Hierarchical AI Multi-Agent Fundamental Investing: Evidence from China's A-Share Market

Battle Count: 80

Highly relevant to quantitative trading. The paper presents a complete end-to-end trading system with multi-source data integration (macro, industry, firm fundamentals, technicals, news, reports), RL-based portfolio construction, and risk management. It demonstrates strong out-of-sample performance (55.41% CR, 1.96 Sharpe in testing) on a realistic stock universe (CSI 300). The hierarchical design mirrors real-world fund management processes. The ablation study provides actionable insights into component contributions. The system addresses practical concerns like regime shifts, information leakage, and governance.

High implementation complexity due to: (1) Five distinct agent types requiring different data pipelines and models; (2) LLM integration with data obfuscation for News and Report agents; (3) PPO-based RL with action simulation, behavioral cloning, and actor-critic architecture; (4) Multi-level data processing (macro, industry, firm); (5) Risk scaling with volatility targeting; (6) Merrill Lynch investment clock regime classification; (7) Dynamic weight allocation with Top-K selection and K-means clustering reference. Requires expertise in RL, NLP/LLMs, financial data engineering, and portfolio management.

The paper provides detailed methodology, algorithm pseudocode (Algorithm 1), data descriptions, and evaluation metrics. However, no GitHub repository is mentioned. The use of Qwen3-32B as base LLM is specified, and data sources (CSI 300, CSRC industry classification, Baidu/Google news) are identified. Training/testing split is clearly defined (2019-2023 train, 2024 test). Ablation studies are provided. However, specific hyperparameters, code, and exact data processing pipelines are not fully disclosed.