There's good and bad news if you want to copy JPMorgan's new AI agent that outperforms the 60/40 portfolio

By Steve Goldstein

JPMorgan says AI can beat a standard investing model. There is just one catch if you try it yourself - models that disagree.

JPMorgan says it has designed an artificial-intelligent agent to evaluate markets and economic data and use it to outperform the standard 60/40 investing model, with less risk to boot.

MarketWatch spent the better part of a day recreating it. The good news: you can, even without access to expensive feeds of data or costly models, get an AI to tell you identify the current macroeconomic regime.

The bad news: it's very much an art and not a science. The AI chatbots don't entirely agree with each other, and in fact will sometimes even disagree with themselves the more times you ask.

First, a description of what JPMorgan strategists led by Thomas Salopek actually did.

They have data showing how different asset classes behave in different market regimes, so they wanted to find out how well an AI agent can do identifying the current macro regime.

And, they did. "The AI agent can be set up with a process to be empowered to make decisions under uncertainty, producing outperformance vs a reasonable benchmark," the strategists say.

JPMorgan took pains to test the model to only use information available on that date.

"If we ask the LLM model 'what regime is it?' as of March 2020 and it does not need to reason from the data, it has read thousands of articles of the COVID crash," they point out.

So strategists then ask the models to take in data, and say the likelihood they fit one of these four categories:

-- "Goldilocks" - growth above trend, inflation falling toward target, policy supportive/neutral; risk assets benefit, duration neutral.

-- "Reflation" - growth accelerating AND inflation rising, policy tightening behind the curve; cyclicals/credit benefit, long-duration govt suffers.

-- "Stagflation" - growth weak/falling while inflation high/sticky, policy constrained; defensive govt and real assets favoured, equities/credit lag.

-- "Risk_Off" - growth slowing sharply or financial stress dominant (widening credit spreads, equity drawdown, vol spike); long-duration govt (US/JPY) is the haven, HY and EM under pressure.

Their findings, after running it through two recent Anthropic and two OpenAI models, is that they all produce similar, though not identical, results.

-- 2001-2003 is dominated by risk-off and stagflation, reflecting the dot-com bubble unwind.

-- 2004-2007 is mostly Goldilocks and reflation. Think, latter Greenspan era.

-- 2008-2009 shows a shared risk-off block, as the global financial crisis hit.

-- 2010s are mostly Goldilocks and see the highest agreement among the models

-- 2020 produces a clean risk-off spike followed by Goldilocks, reflation, and finally stagflation

-- 2026 starts in Goldilocks and transitions into reflation.

And then they tested it, using a fixed investing strategy for each regime.

The returns, on average, beat the 60/40.

The AI chatbots beat a simple 60/40 model

The JPMorgan strategists did wonder if the AI chatbots cheated. "Although the data is lagged and the prompt is date-anonymized, the LLM models are still trained on data after the cut-off point and may implicitly recall the outcome of recognizable historical episodes (e.g., 2008, COVID, etc.)," the strategists said.

So, could MarketWatch recreate what JPMorgan did? Keep in mind this reporter has a precisely zero-dollar budget for data and AI. There is enterprise access to Gemini, which is helpful.

The first challenge is to obtain all the economic and financial market data. The economic data is pretty easy to get, in an automated way, thanks to the wonderful FRED tool from the St. Louis Fed. (One unfortunate drawback is that FRED no longer includes the Institute for Supply Management's manufacturing index, so here the Philadelphia Fed manufacturing series is substituted.)

On the markets side of things, it gets a bit more challenging, but exchange-traded funds were used to create JPMorgan's rates/FX/commodities/risk panel: the iShares 7-10 Year Treasury Bond ETF IEF, the iShares 1-3 Year Treasury Bond ETF SHY, the Invesco DB US Dollar Index Bullish Fund UUP, the Invesco DB Commodity Index Tracking Fund DBC, the Cboe Volatility Index VIX, the iShares iBoxx $ High Yield Corporate Bond ETF HYG and the iShares iBoxx $ Investment Grade Corporate Bond ETF LQD.

What MarketWatch found was that it's probably best to retrieve the data first, and then feed it into the AI tool of choice with the JPMorgan instructions. The Python script is available upon request. (You can create an entire Python script that both gets the data and evaluates it, but if you're doing it on the cheap, you're probably not getting the best models.)

The AI models, in broad strokes, seem to agree.

Claude's assessment of the economy and markets.

ChatGPT, with what it assessed to be 73% confidence, said there's a 43% chance of a reflation regime next month and a 31% chance of Goldilocks. Claude, at 70% confidence, said there's a 60% chance of a reflation regime but put stagflation at second-most likely, at 19%.

Grok and Gemini also favored reflation during MarketWatch's testing.

JPMorgan's testing of models, done at the end of June, also found all of the models favoring reflation.

The JPMorgan strategists were at pains to say investors should not just blindly do what an AI recommends. Here, AI seems to be saying, most of your money should be in stocks, with some debate over how much into government bonds and how much corporate credit.

-Steve Goldstein

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

(END) Dow Jones Newswires

07-13-26 0417ET