Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks
| Authors | Kunihiro Miyazaki et al. |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2602.23330 |
| Download | |
| Categories | cs.AI |
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.
Engineering Breakdown
Plain English
This paper presents a multi-agent LLM system for automated stock trading that breaks down investment analysis into fine-grained, explicit tasks rather than relying on abstract high-level instructions. The authors evaluated their framework on Japanese stock market data (prices, financial statements, news, macro indicators) using leakage-controlled backtesting to prevent look-ahead bias. The key finding is that decomposing trading workflows into specific, granular subtasks improves inference performance and decision transparency compared to coarse-grained agent instructions that mainstream multi-agent trading systems typically use. The paper addresses a real gap: existing systems focus on mimicking analyst and manager roles abstractly, but miss the concrete operational details that real investment teams follow.
Core Technical Contribution
The core novelty is replacing abstract role-based instructions (analyst→manager pipelines) with an explicit task decomposition layer that breaks investment analysis into fine-grained, sequenced subtasks. Rather than telling an agent "act as an analyst" or "make trading decisions," the framework defines specific computational steps: data retrieval, fundamental analysis, sentiment analysis, risk assessment, and position sizing as separate, measurable tasks with clear inputs and outputs. This task-oriented design enables better control flow, intermediate validation, and more interpretable decision paths through the agent pipeline. The insight is that LLMs perform better with concrete task specifications than with role abstractions, especially in high-stakes domains where transparency and accountability matter.
How It Works
The system ingests four data streams—stock prices, financial statements, news articles, and macroeconomic indicators—and routes them through specialized agent nodes, each responsible for a fine-grained analysis task. Each agent receives task-specific prompts and context windows designed for that particular computation (e.g., a fundamental analysis agent gets balance sheets and income statements; a sentiment agent gets news text). The agents produce intermediate outputs (scores, flags, assessments) that feed into downstream agents in a directed acyclic graph structure, allowing the manager agent to make final trading decisions based on aggregated, auditable reasoning from each component. The backtesting framework prevents data leakage by strictly separating training periods from test periods and ensuring agents cannot see future information. Output is a trading signal (buy/hold/sell) with a trace of reasoning from each task stage, enabling post-hoc analysis of why a trade was recommended.
Production Impact
For teams building automated trading systems, this approach directly addresses the interpretability and stability problems plaguing current multi-agent deployments: you get auditable decision chains where regulators and risk officers can see exactly which analysis component influenced each trade. Implementation would require building a task orchestration layer (DAG executor) atop your LLM inference pipeline, standardizing data feeds to multiple agents in parallel, and adding intermediate result validation—adding ~15-25% compute overhead but gaining transparency that makes compliance sign-off faster. The fine-grained structure also isolates failure modes: if the sentiment analysis component drifts, you can retune it independently rather than retraining the entire agent pipeline. However, you must invest in careful prompt engineering and validation for each subtask; this is not a black-box plug-and-play solution. Production deployment would require real-time data infrastructure, backtesting frameworks with strict leakage controls, and monitoring systems that track each agent's output distribution over time.
Limitations and When Not to Use This
The paper evaluates only on Japanese equities, leaving open questions about generalization to other asset classes (crypto, bonds, FX), markets with different liquidity profiles, or non-English financial text—multilingual fine-tuning and domain adaptation are not addressed. The abstract does not specify which LLM backbone was used (GPT-4, Llama, proprietary), so reproducibility and cost estimates are unclear; the approach may be sensitive to model choice in ways that aren't discussed. The leakage-controlled backtesting prevents data snooping, but the paper doesn't address concept drift in financial markets—agent performance likely degrades over time as market regimes change, and the static task decomposition may not adapt to structural breaks in correlations or volatility patterns. Finally, the framework assumes access to complete, timely financial data and news feeds; real production systems suffer from data quality issues, delays, and missing values that the paper's idealized setup likely does not encounter.
Research Context
This work builds on the recent wave of multi-agent LLM systems (ReAct, AutoGPT, multi-role agent teams) but specifically targets the financial domain where transparency and fine-grained decision-making are regulatory and practical requirements. It contrasts with and improves upon mainstream financial AI approaches that treat trading as a monolithic policy learned end-to-end or that use simple role templates (analyst-manager) without decomposing workflows into measurable subtasks. The research opens a direction for task-aware agent design: instead of scaling agent capability through scale (bigger models, more in-context examples), decompose the task explicitly and let specialized, smaller agents handle each piece. This connects to broader work on mechanistic interpretability and modular neural systems, suggesting that financial AI and possibly other high-stakes applications benefit from explicit task graphs over end-to-end learned policies.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::