
By Associate Professor JI Yan, Department of Finance, HKUST Business School
Algorithms are transforming market efficiency and risk dynamics, presenting both opportunities and risks. Understanding AI collusion in financial market and its implications is essential for navigating the evolving landscape.
Algorithmic trading has been prevalent in financial markets worldwide for over two decades. The use of computer algorithms allows market participants to automatically execute trading decisions and efficiently manage their orders, increasing market liquidity and enhancing price informativeness.1
Recently, breakthroughs in AI technologies, particularly reinforcement learning and deep neural networks, have captured the interest of major hedge funds and investment powerhouses. These entities are now leveraging AI to enhance their algorithmic trading, enabling algorithms to trade intelligently through self-learning in dynamic environments rather than relying on rigid, hard-coded protocols. With the U.S. Securities and Exchange Commission's (SEC) approval of Nasdaq's reinforcement learning-based, AI-driven order type, the integration of AI in trading is gaining significant momentum.2 AI-powered trading presents new regulatory challenges and has the potential to fundamentally reshape financial markets.
One of the key roles of financial markets is price discovery. By aggregating information from sophisticated market participants who trade for profit based on informative signals, market prices reflect the fundamental values of underlying assets. Over the years, this powerful source of information has become increasingly informative due to reduced information costs and transaction fees, improved data access, and a rise in institutional ownership.3 The integration of algorithmic trading with AI could significantly enhance market efficiency, thanks to AI's ability to analyze vast amounts of data and make predictions. However, recent research indicates that AI-powered trading also carries risks. If the development and adoption of AI are dominated by a few leading entities, market prices may become less informative regarding asset fundamentals.
AI collusion in financial markets
A crucial feature of financial markets is information asymmetry. Investors can profit from trading an asset if they possess more knowledge about its value than other investors. Trading becomes highly strategic when multiple investors are informed about the asset’s value. If these informed investors trade the asset intensively, their private information will quickly be reflected in the market, leading to significant price changes that can render their trading unprofitable. To maximize profits, informed investors may coordinate their trading to allow their private information to be gradually absorbed by the market without causing immediate, substantial price fluctuations. This coordinated trading behavior among informed investors is known as collusion.
Such collusion is typically challenging to achieve. Explicit collusion is illegal, and implicit collusion is difficult to maintain, as informed investors always have the incentive to deviate from the agreement to enhance their short-term profits. Moreover, effective monitoring of each other's behavior is hindered by significant information asymmetry in financial markets.
While collusion in financial markets might be challenging for humans, my recent working paper with Wharton finance professors Winston Wei DOU and Itay GOLDSTEIN indicates that such collusion may emerge when trading orders are executed by AI-powered algorithms.4 We developed a trading laboratory where AI algorithms trade assets using basic reinforcement learning techniques. Over time, these algorithms adapt their behavior through self-learning and learn to coordinate independently, even without direct instructions or communication. This AI collusion suggests that market liquidity and price informativeness may be negatively impacted.
Mechanisms behind AI collusion in financial markets
AI algorithms differ from human traders in that they do not simply mimic human behavior. Traditional theories and experimental studies on human behavior fall short in explaining the actions of AI traders and the market equilibria they may form. AI operates with a distinct form of intelligence, where decision-making is guided by pattern recognition rather than emotions or logical reasoning, making it unaffected by higher-order beliefs. Therefore, understanding the dynamics of financial markets with AI-powered trading algorithms requires an understanding of algorithmic behavior that resembles the “psychology” of machines.5 This approach parallels how decision theory and psychology have provided valuable insights into modelling human behavior within an economic context.
Economic theories of learning in games provide valuable insights into mechanisms that may drive AI collusion in financial markets. As highlighted in our working paper, the emergence of AI collusion can result from either sophisticated artificial intelligence or from what can be termed "artificial stupidity," depending on prevailing market conditions.
