Fine-grained task decomposition makes LLM trading agents more profitable in backtests on Japanese stocks, improving risk-adjusted returns versus coarse-grained agent designs; improved alignment of intermediate analyses with decision rules appears to drive the gain.

Main Finding

Fine-grained decomposition of investment tasks into explicit, expert-style prompts for a multi-agent LLM trading system materially improves trading performance and interpretability relative to conventional coarse-grained agent instructions. The authors demonstrate this in a leakage-controlled backtest on TOPIX‑100 stocks (Sep 2023–Nov 2025): fine-grained task designs yield higher risk-adjusted returns (Sharpe ratios), and alignment of intermediate analytic outputs with downstream decision preferences is a key driver of performance. Portfolio optimization that exploits low correlation and output variance across the system further enhances results.

Key Points

• Problem: Most LLM multi-agent trading systems use coarse, high-level role instructions; this can reduce inference quality and obscure intermediate reasoning.
• Proposed approach: Explicitly decompose investment analysis into fine-grained, expert-like tasks across specialized agents rather than providing broad role prompts.
• Agent architecture: Hierarchical manager–analyst structure:
  • Level 1 analysts: Quantitative, Qualitative, News, Technical agents produce scores (0–100) and textual rationales.
  • Level 2: Sector agent adjusts analyst outputs by sector benchmarks; Macro agent assesses macro regime.
  • Level 3: Portfolio Manager (PM) synthesizes and selects long/short positions.
• Experimental focus: Technical and Quantitative agents compared under two prompt granularities (fine-grained vs coarse-grained), because they directly supply numerical signals for execution.
• Main empirical results:
  • Fine-grained task prompts significantly improve Sharpe ratios versus coarse-grained prompts.
  • Intermediate textual outputs align better with downstream scoring when tasks are fine-grained; this alignment correlates with portfolio performance.
  • Adding portfolio optimization that leverages low correlation between system outputs and index (and per-stock output variance) further improves returns.
• Ablation analysis: Systematically removing/replacing agents shows distinct functional contributions from agents and reveals which components are most valuable.
• Practical considerations: Reproducibility—the authors commit to releasing code and prompts upon acceptance.

Data & Methods

• Investment universe: TOPIX 100 (large-cap Japanese equities).
• Backtest design:
  • Period: Sep 2023 – Nov 2025 (27 months). Chosen to lie after the LLM (GPT‑4o) knowledge cutoff (Aug 2023) to avoid model training leakage.
  • Strategy: Market‑neutral long–short equal-weight portfolio. Rebalance monthly at the open of the first business day (signals formed after close of last business day).
  • Portfolio construction: Select equal numbers of highest‑scored longs and lowest‑scored shorts based on PM outputs.
• Models and inference:
  • LLM: GPT‑4o used as inference engine. Temperature set to 1. Outputs aggregated (median) across runs to mitigate stochasticity; authors argue temperature>0 preserves useful diversity.
  • Leakage control: Agents only receive data publicly available up to each decision point; backtest intentionally after model cutoff date.
• Data sources:
  • Price data: Yahoo Finance (daily closes used for signals; opens used for execution/performance).
  • Financial statements: EDINET API (FSA Japan) — numerical and textual filings.
  • News: Ceek.jp aggregator (Nikkei, Reuters, Bloomberg JP headlines/previews).
  • Macroeconomic indicators: FRED and Yahoo Finance (rates, CPI, yields, FX, indices, volatility).
• Fine-grained task details:
  • Technical Agent (fine-grained): Receives precomputed, normalized technical indicators (multi-horizon RoC, Bollinger Z‑scores, MACD, RSI, KDJ, etc.) and asked to produce 0–100 attractiveness scores plus rationale.
  • Technical Agent (coarse): Receives raw price time series (1 year) and a high-level instruction to analyze technicals.
  • Quantitative Agent (fine-grained): Receives computed financial ratios across five dimensions — Profitability (ROE, ROA, FCF margin), Safety (equity ratio, current ratio, D/E), Valuation (P/E, EV/EBITDA), Efficiency, and Growth (CAGR, EPS growth) — with TTM adjustments and YoY changes.
  • Quantitative Agent (coarse): Receives raw balance sheet / income statement / cash flow numbers and historical EPS points; asked to evaluate fundamentals.
• Evaluation metrics:
  • Quantitative: Monthly Sharpe ratio (mean monthly return / monthly return SD).
  • Qualitative: Analysis of agent scores and textual rationales to assess interpretability and alignment.
• Extra processing: Missing metric handling via NaNs passed to LLM; indicators normalized/ratio-adjusted to remove price-level bias. Portfolio optimization additionally leverages system output variance and correlation structure.

Implications for AI Economics

• Task design matters economically: The granularity of prompts is an economically relevant design choice. Fine-grained, SOP-like decompositions improve predictive and allocative performance of LLM agent teams, implying that human expert process encoding into AI systems can raise expected financial returns.
• Interpretability and auditability: Fine-grained tasks produce intermediate outputs that are more meaningful and inspectable. For institutional adoption and regulatory scrutiny in finance, structured intermediate reasoning improves transparency and may reduce operational risk.
• Alignment of objectives across layers: The finding that alignment between analytic outputs and downstream decision preferences drives performance highlights a form of “mechanism design” within multi-agent AI — agents’ scoring conventions and outputs must be economically aligned with portfolio objectives to realize value.
• Diversification role of LLM signals: System outputs can be low‑correlation signals relative to the market index; LLM-based teams may therefore provide genuine diversification and risk‑adjusted benefit when combined with portfolio optimization.
• Design blueprint for practitioners: The study suggests concrete implementation principles for production teams: hierarchical manager–analyst structures, encoded expert tasks (SOPs), careful normalization of inputs, ensemble aggregation, and leakage controls.
• Cautions and externalities:
  • Generalizability: Results are on large-cap Japanese equities with monthly rebalancing and a specific LLM (GPT‑4o); transferability to other markets, frequencies, or model families requires validation.
  • Data leakage and model knowledge: Strict temporal controls are necessary; otherwise backtests can be misleading. As LLMs evolve, training‑data leakage risks remain a major evaluation threat.
  • Market impact and equilibrium effects: If many agents adopt similar SOP-encoded strategies, returns may erode; systemic effects deserve study.
  • Operational frictions: The paper does not emphasize transaction costs, market impact, or slippage in depth — these will matter in deployment.
• Research directions: Quantifying marginal value of agent roles (via ablation), formalizing alignment metrics between intermediate outputs and final decisions, adapting SOPs dynamically, combining human oversight with agent teams, and evaluating longer time-series and different market microstructures.

If you want, I can (a) extract the exact prompt templates for the Technical/Quantitative agents from the paper appendix and summarize them, (b) sketch a toy implementation pipeline to reproduce the experiments, or (c) produce a short critique listing limitations and robustness checks the authors could add. Which would be most useful?