AI trading agents learn to coordinate without talking in simulated markets

The AI traders were trained to maximize their own profits, not to work together. Yet in synthetic-market experiments, they learned coordinated trading strategies without exchanging messages or sharing an intent to collude. The research found that this coordination amplified market instability when enough investors chased price trends—not evidence of manipulation in actual markets.

Knowledge at Wharton described the findings on October 13, 2026. The paper, Financial Market Fragility in the Era of AI Planning, is by Winston (Wei) Dou, Itay Goldstein, and Yan Ji. It combines a theoretical trading framework with reinforcement learning: training through repeated trials and rewards, across millions of trading episodes.

The framework lets traders connect decisions today with returns several steps later. Its informed speculators have private, price-relevant information. Alongside them are momentum investors, who buy when prices rise and sell when prices fall, and traders whose limits prevent them from fully correcting mispricing. The AI speculators learned rules that used prices as coordination signals.

The strategy starts with aggressive trading in one direction, creating a price bubble above or below an asset’s underlying value. Momentum investors follow the movement and deepen the distortion. Once the bubble forms, the informed traders gradually unwind their positions together, extracting profits at the trend followers’ expense.

During a rising bubble, each participant has a reason to sell before the others and secure profits before prices fall. Earlier, a trader could contribute little to building the bubble while benefiting from everyone else’s effort. Simply suspending coordination may impose too small a loss to discourage that free-riding.

The researchers describe a stronger response: aggressive sales that drive down the price and erase free-riding gains. In the simulations, other agents responded to a deviation during bubble creation with exceptionally aggressive punishment trades. Accepting that short-term cost enabled a faster return to the more profitable coordinated strategy.

Without AI planning, agents failed to learn the coordination across time needed to profit from pump-and-dump dynamics. With planning, they learned multistep manipulation strategies. The theoretical analysis also established two benchmarks:

• No coordination: informed speculators trade without coordinating with one another.
• A perfect cartel: informed speculators coordinate completely, effectively acting as a monopoly.

The authors used synthetic environments because researchers lacked enough real-time data on AI-informed trading. Goldstein noted that much of that data is proprietary and market participants are reluctant to share it.

Goldstein’s recommendation for regulators is to watch price patterns for noticeable changes, including greater volatility or pump-and-dump behavior.