JPMorgan shares ticked higher on Friday after the bank disclosed that AI-powered investing agents beat a classic 60/40 portfolio in historical simulations. The note revealed systems designed to rotate between equities and fixed income in response to evolving market conditions.
In backtests spanning the past two decades, the highest-performing AI system outpaced a 60/40 mix by 0.7 percentage point per year while registering lower volatility, the bank's strategists said. The same system also outperformed the firm's rules-based market regime model, according to the strategists led by Thomas Salopek.
Those results come from historical simulations rather than real-world trading. The strategists warned against treating the backtests as definitive evidence that AI can reliably beat markets in live conditions. "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 reportedly wrote Thursday, describing the project as the firm's initial attempt to construct an AI system for identifying market regimes.
The strategists also flagged important caveats around wider adoption of such approaches. "We strongly caution against uncritically accepting what amounts to in-sample, overly confident answers of AI," they wrote. "Agentic AI needs to be grounded in a well thought-out asset allocation process, rather than naively assuming the agent can be the source of the domain knowledge."
Bank researchers built the agents to shift exposures between stocks and bonds based on signals tied to changing market environments. The work represents an exploratory effort to see whether machine-driven decision agents can improve capital allocation across markets.
Financial institutions have spent the past two years incorporating large language models and related systems into research, coding and internal investing tools. The bank's note indicates those efforts are evolving into trials that test whether such systems can take on capital allocation roles - albeit under careful oversight and with explicit recognition of limitations from backtested results.
Context and next steps
The bank described this as an initial foray rather than a finished product. Strategists emphasized the need to embed agentic AI within established allocation frameworks and to guard against overconfidence in in-sample performance. The findings will likely shape further internal testing rather than immediate changes to client-facing products.