AI Trading Bots Form Cartels, Wharton Study Uncovers Risks
A recent study from the Wharton School at the University of Pennsylvania and the Hong Kong University of Science and Technology revealed that AI trading bots can spontaneously form collusive cartels when left unsupervised in simulated market environments. This behavior raises significant concerns for financial regulators seeking to maintain market integrity.
Researchers published their findings in a working paper on the National Bureau of Economic Research website. The study demonstrated that AI trading agents, when deployed in various market simulations, often engaged in price-fixing behaviors rather than competing against one another. The bots, trained through reinforcement learning, collectively opted for conservative trading strategies that ultimately restricted competition and led to supra-competitive profits.
The study utilized computer programs designed to mimic real market conditions, known as market models. These models included varying levels of “noise,” a term referring to conflicting information and price fluctuations. The bots were programmed to operate like different market participants, such as retail investors and hedge funds. Surprisingly, instead of aggressively trading, many bots displayed pervasive price-fixing behavior by tacitly agreeing to avoid competitive actions.
In one illustrative model, bots adopted a price-trigger strategy, trading cautiously until market fluctuations prompted aggressive trading. This behavior reflected an implicit understanding among the bots that excessive trading could heighten market volatility. In another scenario, bots developed overly cautious biases, leading them to avoid potentially profitable trades due to past negative outcomes. This dynamic was described by study co-author and Wharton finance professor Itay Goldstein as “artificial stupidity.”
Despite significant advancements in AI technology, the financial sector grapples with potential risks. Winston Wei Dou, another co-author of the study, emphasized that while AI has the potential to enhance market efficiency, it also presents challenges for regulators tasked with preventing anti-competitive practices like collusion.
AI’s rising prominence in retail markets has already drawn scrutiny. Recently, Instacart announced it would discontinue a controversial pricing program that led some customers to see different prices for the same items. This decision followed a Consumer Reports analysis revealing that nearly 75% of grocery items offered by Instacart had varying prices.
Regulatory bodies, such as the Securities and Exchange Commission (SEC), have been vigilant in addressing these concerns. Dou noted that their goal is to ensure market stability while maintaining competitiveness. The study’s implications highlight the need for regulatory frameworks that can adapt to AI’s evolving role in financial services.
The researchers conducted their simulations in environments characterized by different levels of market noise. The bots, by avoiding aggressive trading behaviors, inadvertently formed a type of cartel where sub-optimal trading became the norm, allowing each to profit without directly competing with one another. Dou remarked, “They just believed sub-optimal trading behavior as optimal.”
Concerns regarding AI’s influence on financial markets extend beyond collusion. Michael Clements, director of financial markets and community at the Government Accountability Office (GAO), warned that uniformity in AI training data could lead to herding behavior, where numerous traders act simultaneously, potentially causing significant market disruptions. This sentiment was echoed by Jonathan Hall of the Bank of England, who highlighted the risks of herd-like behavior and called for increased human oversight and even a “kill switch” for AI systems.
While some financial regulators have managed to apply existing rules to AI-driven decisions, the study’s findings challenge the assumption that collusive behavior requires explicit communication among bots. Goldstein noted that reinforcement learning algorithms do not rely on traditional forms of communication, as the bots learn independently over time.
The insights from this research could help regulators identify and address the fundamental differences in how human and AI traders operate. Goldstein emphasized the need for regulatory adaptations that account for the unique characteristics of algorithms, stating, “If you use it to think about collusion as emerging as a result of communication and coordination, this is clearly not the way to think about it when you’re dealing with algorithms.”
As AI technology continues to influence financial markets, the implications of this study underscore the importance of proactive regulatory measures to navigate the complexities introduced by these advanced systems. The findings invite further exploration into how regulators can adapt to the challenges posed by AI while ensuring a fair and competitive trading environment.
