Fine-grained task decomposition for LLM trading agents materially improves risk-adjusted returns in historical Japanese equity backtests; better semantic alignment between analysis steps and decision layers appears to be the mechanism.

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

Kunihiro Miyazaki, Takanobu Kawahara, Stephen Roberts, Stefan Zohren · June 16, 2026 · The Journal of Financial Data Science

openalex quasi_experimental medium evidence 7/10 relevance Full text usable extracted full text DOI Source PDF

Decomposing LLM-based trading analysis into fine-grained, domain-informed tasks significantly improves risk-adjusted returns in leakage-controlled backtests on Japanese equities, with semantic alignment between intermediate outputs and decision layers explaining much of the performance gain.

The advancement of large language models (LLMs) has accelerated the development of autonomous financial trading systems. While mainstream approaches deploy multi-agent systems with hierarchical decision architectures, they often rely on coarse-grained instructions that underspecify the analytical procedures, leading to degraded inference quality and reduced transparency. We propose a structured decision architecture that explicitly decomposes investment analysis into fine-grained, domain-informed tasks assigned to specialized inference modules, rather than providing abstract role-level 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, validated through bootstrap confidence intervals, multiple-testing corrections, and subperiod stability analysis, demonstrate 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 semantic alignment between analytical outputs and downstream decision layers is a critical driver of system performance, consistent with a structured regularization interpretation. Moreover, we conduct standard portfolio optimization, exploiting low correlation with the stock index and the variance of each system’s output, achieving superior risk-adjusted performance. These findings contribute to the design of structured inference architectures and task decomposition strategies for LLM-based financial decision systems.

Summary

Main Finding

A hierarchically structured multi-agent LLM trading system that decomposes investment analysis into fine-grained, domain-informed tasks substantially improves risk‑adjusted performance, interpretability, and stability versus conventional coarse-grained role prompts. In a leakage-controlled backtest on TOPIX‑100 stocks (Sep 2023–Nov 2025) the fine‑grained architecture produced materially better Sharpe/PSR outcomes, a higher probability of positive performance, lower output/hallucination variance, and reduced portfolio turnover.

Key Points

Architectural innovation: A bottom‑up hierarchy (Level 1 analyst agents → Level 2 Sector & Macro agents → Level 3 Portfolio Manager) assigns narrowly specified, expert-like subtasks to specialist agents (Quantitative, Qualitative, News, Technical).
Fine‑grained task decomposition acts as “structured regularization”: constraining agent reasoning paths reduces output variance and hallucinations compared to coarse, underspecified prompts.
Semantic alignment: Explicit intermediate outputs (scores and textual rationales) improve alignment between upstream analyses (notably technical analysis) and downstream decision layers, stabilizing decisions.
Empirical validation: Gains validated with bootstrap confidence intervals, multiple‑testing corrections, and subperiod stability checks rather than anecdotal performance.
Practical portfolio construction: Market‑neutral, equal‑weight long/short portfolios rebalanced monthly; ensemble median across stochastic LLM outputs used to mitigate extreme variability.
Ablation and diagnostics: Systematic removal of modules shows each component’s structural contribution; analysis of intermediate texts helps interpret failure modes and information flow.
Real‑world considerations: The system exploits low correlation between model signals and the market index; variance of model outputs is used in portfolio optimization to enhance risk‑adjusted returns.

Data & Methods

Investment universe: TOPIX 100 (large‑cap Japanese stocks).
Backtest period: Sep 2023 – Nov 2025 (27 months). To avoid knowledge leakage, experiments start after the model knowledge cutoff and agents only receive information available up to each decision date.
LLM & inference settings: GPT‑4o (knowledge cutoff Aug 2023), temperature = 1, ensemble median aggregation across outputs.
Inputs: stock prices, financial statements, news text, macroeconomic indicators.
Portfolio design: Monthly rebalancing at the first business‑day open; market‑neutral long/short equal‑weight portfolios to isolate stock selection skill.
Evaluation metrics: Sharpe ratio (primary), Sortino, maximum drawdown, Calmar, win rate, return skewness/kurtosis, probabilistic Sharpe ratio (PSR), turnover‑adjusted performance.
Statistical controls: Bootstrap confidence intervals, multiple‑testing corrections, subperiod analyses, and ablation studies of agent roles.
Theoretical framing: Formalized multi‑agent system as constrained policy representation; fine‑grained prompting interpreted as narrowing the effective policy class (structured regularization).

