Researchers Built an AI Trader That Rewrites Its Own Rules

Most AI trading bots run on instructions written by humans and locked in before the bot ever touches a market. EvolveTrade changes that: it treats those instructions as something the AI is allowed to rewrite. Every so often, a second AI (the "Policy Agent") reads through the log of past decisions and the actual profit or loss they produced, then edits the rulebook for the next round of trades. The base AI model never changes — only the playbook does.

In tests spanning several market conditions and two different AI models, the self-updating version usually did better on two measures: Sharpe Ratio (return measured against how bumpy the ride was) and total cumulative return. Notably, the bots that evolved started writing and running more of their own code to analyze data, and they switched on the kinds of calculations that suited whatever market they were in. The authors traced specific rule changes back to specific differences in returns.

The catch is that this is a research paper, not a product. The trades were simulated, not made with real money at a real brokerage. There's no word on trading fees, taxes, or what happens when markets do something nobody has seen before. "Most evaluated settings" means not all of them — in some conditions, the fixed-rule bot still won. And a bot that adapts quickly can also adapt itself into bad habits.

Still, the bigger idea travels well beyond finance. The trick here is letting AI improve the reusable procedure it follows — how it gathers evidence, checks its work, and manages risk — rather than just swapping in a newer model. That same approach could show up in customer service, medical triage, or any job where an AI follows a set of steps.