JPMorgan Chase has built AI-powered investing agents that outperformed a traditional 60/40 stock-bond portfolio in backtests, offering an early look at how Wall Street is testing artificial intelligence for one of finance’s hardest calls: where to put money.

The bank’s researchers created a group of AI agents that shift allocations between stocks and bonds depending on market conditions, Bloomberg reported, citing a Thursday, 9 July, note by JPMorgan strategists led by Thomas Salopek.

The best-performing system beat a traditional portfolio of 60% stocks and 40% bonds by 0.7 percentage point a year over backtests spanning about two decades. It also delivered lower volatility and outperformed JPMorgan’s own rules-based market-regime model, according to the report.

The finding comes with a large caveat. The results were based on historical simulations, not live money management. A backtest measures how a strategy would have performed using historical market data, rather than real money in live markets.

The caveat

JPMorgan warned that the test should not be read as proof that AI can consistently beat markets.

“We strongly caution against uncritically accepting what amounts to in-sample, overly confident answers of AI,” the strategists wrote, according to Bloomberg.

In-sample results come from the same historical data used to design or train a model, making them especially vulnerable to overfitting. In investing, that means a strategy may look strong in the past but fail when market conditions change.

The experiment still marks a significant step in Wall Street’s use of generative AI. Banks and asset managers have spent the past two years using large language models for research, coding, client service, document review and internal workflow automation.

JPMorgan’s test points to a more consequential phase, where AI systems are being asked not just to assist human analysts but to make structured investment decisions under changing market conditions.

The JPMorgan team used AI agents powered by models from OpenAI and Anthropic. The agents classified markets into four regimes: Goldilocks, reflation, stagflation and risk-off.

The system tried to identify whether the market backdrop favored risk-taking or caution. In a strong growth environment, it could favor equities. When the outlook weakened or inflation pressures changed, it could move more toward bonds.

All eight AI agents tested beat the 60/40 benchmark on a risk-adjusted basis, Bloomberg reported. They also outperformed JPMorgan’s existing rules-based market-regime framework, suggesting the agents were able to improve on a model already used for asset-allocation guidance.

Reality check

That is the attractive part of the test. Financial markets are full of strategies that looked brilliant in simulation and ordinary, or worse, once exposed to real capital, transaction costs, crowding and changing investor behavior.

JPMorgan’s note appears to acknowledge that risk. The strategists said agentic AI should be grounded in a clear asset-allocation process rather than treated as the source of market expertise.

The broader concern is not only whether one bank’s AI agents can outperform but what happens if many firms use similar models, similar data and similar prompts to make similar allocation decisions.

Regulators and central banks have increasingly warned that wider use of AI in finance could make markets faster and more efficient, but also more correlated during stress.

The Financial Stability Board has said rapid AI adoption in finance, combined with limited data on how firms use the technology, creates a need for stronger monitoring and supervisory capacity.

The Bank of England has also warned that AI could create new financial-stability risks if it contributes to crowded trades, cyber vulnerabilities or excessive confidence in model-driven decisions.