Original title: Can Reinforcement Learning Efficiently Discover Price Manipulation?
Ioanna-Yvonni Tsaknaki, Andrea Macrì, Fabrizio Lillo · July 2026
Financial markets can suffer from price manipulationTrading in a way that intentionally moves prices to profit from the movement., where traders intentionally distort asset prices to make unfair profits. Researchers wanted to see if an artificial intelligence could learn these manipulative trading tricks on its own better than a traditional mathematical model. They tested this in a simulated single-asset market. In this setting, trades move prices permanently in a non-linear way and temporarily in a linear way. One method used a standard model that knew the market rules but had to estimate noisy details. The other was a model-free reinforcement learningA type of AI that learns from experience without needing to know the rules of the environment. agent that learned purely by trial and error from trading data. They found that with medium price volatility, the reinforcement learning agent discovered profitable manipulative tricks with limited data and beat the traditional model. However, when price volatility was small, the traditional model won. When volatility was large, neither method could find manipulation opportunities.
You are trying to learn a new card game by watching others play. You don't know the rules, so you guess and see if you win. Sometimes you find a clever move that works even though you don't understand why. Here, the reinforcement learning agent learns to manipulate prices by trial and error, without knowing the market model, and finds profitable strategies that a method relying on estimated rules might miss.
This matters because financial markets are vulnerable to manipulation. If RL agents can discover manipulation strategies without understanding the market, they could be used for harmful purposes. Regulators need to be aware that even model-free learning can find exploits, and safeguards are necessary.
The results are limited to a specific market model (Almgren-Chriss) and simulated data. RL only outperforms for intermediate volatility; for low volatility, model-based is better, and for high volatility, neither works. The paper does not test real market conditions or consider regulatory constraints.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Can Reinforcement Learning Efficiently Discover Price Manipulation? — Ioanna-Yvonni Tsaknaki, Andrea Macrì, Fabrizio Lillo
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.