Research
JPMorgan's AI agents beat the 60/40 portfolio in twenty years of backtests, with real caveats attached
Eight AI agents built to classify market regimes and shift stock-bond allocation outperformed a static 60/40 split by 0.7 percentage points annually across two decades of historical simulation. JPMorgan is explicit this is research, not a product.
JPMorgan's research team built eight AI-powered agents, using OpenAI and Anthropic models, tasked with classifying prevailing market conditions into one of four regimes: goldilocks, reflation, stagflation and risk-off, and dynamically shifting allocation between equities and fixed income in response. Backtested across roughly twenty years of market history, every one of the eight agents outperformed a static 60/40 stock-bond portfolio on a risk-adjusted basis, beating it by 0.7 percentage points of annualised return while carrying lower volatility. All eight also beat JPMorgan's existing rules-based regime model.
Seventy basis points sounds like a rounding error until you sit inside institutional asset management, where that margin compounded over two decades, with lower volatility alongside it, is the kind of outperformance that justifies an entire product line. The result is not that AI predicted the market. It is that a regime-classification approach, deciding how much risk to carry based on a read of current growth and inflation conditions, worked better when an AI agent made that judgement than when a fixed rules-based model did, across a long and varied historical sample that included multiple genuine regime shifts.
The caveats JPMorgan itself attaches are the part of this story worth taking seriously rather than skipping past. This is a historical backtest, not live trading performance, and the two routinely diverge: backtests cannot suffer from execution slippage, data leakage is a persistent risk in any regime-classification model trained and tested on overlapping history, and a model that correctly classified past regimes has not been proven to correctly classify regimes it has never seen before, which is the entire point of a regime shift. JPMorgan has been explicit that this is research-stage work with no indication of an imminent client-facing product.
The signal worth extracting is narrower and more durable than the headline number. A large institution with genuine research resources found that an AI agent making qualitative regime judgements outperformed a rules-based system encoding the same logic mechanically. That is a real data point on where AI adds value in asset allocation, judgement under ambiguity, rather than the more commonly hyped claim of AI predicting individual security prices. Investors should read this as evidence for a specific, narrower thesis about AI in portfolio construction, not as proof that AI has cracked market timing generally.
Read the original: Crypto Briefing - JPMorgan's AI agents beat the 60/40 portfolio in twenty years of backtests, with real caveats attached. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.