AI Outperforms Humans in Stock Investing

A new AI system has demonstrated superior stock-picking capabilities compared to human advisors and traditional methods, achieving a 167.52% cumulative return in China's A-share market from 2019 to 2024. This hierarchical multi-agent framework, developed by researchers from the University of Hong Kong and Hong Kong University of Science and Technology, integrates macroeconomic trends, firm fundamentals, technical data, and news sentiment to construct and manage portfolios. For general investors, this offers a transparent, cost-effective alternative that mitigates common biases like overconcentration and under-diversification, potentially democratizing access to sophisticated investment strategies.

The key finding is that this AI system consistently outperformed benchmarks, including the CSI 300 index and state-of-the-art models like MASS, with higher risk-adjusted returns and better drawdown control. For instance, it recorded a Sharpe ratio of 1.84 and a Sortino ratio of 3.05 during the training period, indicating strong performance relative to risk. In out-of-sample testing from January to December 2024, it maintained a cumulative return of 55.41%, surpassing the best baseline by 38.88 percentage points. This success stems from its ability to unify diverse data sources—from inflation rates to corporate reports—into a cohesive decision-making process.

Methodologically, the system employs a five-layer hierarchy inspired by real-world investment firms. At the top, a Macro agent screens sectors based on economic indicators like CPI and PMI, filtering out unfavorable industries. Within selected sectors, specialized agents—Fundamental, Technical, News, and Report—score individual stocks. The Fundamental agent analyzes financial health using metrics such as return on equity, while the Technical agent identifies trends via tools like moving averages. The News and Report agents use large language models (LLMs) to process media sentiment and corporate disclosures, converting unstructured text into quantifiable signals. A Portfolio agent then aggregates these scores using reinforcement learning to optimize allocations, and a Risk Control agent adjusts exposures in response to market volatility, ensuring stability during downturns.

Results from extensive back-testing show the system's robustness. Ablation studies revealed that removing any component—such as risk scaling or text processing—significantly degraded performance, underscoring the synergy between modules. For example, without risk control, maximum drawdown increased, highlighting its role in capital preservation. The framework's modular design also allows for interpretability, as each agent's contribution can be analyzed separately, addressing the 'black box' issue common in AI systems. This transparency is crucial for regulated environments where accountability is paramount.

In context, this AI approach matters because it addresses limitations in human advising, such as conflicts of interest and high costs, while outperforming robo-advisors in diversification and adaptability. By processing vast datasets rapidly, it helps investors navigate volatile markets like China's, where conditions shift frequently. The system's scalability and low operational costs could make advanced investing accessible to a broader demographic, not just affluent individuals, potentially reducing wealth disparities. Moreover, its hierarchical structure allows for future integration with human experts, fostering a collaborative 'copilot' model in finance.

Limitations include the system's reliance on historical data from 2019 to 2024, which may not capture all market regimes, and its focus on China's A-share market, raising questions about generalizability to other regions. The paper notes that while the AI reduces overfitting through rigorous testing, unforeseen economic shocks could challenge its adaptability. Future work will explore extensions to other asset classes and international markets, as well as incorporating human oversight to enhance decision-making in complex scenarios.