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

Battle Count: 74

Highly relevant to quantitative trading. The paper directly addresses the design of LLM-based trading agents, evaluates risk-adjusted returns (Sharpe ratio) through backtesting, constructs long-short portfolios, performs portfolio optimization with risk parity, and benchmarks against market indices. It provides practical insights into agent architecture, task design, and information flow for trading systems. The focus on interpretability and real-world deployment constraints makes it particularly relevant for institutional quantitative trading applications.

Moderate to high complexity. Requires setting up a 7-agent hierarchical system with specific prompt engineering for each agent, integrating multiple data sources (stock prices, financial statements, news, macro data), implementing technical indicator calculations (RoC, Bollinger Bands, MACD, RSI, KDJ), computing financial metrics across five dimensions, performing monthly rebalancing with long-short portfolio construction, and conducting portfolio optimization with risk parity weighting. However, no model training is required as it uses pre-trained LLMs for inference. The main complexity lies in prompt design, data pipeline construction, and the hierarchical agent coordination logic.

The authors commit to releasing implementation codes and prompts upon acceptance. They use open data sources (Yahoo Finance, EDINET API, FRED, Ceek.jp News). The experimental setup is well-documented with specific model (GPT-4o), temperature settings, backtesting period, and evaluation protocols. However, the code is not yet publicly available at the time of the paper.