical framework where specialized agents, particularly the technical agent, drive risk-adjusted performance through precise financial and technical indicators.
• The paper shows that explicit task decomposition enhances interpretability and auditability, offering a robust system architecture adaptable to diverse financial markets.

Expert Investment Teams via Multi-Agent LLMs with Fine-Grained Trading Tasks

System Architecture and Task Decomposition

This work presents a hierarchical multi-agent LLM-based trading system, introducing explicit fine-grained task decomposition into the prompt design and decision workflow. Unlike prior approaches that allocate coarse-grained roles to agents, the proposed framework operationalizes investment analysis through concrete, low-level tasks that closely align with professional financial practices.

At the foundation, four specialist agents—Quantitative, Technical, News, and Qualitative—process heterogeneous company-level information. The Quantitative and Technical agents score securities based on structured financials and technical signals, while News and Qualitative agents extract latent risk/catalyst features from unstructured disclosures and news. Outputs are passed, aligned by sector, to a Sector Agent for benchmarking and adjustment. A Macro Agent independently processes macroeconomic signals. The Portfolio Manager (PM) synthesizes all inputs to allocate long/short positions on the TOPIX 100 universe. The system adopts strict look-ahead control to preclude data leakage and employs GPT-4o as the reasoning engine.

Figure 1: Architecture of the multi-agent hierarchical LLM trading system, illustrating agent roles and information flow.

Methodology: Task Granularity in Prompt Design

Fine-grained prompts for the Quantitative and Technical agents deliver predefined financial and technical indicators, such as normalized multi-scale momentum, volatility, classic oscillators (MACD, RSI, KDJ for Technical), and exhaustive profitability, valuation, leverage, and growth ratios for Quantitative. In contrast, coarse-grained prompts only provide high-level or raw unstructured data (e.g., 1-year price series without indicator computation or raw financial statements).

This explicit task decomposition is motivated by evidence that concrete procedural breakdowns enhance LLM reliability, reasoning, and interpretability, consistent with findings in non-financial multi-agent domains.

Empirical Results: Trading Performance and Ablation Analyses

Rigorous backtesting using the Japanese TOPIX 100 over a 27-month hold-out period demonstrates that architectures with fine-grained task specifications yield statistically significant improvements in out-of-sample Sharpe ratios compared to coarse-grained baselines. The advantage is persistent across portfolio sizes ($N = 20$ to $50$), with performance uplifts of 8 to 26 basis points in median Sharpe (Mann-Whitney $p<10^{-4}$ in most cases). The only exception arises with the smallest portfolios ($N=10$), where noise dominates.

Figure 2: Sharpe ratio distributions for fine-grained (pink) vs. coarse-grained (blue) task agents, with statistical significance scores.

Leave-one-out ablation reveals that the Technical Agent is the primary driver of the fine-grained architecture’s risk-adjusted returns. Removing the Technical Agent in a fine-grained setting drastically reduces Sharpe, especially for larger portfolios; removal of other agents (Quantitative, Qualitative, Macro, News) frequently results in negligible or even improved performance, suggesting redundancy or increased noise. This highlights that effective signal transmission is agent- and task-dependent rather than a mere function of component diversity.

Semantic Analysis and Interpretability

Textual analysis of agent rationale confirms that fine-grained prompts induce more detailed, domain-relevant reasoning (e.g., "momentum," "volatility," "soundness"), while coarse-grained setups revert to generic or superficial descriptors (e.g., "trend," "increase," "stability"). Vector similarity analysis of agent outputs using LLM embeddings shows that fine-grained tasking yields improved semantic alignment, particularly for Technical signals, between lower-level agents and the Sector Agent, evidencing enhanced interpretability and signal propagation in the decision hierarchy.

Portfolio Optimization and Practical Implications

Blending LLM-driven agent strategies with the market index, via risk parity allocation, yields further improvements in the net Sharpe ratio, with combined portfolios outperforming either constituent alone despite transaction costs. Notably, the agent composite exhibits only moderate correlation ($\approx 0.4$) with the underlying index, confirming substantial diversification potential.

Figure 3: Sharpe ratio as a function of allocation ratio between TOPIX 100 and the LLM agent composite; both gross and cost-adjusted statistics are shown.

These results suggest that, even with substantial agent heterogeneity and stochasticity, prompt-level task engineering meaningfully drives both systematic and idiosyncratic alpha sources.

Implications, Limitations, and Future Directions

The research demonstrates that for LLM-based financial multi-agent systems, architectural gains stem not only from advanced language reasoning but from task granularity—i.e., the explicit injection of robust, domain-specific feature engineering via prompts. This has direct implications for the practical deployment of LLM agents in finance: task decomposition should prioritize actionable, expert-designed subtasks rather than high-level roles or end-to-end completion.

Interpretability is enhanced, as the system admits detailed tracking of rationale through all agent layers, facilitating post-hoc auditability critical for institutional adoption. The analyses recommend agent frameworks built around workflow decomposition rather than strict data-modality role assignment.

However, model performance could be influenced by LLM-specific linguistic biases and temporal generalization limits, as backtesting horizons remain constrained by LLM knowledge cutoffs. Future work should address these via longer-horizon simulation using time-aware LLMs and further scrutiny of language-induced signal transmission. The architecture is extensible to other markets and alternative LLMs, but the findings are most robust within the Japanese equities context due to corpus and deployment constraints.

Conclusion

This study establishes that explicit fine-grained task decomposition in LLM-based trading agent prompts enhances both risk-adjusted performance and interpretability over traditional coarse role-based decomposition. Notably, technical analysis, when cleanly encapsulated in prompt-engineered agents, delivers substantial predictive value. The findings have immediate consequences for the design of systematic LLM agent architectures in finance, favoring expert task workflows over high-level abstractions to maximize real-world efficacy and auditability (2602.23330).