JPMorgan's AI agents top 60/40 portfolio in two-decade backtests

A tier-one bank publishing a study where off-the-shelf AI agents beat its own hand-crafted quant model is the sort of small result that says more about where finance is heading than most flagship AI announcements. Bloomberg reported that a JPMorgan team led by strategist Thomas Salopek built eight AI investing agents on top of OpenAI and Anthropic models, and all eight beat a traditional 60/40 stocks-and-bonds portfolio on a risk-adjusted basis in backtests spanning the past two decades. The best of them topped the benchmark by 0.7 percentage point a year at lower volatility, and also beat JPMorgan's own rules-based market regime model.

The setup is worth understanding before drawing conclusions. The agents classify each moment in markets into one of four regimes based on growth and inflation, Goldilocks, reflation, stagflation, and risk-off, and rotate the allocation accordingly, favoring equities during periods of strong growth and increasing fixed-income exposure as the outlook deteriorates. It is the same regime-switching framework quant desks have used for years, but with the classification and reallocation call handed to a foundation-model agent rather than to a hand-coded rule.

Why this matters for anyone not sitting at JPMorgan: the bank's own rules-based regime model was the internal baseline, and generic frontier models running on top of it apparently did the job better. If that generalises, the moat in allocation strategy shifts away from bespoke quant models and toward data pipelines, prompt design, and orchestration. Asset managers that were planning to build their own regime engines now have to weigh whether the effort is still justified.

The honest caveat is that JPMorgan itself is putting up the biggest warning sign. Salopek and his colleagues wrote that they "strongly caution against uncritically accepting what amounts to in-sample, overly confident answers of AI", and said they are "wary to hand off asset allocation decision-making to an agent". These are historical simulations, not live money, and the reporting does not give you transaction costs, model versions, or how much the eight agents diverged from each other. What is worth watching is whether a firm actually lets one of these agents move real capital in production, and how the numbers hold up when the market it sees is not one it has already read about.