In less noisy markets, algorithms can develop price-trigger strategies to maintain collusion, demonstrating artificial intelligence. Essentially, these algorithms learn to trade the asset as if adhering to an implicit agreement. However, if market prices become abnormal, the algorithms interpret this as a breach of the agreement, prompting all algorithms to trade aggressively in a punitive manner. These strategies align with theoretical mechanisms in dynamic games with imperfect monitoring, where a return to non-collusive competition occurs when the market price significantly deviates from its expected collusive level.6
In contrast, in highly noisy markets, algorithms struggle to accurately grasp the underlying environment. Their learning can be biased due to the inherent statistical limitations of basic reinforcement learning algorithms in a multi-agent dynamic system. Surprisingly, these biases, stemming from artificial stupidity, can lead algorithms to trade in a coordinated manner that ultimately increases profits for all participants. This steady-state equilibrium formed by AI algorithms resembles the experience-based equilibrium or self- confirming equilibrium found in economic theory.7
An intriguing aspect of AI collusion is that it does not require identical algorithms. Different algorithms can still learn to collude, albeit to varying degrees. However, algorithmic homogenization plays a crucial role in facilitating AI collusion, which can occur when algorithms are developed from similar foundational models, effectively creating a hub-and-spoke conspiracy.
Challenges and implications
Beyond financial markets, there are growing concerns and evidence of AI-driven market manipulation in retail markets globally. The U.S. Federal Trade Commission recently filed a lawsuit against Amazon, alleging that the company employed a secret algorithm to manipulate prices.8 Additionally, recent studies indicate that algorithmic pricing used by retail gasoline stations in Germany leads to anti-competitive outcomes.9
While there is currently no evidence of AI collusion negatively impacting financial markets, the threat remains significant as leading digital trading platforms increasingly endorse reinforcement learning-based AI trading bots. The potential for AI collusion in financial markets has raised concerns for the U.S. SEC and regulators worldwide. SEC Chair Gary GENSLER particularly cautioned against “the possibility of AI destabilizing the global financial market if big tech-based trading companies monopolize AI development and applications within the financial sector.”
References
1. Terrence Hendershott, Charles M. Jones, and Albert J. Menkveld, 2011, “Does Algorithmic Trading Improve Liquidity?” Journal of Finance, 66(1), pages 1-33
2. Nasdaq gets SEC nod for first exchange AI-driven order type, https://www. reuters.com/technology/nasdaq-gets-sec-nod-first-exchange-ai-driven-order- type-2023-09-08/
3. Jennie Bai, Philippon, Thomas, and Alexi Savov, 2016, “Have financial markets become more informative?” Journal of Financial Economics, 122(3), pages 625- 65
4. Itay Goldstein, Winston Dou, and Yan Ji, 2024, “AI-Powered Trading, Algorithmic Collusion, and Price Efficiency,” Working paper
5. Itay Goldstein, Chester S Spatt, and Mao Ye, 2021, “Big Data in Finance,” Review of Financial Studies, 34(7), pages 3213–3225
6. Edward Green and Porter, Robert, 1984, “Noncooperative Collusion under Imperfect Price Information,” Econometrica, 52(1), pages 87-100
7. Drew Fudenberg and David Levine, 1993, “Self-Confirming Equilibrium,” Econometrica, 61(3): 523–45.
Chaim Fershtman and Ariel Pakes, 2012, “Dynamic Games with Asymmetric Information: A Framework for Empirical Work,” Quarterly Journal of Economics, 127(4), pages 1611–1661
8. Amazon Used Secret ‘Project Nessie’ Algorithm to Raise Prices, https://www. wsj.com/business/retail/amazon-used-secret-project-nessie-algorithm-to- raise-prices-6c593706
9. Stephanie Assad, Robert Clark, Daniel Ershov, and Lei Xu, “Algorithmic pricing and competition: Empirical evidence from the German retail gasoline market,” Journal of Political Economy, 132, 723-771