# Toward Expert Investment Teams: A Multi-Agent LLM System with Fine-Grained Trading Tasks Source: https://www.youtube.com/watch?v=VR5TbyMdGVU Recap page: https://rapidrecap.app/video/VR5TbyMdGVU Generated: 2026-02-28T17:03:43.466+00:00 --- ## Quick Overview Researchers from Japan Digital Design and the University of Oxford developed a multi-agent LLM system that strictly mimics the daily operations of a real institutional investment team, achieving superior performance metrics like a Sharpe Ratio of 1.91 compared to the baseline of 1.38, primarily by using fine-grained instructions and rigorous testing methodologies like ablation studies. **Key Points:** - The multi-agent LLM system was designed to mimic the stratified structure of an expert investment team, featuring a Level 1 agent for information aggregation and a Level 2 'Adjustment Layer' agent. - The system uses GPT-4 as the underlying model and was trained using 252 days of raw closing prices from Japan's largest cap stocks (September 2023 to November 2025). - The fine-grained setup, which included specific instructions like setting the temperature parameter to one, yielded a superior Sharpe Ratio of 1.91, significantly outperforming the baseline of 1.38. - The Level 1 agent focuses solely on information aggregation (technical and macro data), while the Level 3 Portfolio Manager agent synthesizes outputs from lower levels to make the final decision. - A critical finding was that removing the Level 1 technical agent caused performance to plummet, proving that specialized agents are necessary and that simply throwing more compute at a large model is insufficient. - The system successfully avoided look-ahead bias by relying only on data available up to the simulation date and explicitly excluded data from future events during training. ![Screenshot at 00:00: The introductory screen features an image of two people podcasting with the text "BECOME A MEMBER TODAY!" overlayed, setting the stage for a discussion about AI research, which is visually framed by an oscilloscope-like grid.](https://ss.rapidrecap.app/screens/VR5TbyMdGVU/00-00-00.jpg) **Context:** This video discusses a research paper from collaborators at Japan Digital Design and the University of Oxford detailing a novel multi-agent Large Language Model (LLM) system designed to replicate the complex, hierarchical decision-making process of expert investment teams. The core innovation lies in structuring the agents to handle specific, fine-grained tasks—from data aggregation to final investment decisions—rather than relying on broad, generalized LLM outputs. ## Detailed Analysis The researchers developed a multi-agent LLM system to address the hype surrounding autonomous AI agents in finance. The system strictly mimics the structure of an institutional investment team, using GPT-4 as the base model. The simulation period covered 252 trading days from September 2023 to November 2025, using raw closing prices for Japan's largest cap stocks. The methodology involved three hierarchical levels: Level 1 agents (technical and quantitative) focused on aggregation, Level 2 (Adjustment Layer) synthesized this data using specific instructions (like setting temperature to 1), and Level 3 (Portfolio Manager) made the final decision. A key finding was that the fine-grained, structured approach significantly improved performance, yielding a Sharpe Ratio of 1.91 compared to the baseline of 1.38 (which used a massive model without structured decomposition). The researchers also conducted ablation studies, finding that removing the specialized Level 1 technical agent caused performance to collapse, demonstrating that specialization within the multi-agent framework is crucial and that simply using more raw compute is not the answer. Furthermore, the system was rigorously tested against look-ahead bias, ensuring it only used data available at the time of prediction, which is vital for real-world financial modeling. ### Research Overview - Paper from Japan Digital Design and Oxford - System is a multi-agent LLM mimicking expert investment teams - Goal is to perform reliable trading tasks ### Methodology and Training - Used GPT-4 as the base model - Trained on 252 days of Japanese large-cap stock closing prices (Sept 2023 - Nov 2025) - Agents were given fine-grained, specific instructions (e.g., temperature=1) ### Agent Hierarchy - Level 1 agents aggregate information (technical analysis, macro data); Level 2 agent adjusts; Level 3 agent (Portfolio Manager) makes final decision ### Key Results - Achieved Sharpe Ratio of 1.91, beating the baseline of 1.38 - Risk-adjusted returns significantly outperformed the baseline index ### Ablation Study Findings - Removing the Level 1 technical agent caused performance to 'plummet' - Proved that specialized agents are more important than just generalized AI power ### Bias Mitigation - System avoided look-ahead bias by excluding future data and used realistic transaction costs (10 basis points) ![Screenshot at 00:00: The opening visual displays the podcast/analysis theme with a prominent call to action: "BECOME A MEMBER TODAY!"](https://ss.rapidrecap.app/screens/VR5TbyMdGVU/00-00-00.jpg) ![Screenshot at 00:24: The speaker references the foundation of the research, stating the paper 'grounds those lofty claims in rigorous mathematical testing.'](https://ss.rapidrecap.app/screens/VR5TbyMdGVU/00-00-24.jpg) ![Screenshot at 01:12: A graphic representation of the multi-layered structure, explaining that the system moves from chain of command \(Level 1\) to the Adjustment Layer \(Level 2\).](https://ss.rapidrecap.app/screens/VR5TbyMdGVU/00-01-12.jpg) ![Screenshot at 04:57: The speaker highlights the final safety measure: the system strictly prevents look-ahead bias by avoiding feeding future data into the simulation.](https://ss.rapidrecap.app/screens/VR5TbyMdGVU/00-04-57.jpg) ![Screenshot at 09:09: The comparison between the fine-grained system \(Sharpe Ratio 1.91\) and the baseline \(1.38\) is introduced, indicating superior performance.](https://ss.rapidrecap.app/screens/VR5TbyMdGVU/00-09-09.jpg)