JPMorgan AI Agents Outperform 60/40 Portfolio by 0.7% Annually

JPMorgan's research team built AI-powered investing agents using models from OpenAI and Anthropic that classify markets into four regimes (Goldilocks, reflation, stagflation, risk-off) and dynamically allocate capital across equities and bonds. In backtests spanning two decades, the best-performing agent outpaced a traditional 60/40 stock-bond portfolio by 0.7 percentage points annually while delivering lower volatility. All eight AI agents tested beat both the 60/40 benchmark and JPMorgan's own rules-based market regime model, suggesting the technology can improve on existing asset-allocation frameworks.

Strategists led by Thomas Salopek emphasized critical caveats: the results are from historical simulations, not live trading, and JPMorgan cautions against treating backtests as proof of consistent outperformance. The team also acknowledged systemic risks if crowded trading emerges from multiple firms using similar AI models, which could amplify market stress and make markets easier to manipulate. JPMorgan stated it remains wary of handing off real asset-allocation decisions to agents without domain-specific guardrails.

For practitioners, this validates the commercial promise of agentic AI for capital allocation while surfacing a critical operational risk: the shift toward autonomous decision-making at scale across the financial sector. Architects should note that regime-based allocation frameworks are now table-stakes for AI in finance; the question is how to layer human oversight, prevent herding, and ensure models degrade gracefully when conditions diverge from training data.