Implications for AI Economics

Design principle: Task granularity matters — encoding expert workflows into LLM agent prompts can act as an inductive bias that improves generalization and reduces model risk for automated trading systems.
Interpretability & governance: Producing structured intermediate outputs enables auditability and root‑cause analysis, aiding compliance and operational risk management in asset management.
Model aggregation & risk allocation: LLM output variance and correlation with benchmarks are useful inputs for portfolio construction and risk budgeting (e.g., using output variance to size positions or as an input to optimization).
System engineering tradeoffs: Fine‑grained, specialist modules increase control and stability but require careful domain engineering (expert prompts, module interfaces); this raises implementation costs and ongoing maintenance needs.
Evaluation standards: Leakage controls, robust statistical validation (bootstrap, multiple‑testing), ablation studies, and inspection of intermediate outputs should become standard practice when assessing LLM‑based trading strategies.
Market effects & scalability: If broadly adopted, more stable and less noisy LLM signals could reduce turnover and market impact per manager but may change cross‑sectional return dynamics; monitoring of cross‑strategy correlations will be important.
Caution on generalizability: Results are from TOPIX‑100 and a specific period; practitioners should test across markets, regimes, and longer horizons before deployment.

Assessment

Paper Typequasi_experimental Evidence Strengthmedium — The paper uses rigorous statistical tools (bootstrap CIs, multiple-testing correction, subperiod robustness) which strengthen confidence in the backtest findings, but evidence is limited to historical simulation rather than live deployment; risks of overfitting, unmodelled transaction costs/market impact, sensitivity to model/hyperparameter choices, and single-country data reduce causal certainty and external validity. Methods Rigormedium — The authors implement several good-practice checks (leakage control, bootstrap inference, multiple-testing correction, subperiod analysis) and analyze intermediate outputs to probe mechanisms, but the description lacks evidence of live out-of-sample trading, nested hyperparameter tuning, explicit modeling of transaction costs/liquidity/market impact, and details on sample size/time horizon and model selection procedures, which limits methodological rigor. SampleHistorical Japanese equities dataset including prices, financial statements, news articles, and macroeconomic variables; used in leakage-controlled backtests comparing agent architectures (exact time span, number of stocks, frequency, and preprocessing details not specified in summary). Themesinnovation productivity IdentificationComparative backtest experiment that evaluates two agent architectures (fine-grained, domain-informed task decomposition vs conventional coarse-grained multi-agent designs) on the same historical Japanese equity dataset; uses leakage-controlled backtesting to avoid look-ahead bias, reports bootstrap confidence intervals for performance metrics, applies multiple-testing corrections, and performs subperiod stability checks to assess robustness of effects across time. Portfolio-level optimization and low-correlation analysis are used to translate model outputs into traded returns and to compare risk-adjusted performance. GeneralizabilityLimited to Japanese equity market — may not generalize to other countries, asset classes, or market structures, Historical backtest setting — results may not hold in live trading due to market impact, slippage, and changing market regimes, Dependent on the specific LLM(s), prompts, and task-module implementations used; results may vary with model size, language, or training data, Unclear how robust findings are to different hyperparameter tuning, re-training frequency, and ensemble choices, Performance may rely on availability and language of news/financial disclosure (Japanese-language data), limiting transferability

Claims (8)

Claim	Direction	Outcome	Confidence & Evidence	Details
Fine-grained task decomposition significantly improves risk-adjusted returns compared to conventional coarse-grained designs. Decision Quality	positive	risk-adjusted returns	Reading fidelity high Study strength medium	not reported 0.48
A structured decision architecture that explicitly decomposes investment analysis into fine-grained, domain-informed tasks assigned to specialized inference modules (rather than abstract role-level instructions) improves inference quality and transparency. Decision Quality	positive	inference quality and transparency (as improved by the architecture)	Reading fidelity high Study strength medium	not reported 0.48
Experimental results were validated through bootstrap confidence intervals, multiple-testing corrections, and subperiod stability analysis. Decision Quality	null_result	statistical robustness of experimental results	Reading fidelity high Study strength medium	not reported 0.48
Semantic alignment between analytical outputs and downstream decision layers is a critical driver of system performance, consistent with a structured regularization interpretation. Decision Quality	positive	system performance (driven by semantic alignment)	Reading fidelity high Study strength medium	not reported 0.48
Standard portfolio optimization exploiting low correlation with the stock index and the variance of each system’s output achieves superior risk-adjusted performance. Decision Quality	positive	risk-adjusted portfolio performance	Reading fidelity high Study strength medium	not reported 0.48
Mainstream multi-agent hierarchical decision architectures often rely on coarse-grained instructions that underspecify analytical procedures, leading to degraded inference quality and reduced transparency. Decision Quality	negative	inference quality and transparency of mainstream architectures	Reading fidelity high Study strength low	not reported 0.24
The proposed structured inference architecture and task decomposition strategy contribute to the design of LLM-based financial decision systems. Other	positive	design contributions to LLM-based financial decision systems	Reading fidelity high Study strength low	not reported 0.24
The system outputs exhibit low correlation with the stock index (enabling diversification benefits used in portfolio optimization). Other	positive	correlation between system outputs and the stock index	Reading fidelity medium Study strength medium	not reported 0.29