AI agents beat the 60/40 — in a backtest. Does it matter?
JPMorgan says AI agents allocating between stocks and bonds beat the classic 60/40 portfolio over 20-year backtests. The question the desk cares about: is this an overfit demo everyone should laugh at, or the first datapoint of a structural shift in who runs allocation?
The chain of the argument
The underlying report, surfaced by Walter Bloomberg (@DeItaone), is more careful than the headline: JPMorgan tested AI agents that independently shift money between stocks and bonds; the best one beat a 60/40 portfolio by 0.7 percentage points a year with lower volatility, and all eight agents produced stronger risk-adjusted returns. JPMorgan itself flagged that these are simulations, not live performance, and that AI allocators could crowd trades and amplify market stress. Then Polymarket compressed it into a "JUST IN" that did 459k views, and the replies turned into a referendum on backtests. The skeptics piled on overfitting and the low bar. @hegdedarsh read the report itself and pulled out a different finding — the off-the-shelf models beat JPMorgan's own quant systems. CryptoTweets zoomed out: JPMorgan, Robinhood, and Coinbase all moved AI agents deeper into trading in the same month.
Every allocator with a backtest beats 60/40. Almost none beat it live.@Ferbin08
The two sides
For — the signal is real
The structural shift crowd
All eight agents delivered stronger risk-adjusted returns with lower volatility — breadth across designs, not one lucky configuration. And this comes from the bank's own published research, caveats included.
@DeItaoneThe buried headline: generic off-the-shelf models from OpenAI and Anthropic outperformed JPMorgan's own custom rules-based quant models. The multi-million-dollar moat of bespoke Wall Street quant infrastructure is evaporating.
@hegdedarshNot a lone experiment: Robinhood has opened 70,000+ agentic accounts since late May and will let users connect third-party AI agents to trade crypto via its Model Context Protocol; Coinbase shipped agent trading tools in June. This is becoming Wall Street infrastructure.
@CryptoTweetsThe edge isn't magic — it's the ability to process existing data faster and more accurately than humans, now demonstrated at the scale of a major bank.
@adelbucettaAgainst — it's an overfit demo
The backtest skeptics
The models were trained on the same 20 years they were tested on. The backtest grades an AI on history it has effectively memorized.
@bajamojiAny 100% equity portfolio beat 60/40 over the past two decades — record stock performance against historically poor bond returns. The bar was on the floor.
@CrabbyChicken76A 20-year backtest of an AI agent is a fit-to-history exercise with extra steps. The real test is 20 years forward, with no knowledge of what already happened.
@Ferbin08Backtests don't capture regime shifts, liquidity shocks, or the cost of changing AI models over two decades. Real edge is adaptability, not historical averages.
@DinoLeadingNewsBacktests struggle with data leakage. Treat the result as a big maybe.
@shoughtonjrJPMorgan's own strategists warned against trusting in-sample historical simulations. Overfitting 20 years of hindsight is easy — let it trade real capital through a black swan.
@hegdedarshThe exchange, in their words
Why this is an investment question, not a meme
The skeptics win the narrow argument — a 20-year in-sample backtest run by models trained on that same history, against a benchmark that anything equity-heavy beat, proves close to nothing about future returns. But the note-worthy part was never the 0.7 points. It's two structural facts underneath. First, commodity frontier models matched or beat bespoke quant infrastructure inside one of the world's most sophisticated banks — which reprices what proprietary quant capability is worth, and what active allocation fees can charge. Second, the execution rails for agentic money are being wired right now: a major bank testing AI allocators, and retail brokers connecting third-party agents to live crypto and equity accounts in the tens of thousands. Whether or not the backtest survives contact with reality, agent flow is becoming a market participant.
The TT desk thoughts
Fade the performance claim, position for the plumbing. Nobody should allocate a dollar on this backtest, and any product sold on "our AI beat the market historically" deserves the raccoon treatment the quote tweets gave it. But the direction of travel is investable: value migrates from proprietary quant moats toward the toll booths agent flow must cross — execution venues, agent-connectivity layers like Robinhood's MCP, and the crypto rails where agents can hold and move money natively. The under-priced risk sits on the other side: if allocators converge on the same handful of frontier models, they converge on the same trades — JPMorgan said this itself. Same-model crowding is a new systemic correlation factor, and drawdowns in an agent-allocated market will be faster and more synchronized than the backtests that sold the products. Own the rails, not the track record.
Keep reading
Why AI agents need crypto rails · Robinhood Chain, week one · All notes