Is the AI fund manager more powerful? JPMorgan backtesting: Annualized returns outperform classic investment portfolios with lower volatility

AI is moving towards the core of Wall Street’s investment decision-making. The team of JPMorgan strategist Thomas Salopek recently completed a backtesting experiment on AI investment agents, applying AI systems to market mechanism identification for the first time. The team built multiple AI agents capable of dynamically adjusting equity and bond allocations based on market conditions to explore the feasibility of autonomous investment decision-making.

The backtest results showed that the best-performing system in the simulated environment of the past two decades achieved an annualized return 0.7 percentage points higher than the traditional 60/40 equity-bond portfolio, with lower volatility, and outperformed JPMorgan’s existing rule-based market mechanism model as well.

While Wall Street is accelerating the deployment of AI in analysis, programming, and investment tools, this experiment marks a further extension of AI applications into the core decisions of capital allocation. However, JPMorgan clearly warns that these results should not be interpreted as evidence that AI has the sustained ability to outperform the market, as related research is still in its early stages.

Impressive Historical Simulation, No Live Trading Verification

The AI investment agent developed by JPMorgan researchers has a core function of dynamically adjusting equity and bond allocation ratios based on changes in market conditions. In historical simulations covering the past two decades, the optimal system achieved an annualized excess return of 0.7 percentage points, with lower volatility, surpassing the bank’s existing rule-based market mechanism models.

The strategist team pointed out in the report that this AI agent is designed to make decisions under uncertainty and can perform better than reasonable benchmarks. This is also the first time JPMorgan has publicly disclosed its research results in the AI-driven capital allocation field, marking a key step forward in the bank’s exploration of intelligent investment decision systems.

Despite positive backtest results, JPMorgan maintains caution in interpreting the conclusions. The bank emphasized that all of the above results are purely from historical simulations and have not yet been verified by actual market trading, hence these outcomes should not be used to infer any inherent ability of AI to continuously outperform the market.

The strategist team also warned in the report that market participants should avoid blindly accepting overconfidence in AI judgments based on in-sample backtest results. They believe that agent-based AI systems must be built on rigorous and prudent asset allocation processes, rather than simply assuming the agent itself as a source of expertise.

AI Consensus Risks Rising: Automated Trading Heading into the “Deep Waters” of Decision-Making

As enthusiasm for AI investment tools continues to rise on Wall Street, academia’s caution towards their potential systemic risks is also ramping up. According to Bloomberg, more and more studies are focusing on a core question: What changes will occur in market operation logic when a large number of institutions deploy similar AI models for investment decisions?

Researchers point out that while AI technology can significantly improve information acquisition efficiency and decision accuracy, it may also result in increasingly similar position structures and market manipulation risks, especially under stress scenarios where many institutions arrive at similar conclusions simultaneously, potentially amplifying market volatility. JPMorgan’s strategist team also acknowledged the existence of these risks in their recent report.

This JPMorgan test reflects the evolutionary path of AI applications on Wall Street. Over the past two years, major banks have widely integrated large language models into report generation, code writing, and internal investment tool applications. The current tests indicate the industry is assessing whether AI systems can move from assisting employees with decisions to taking on more decisive core functions, such as cross-market capital allocation.