Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks
Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks
This paper introduces a multi-agent LLM trading system that decomposes investment analysis into fine-grained tasks mimicking real-world analyst workflows, contrasting with typical coarse-grained role assignments. The framework includes seven agents: Technical, Quantitative, Qualitative, News, Sector, Macro, and Portfolio Manager, each receiving detailed prompts with pre-calculated financial metrics or raw data. Backtesting on Japanese TOPIX 100 stocks from September 2023 to November 2025 shows that fine-grained task decomposition significantly improves Sharpe ratios across portfolio sizes (20–50 stocks) compared to coarse-grained instructions. Ablation studies reveal the Technical agent as the primary performance driver, while textual analysis using log-odds ratios and embedding similarities demonstrates that fine-grained prompts enhance the propagation of technical insights to higher-level decision-makers. Additionally, portfolio optimization by blending multiple agent strategies with the TOPIX 100 index yields superior risk-adjusted returns due to low correlation, highlighting practical applicability. The study underscores that prompt design quality—specifically feature engineering embedded in tasks—is critical for LLM-based trading agent performance.
Highlights
- 1Proposes a multi-agent LLM trading framework with fine-grained task decomposition based on real-world investment workflows.
- 2Demonstrates that fine-grained tasks significantly improve risk-adjusted returns compared to coarse-grained instructions in Japanese stock backtesting.
- 3Reveals that the Technical agent is the primary performance driver, with fine-grained design enhancing its signal propagation to higher-level agents.
- 4Conducts comprehensive ablation studies and textual analysis to show that task granularity improves interpretability and information flow.
- 5Validates real-world applicability via portfolio optimization, achieving superior Sharpe ratios by blending agent strategies with a market index.
Methods
- MMulti-agent hierarchical framework with seven specialized agents (Technical, Quantitative, Qualitative, News, Sector, Macro, Portfolio Manager) using fine-grained vs. coarse-grained prompts.
- MBacktesting on TOPIX 100 constituents (Sept 2023–Nov 2025) with market-neutral long-short portfolio, monthly rebalancing, and strict look-ahead bias mitigation.
- MTextual analysis using log-odds ratios and cosine similarity of LLM embeddings to measure vocabulary shifts and information propagation between agents.
- MPortfolio optimization via equal risk contribution weighting across agent strategies and blending with TOPIX 100 index.
Results
- RFine-grained tasks yield statistically higher Sharpe ratios than coarse-grained tasks across portfolio sizes (20–50 stocks), with significance at p<0.05 or better.
- RAblation studies show the Technical agent is crucial; removing it degrades performance significantly in fine-grained settings, while other agents often add noise.
- RText analysis reveals fine-grained prompts produce more nuanced analytical vocabulary (e.g., 'momentum', 'volatility') and improve semantic similarity between Technical and Sector agents.
- RPortfolio optimization combining agent strategies with TOPIX 100 achieves Sharpe ratios superior to either component alone, with low correlation (~0.4) providing diversification benefits.
- RLeave-one-out experiments confirm fine-grained design robustly outperforms coarse-grained across most agent removal scenarios.
Analyze Paper
Generate insights from "Toward Expert Investment Teams:A Multi-Agent LLM System with...".