JPMorgan has created AI agents that outperformed a 60/40 portfolio based on 20 years of data

JPMorgan Chase, the largest U.S. bank by market capitalization, has developed eight AI agents that automatically reallocate funds between stocks and bonds depending on market conditions. Based on 20 years of data, the best of these outperformed the classic 60/40 portfolio by 0.7 percentage points per year. However, the bank cautioned that the result was obtained from simulations, not actual trades, and should not be considered proof that AI can consistently outperform the market.

The team led by Thomas Salopek used models from OpenAI and Anthropic for the agents. The market was divided into four scenarios based on two parameters—GDP growth and inflation: “Goldilocks” (economic growth with low inflation), reflation (government policy to stimulate the economy), stagflation (a combination of economic recession and inflation), and risk-off (a sell-off of risky assets). Participants had to decide how to allocate funds among asset classes depending on the regime: increasing the share of stocks during a strong economy and increasing investments in fixed-income securities when the outlook deteriorated, according to Bloomberg.

In terms of risk-adjusted returns, all eight agents outperformed the 60/40 portfolio. All of them also outperformed JPMorgan’s current algorithm, which identifies these modes based on strict rules and which the bank relies on for asset allocation. “An AI agent can be programmed to operate in a way that allows it to make decisions under conditions of uncertainty and outperform a reasonably chosen benchmark,” wrote JPMorgan strategists, describing the work as the bank’s first attempt to build an AI system for this purpose (quoted from Bloomberg).

Researchers are increasingly asking what would happen if investment decisions were made using similar models. Technology makes investors faster and better informed, but it can also lead to too many people entering the same trades, as well as facilitate market manipulation and exacerbate periods of stress if market participants reach the same conclusions, Bloomberg notes.

JPMorgan’s own strategists acknowledge this risk. “We strongly caution against uncritically accepting what are, in fact, AI-generated responses based on data with which the AI is familiar—and are therefore overly confident,” they said. “An AI agent must rely on a well-thought-out asset allocation strategy, rather than on the naive assumption that the agent itself can be a source of expert knowledge.”

In February 2026, Ray Dalio launched an AI clone to provide advice—a chatbot that answers questions about markets and the economy based on the philosophy of the founder of Bridgewater Associates. The billionaire investor believes that the days of making decisions by keeping the market in mind are over, and that most investors are only now realizing this and are using AI haphazardly.

In his view, everyone will have to develop their own AI systems: retain their own understanding and beliefs, but submit them for verification by a reliable AI assistant. “Simply put, you need to be at the forefront of these technologies, or you’ll lose out. You’ll simply fall behind,” Dalio said at a closed-door event attended by an Oninvest correspondent.

Dalio advises turning to AI not for ready-made answers, but for the sake of a discussion that allows us to explore cause-and-effect relationships. He sees the danger in the opposite—in the habit of taking a machine’s conclusions as instructions for action. According to him, another possible scenario is one in which “the colossal growth in the intellectual capabilities of machines will be accompanied by a complete lack of practical wisdom among people—to such an extent that, by using this super-powerful intelligence, we might simply destroy one another” (quoted from Oninvest).

This article was AI-translated and verified by a human editor