Paper: arXiv 2510.21147

Authors: Chujun He, Zhonghao Huang, Xiangguo Li, Ye Luo, Kewei Ma, Yuxuan Xiong, Xiaowei Zhang, Mingyang Zhao

Abstract

We present a multi-agent, AI-driven framework for fundamental investing that integrates macro indicators, industry-level and firm-specific information to construct optimized equity portfolios. The architecture comprises: (i) a Macro agent that dynamically screens and weights sectors based on evolving economic indicators and industry performance; (ii) four firm-level agents – Fundamental, Technical, Report, and News – that conduct in-depth analyses of individual firms to ensure both breadth and depth of coverage; (iii) a Portfolio agent that uses reinforcement learning to combine the agent outputs into a unified policy to generate the trading strategy; and (iv) a Risk Control agent that adjusts portfolio positions in response to market volatility. We evaluate the system on the constituents by the CSI 300 Index of China’s A-share market and find that it consistently outperforms standard benchmarks and a state-of-the-art multi-agent trading system on risk-adjusted returns and drawdown control. Our core contribution is a hierarchical multi-agent design that links top-down macro screening with bottom-up fundamental analysis, offering a robust and extensible approach to factor-based portfolio construction.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)

Why this score: The paper employs advanced mathematics including reinforcement learning and complex multi-agent coordination, while also presenting rigorous backtesting on a real dataset with outperformance against benchmarks and state-of-the-art methods.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Hierarchical AI Multi-Agent<br>System for Equity Investing"] --> B["Data Inputs:<br>CSI 300 Constituents,<br>Macro, Financial, News Data"]
  B --> C["Methodology: Hierarchical Multi-Agent Architecture"]
  C --> D["Macro Agent:<br>Dynamic Sector Screening<br>& Weighting"]
  C --> E["Firm-Level Agents:<br>Fundamental, Technical,<br>Report, News Analysis"]
  D & E --> F["Portfolio Agent:<br>RL-based Integration<br>& Trading Strategy"]
  F --> G["Risk Control Agent:<br>Volatility Adjustment<br>& Drawdown Control"]
  G --> H["Outcomes:<br>Consistent Outperformance<br>Risk-Adjusted Returns,<br>Superior to Benchmarks"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a multi-agent AI-driven framework for fundamental equity investing that integrates macro, industry, and firm-level information to build optimized portfolios.

Source sentence
We present a multi-agent, AI-driven framework for fundamental investing that integrates macro indicators, industry-level and firm-specific information to construct optimized equity portfolios.

Assumes 5

  • Model parameters are trained on data from Jan 2019 to Dec 2023 and validated on Jan–Dec 2024, implying stationarity across the split. data
    Source sentence
    model parameters are trained exclusively on the sample from January 2019 to December 2023, with performance validated on the out-of-sample period covering January to December 2024.
  • The Risk Control agent can adjust portfolio positions in response to market volatility. model
    Source sentence
    Risk Control agent that adjusts portfolio positions in response to market volatility.
  • Reweighting of agent outputs occurs after sector context has been established, relying on hierarchical separation. model
    Source sentence
    the hierarchical separation ensures that reweighting occurs after sector context has been established.
  • LLM outputs are embedded within the hierarchical workflow rather than used as simple features. model
    Source sentence
    Rather than treating LLM outputs generally as features, we embed them within the hierarchical workflow to allow them to compare with the structure signals in a dynamic way.
Show 1 more
  • The dataset of Chinese A‑share market constituents provides sufficient macro, industry, and firm‑level data for training and testing. data
    Source sentence
    We evaluate the proposed framework using a comprehensive and challenging dataset comprising the Chinese A‑share market from January 1, 2019, to December 31, 2024

Data

Backtest sample: 2024-01 – 2024-12 · CSI 300 constituent stocks · equities · daily

  • Chinese A‑share market (CSI 300 constituents) market-prices unknown
    2019-01-01 to 2024-12-31 · CSI 300 Index constituents
    Source sentence
    We evaluate the system on the constituents by the CSI 300 Index of China’s A‑share market
  • Macro indicators (CPI, money supply, PMI) macro unknown
    Source sentence
    daily return and macro data including CPI, money supply and PMI, so the Macro agent facilitates comprehensive analysis in both macro and industrial perspectives.

Costs

Not modeled A trading or execution result is reported without costs.

Finds 2

Reported performance (Hierarchical AI Multi-Agent Investing): Sharpe 1.96 · total return 55.4% · max drawdown 12.5%

  • In the training period the framework achieved a cumulative return of 167.52%, Sharpe ratio 1.84, Sortino ratio 3.05 and Calmar ratio 1.42, outperforming benchmarks. performance
    Source sentence
    It achieves a cumulative return (CR) of 167.52%, a Sharpe ratio (SR) of 1.84, a Sortino ratio of 3.05, and a Calmar ratio of 1.42.
  • In the out‑of‑sample testing period the model delivered a cumulative return of 55.41%, Sharpe 1.96, Sortino 3.72 and Calmar 4.43, beating standard benchmarks and non‑hierarchical multi‑agent baselines. performance
    Source sentence
    Ours 55.41 55.41 28.20 14.90 1.96 3.72 -12.52 4.43

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 21 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.