[CVPR 2024 Context] Engineering Alpha: Why Fine-Grained Tasks are the Key to Expert LLM Trading Teams

Executive Summary

In the race to build autonomous financial analysts, the bottleneck isn't the model's IQ—it's the clarity of its instructions. This paper presents a breakthrough multi-agent framework that moves away from "vague instructions" toward Fine-Grained Task Decomposition. By mirroring the precise workflows of institutional investment teams, the system achieves a significantly higher Sharpe Ratio on Japanese stock markets (TOPIX 100) compared to traditional agentic setups.

This work marks a transition from "Black Box Agents" to Structured Hierarchical Systems, proving that "how" you ask an agent to think is just as vital as "what" data you give it.

Problem & Motivation: The Danger of Coarse-Grained AI

Prior studies often treat LLMs as "magic boxes," giving them coarse-grained roles like "You are a fundamental analyst, analyze this 10-K report." The researchers identify two fatal flaws in this approach:

1. Performance Degradation: LLMs often abandon complex reasoning midway when instructions are too broad.
2. Interpretability Gap: Coarse instructions lead to surface-level outputs (e.g., "The price is rising"), making it impossible for human managers to trust the underlying logic.

To solve this, the authors propose an "Expert Process" model, where agents are given specific, pre-calculated metrics and standardized protocols.

Methodology: The Hierarchical IQ

The system is organized into a three-level hierarchy, designed to filter noise and amplify "Alpha" signals:

• Level 1 (The Specialists): Technical, Quantitative, Qualitative, and News agents analyze individual stocks. Crucially, the Fine-Grained Technical Agent is fed pre-calculated indicators like RSI and Bollinger Z-scores rather than raw price lists.
• Level 2 (The Coordinators): A Sector Agent benchmarks these scores against industry averages, while a Macro Agent assesses the broad environment (Inflation, Rates).
• Level 3 (The Decision Maker): A Portfolio Manager (PM) Agent synthesizes the top-down and bottom-up views to generate final weights.

Experiments: Does Granularity Pay Off?

The researchers conducted a head-to-head backtest comparing Fine-Grained (specific tasks) vs. Coarse-Grained (raw data/broad tasks).

1. Significant Performance Bump

As the portfolio size increases, the gap between the two methods widens. The Fine-Grained setting achieved a Sharpe Ratio of 0.90 for a 50-stock portfolio, significantly outperforming the coarse baseline.

2. The Power of Technical Analysis

Ablation studies (removing one agent at a time) showed that the Technical Agent was the "MVP." Without it, performance plummeted. Interestingly, the text analysis showed that in fine-grained settings, technical terms like "Momentum" and "Volatility" propagated much more effectively to the PM, influencing the final decision more than surface-level terms like "Price."

Deep Insight: "Alignment" is the Secret Sauce

By analyzing semantic similarities between agents, the authors found that the Sector Agent naturally aligns with fundamental data. However, the fine-grained framework forced the higher-level agents to actually listen to technical insights.

In coarse-grained settings, the LLM often ignores technical nuances because they are "lost in translation" when described in raw text. Structured tasks act as a bridge, ensuring that complex signals reach the final portfolio construction.

Conclusion & Future Outlook

The key takeaway for the industry is clear: Don't just give your agents roles; give them SOPs.

Takeaways:

• Prompt Engineering as Feature Engineering: Pre-calculating indicators is more effective than asking an LLM to calculate them on the fly.
• Hierarchical Trust: Standardized outputs make LLM trading interpretable and auditable—a prerequisite for institutional use.
• Limitations: The study is limited to a 2-year window (2023-2025). Future work must test these agents across different "market regimes" like deep recessions or hyper-inflation.

Final Verdict: This paper provides a blueprint for the next generation of LLM trading systems where the structure of thought is the ultimate competitive advantage.