JPMorgan's AI agents outperform 60/40 portfolio in historical backtests

JPMorgan's AI-powered investing agents outperformed a traditional 60/40 portfolio in historical backtests, though the bank cautioned that the results are based on simulations

As investors increasingly turn to artificial intelligence (AI) for help with everything from stock-picking to risk management, JPMorgan Chase & Co has been testing whether the technology can do something more ambitious: allocate capital itself.

 

The early results are encouraging. Researchers at the bank built a range of AI-powered investing agents that shift between stocks and bonds depending on changing market conditions.

 

In backtests covering the past two decades, the best-performing system outperformed a traditional 60/40 portfolio — 60 per cent in equities and 40 per cent in bonds — by 0.7 percentage point a year while delivering lower volatility. It also outperformed JPMorgan's own rules-based market regime model, according to strategists led by Thomas Salopek.

 

 

The results come with an important caveat. They are based on historical simulations rather than live investing, and JPMorgan warned against treating them as proof that AI can consistently outperform markets. Still, they offer a glimpse of how AI could reshape investment management as automated trading continues to gain momentum.

 

“The AI agent can be set up with a process to be empowered to make decisions under uncertainty, producing outperformance vs a reasonable benchmark,” the strategists wrote in a note on Thursday, describing the work as the firm's first attempt to build an AI system for identifying market regimes.

 

The experiment offers an early indication of Wall Street's next phase of AI adoption. Banks have spent the past two years embedding large language models into research, coding and internal investment tools. Increasingly, they are testing whether those systems can move beyond assisting employees to making one of the industry's most consequential decisions: how to allocate capital across markets.

 

The findings come as a growing body of academic research raises questions about what could happen if investors increasingly rely on similar AI models to make investment decisions. While the technology may make investors faster and better informed, researchers have warned it could also lead to more crowded trades.

 

The JPMorgan strategists also acknowledged those risks.

 

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

 

Even so, the findings add to growing evidence that AI can perform increasingly sophisticated investment tasks. Using agents powered by models from OpenAI and Anthropic, the JPMorgan team designed a system that classifies markets into four regimes based on growth and inflation: Goldilocks, reflation, stagflation and risk-off.

 

The AI agents were then tasked with deciding how to allocate capital across asset classes in each environment — favouring equities during periods of strong growth, for example, and increasing fixed-income exposure as the outlook deteriorated.

 

All eight AI agents tested outperformed the traditional 60/40 portfolio on a risk-adjusted basis. They also beat JPMorgan's existing rules-based market regime model, suggesting the technology was able to improve on a framework already used to guide asset-allocation decisions.

 

“We are enthusiastic about the possibilities of agentic AI, even as we are wary of handing off asset-allocation decision-making to an agent,” Salopek and his colleagues wrote.