Why AI Agents Beat 60/40: Static Allocation Fails Because It Never Questions Its Assumptions

Background

"60/40" is Wall Street's most canonical asset allocation rule—60% stocks, 40% bonds. This ratio has been used since 1942, unchanged for over 80 years. The logic is straightforward: stocks provide growth, bonds provide stability, and they're negatively correlated. When stock markets crash, bonds typically rise, creating a natural hedge.

But 2022 broke this myth. That year, stocks and bonds fell simultaneously—the worst year since 1937. The Federal Reserve's aggressive rate hikes caused long-term bond prices to plummet, while stocks declined due to rising rate expectations. The static 60/40 portfolio couldn't adapt, and many fund managers began asking seriously: is there a better way?

JPMorgan's Experiment

JPMorgan's cross-asset strategy team, led by strategist Thomas Salopek, designed a controlled experiment: build 8 AI agents and give them a simple mandate—autonomously adjust allocation between stocks and bonds based on market conditions.

The AI agents used straightforward judgment criteria, dividing the market into four scenarios (regimes):

1. High growth + Low inflation: Overweight stocks (best economic state) 2. High growth + High inflation: Reduce stocks, increase bonds (growth is offset) 3. Low growth + Low inflation: Pure bond portfolio (recession but no inflation risk) 4. Low growth + High inflation: Diversified allocation (worst case, even boring assets fall)

Importantly, the AI agents' job wasn't to predict—that's critical—but to sense the current scenario and adjust immediately.

Backtesting results: all 8 AI agents outperformed the static 60/40. Remarkably, even JPMorgan's own traditional forecasting models were beaten by these simple adaptive agents.

Why Static Lost to Dynamic

This result seems surprising at first, but the logic is clear:

1. Static allocation embeds an assumption: market environment is constant

The 60 and 40 in 60/40 are derived from long-term statistics from 1942 to the 1980s measuring stock-bond correlation. When market environment changes—say, from low inflation to high inflation, or from stable growth to stagflation—this assumption breaks down.

Yet decision-makers rarely proactively question this assumption. Not from laziness, but because: - Institutional inertia: 60/40 has become the standard allocation for pension funds, insurance companies, and financial advisors. Changing it requires internal approvals and client communication, with high costs - Psychological lock-in: "This is the classic approach" reinforces commitment, making people defensive rather than responsive when faced with contradictory evidence - Regulatory constraints: Many institutional investors have investment policies that lock in allocation ratios; changing them requires board approval

2. AI agents lack these locks

AI agents ask afresh every month or quarter: "What scenario are we in now?" If the answer changes, allocation changes immediately. They lack the psychological resistance of "but this is the classic approach," and they have no political costs.

3. The cost of adaptation isn't as high as imagined

People often worry that frequent adjustments incur trading costs and tax costs. But modern index funds and ETFs have trading costs that are already extremely low. In 20-year backtests, even accounting for trading costs, the AI agents' advantage persists.

Why JPMorgan Didn't Claim "AI Has Beaten Markets"

While publicly releasing these results, JPMorgan explicitly stated: don't treat this as evidence that "AI consistently beats markets." Why?

1. Backtesting bias: Historical data allows us to be "wise after the fact." AI agents adjusted parameters on known outcomes, so of course they look good. Future markets will contain scenarios the JPMorgan team has never seen.

2. Overfitting: All 8 AI agents used similar four-scenario classifications. What if the market's true scenarios number 12? Or aren't classifiable at all?

3. Out-of-sample risk: The 20-year backtest includes the 2008 financial crisis, but not "all bonds fail simultaneously"—complete black swan events like sovereign default or monetary system collapse.

The Real Insight

The core insight of this experiment isn't "AI beats humans," but rather:

Static rules fail when environments change, yet people can't adapt in time because of institutional, psychological, and political reasons. AI agents lack these obstacles, so they can do what humans know they should do but can't.

This recalls management theorist Herbert Simon's concept of "bounded rationality"—people aren't irrational, but rather constrained by institutions, cognition, and decision costs. AI's "advantage" is actually: it has fewer of these human constraints.

But this raises a deeper question: if we hand all decision-making to AI, won't the loss of "human constraints" create new risks? Might AI agents be more fragile than static 60/40 in the face of unseen black swans?

This is the real concern behind JPMorgan's cautious stance.