The Impact and Interactions of Artificial Intelligence Pricing Algorithms in Dealer Market

研究了人工智能定价算法的相互作用及其对交易商市场做市价格的影响。在交易商市场中，交易商，也称为“做市商”，发布他们愿意购买或出售资产的价格。债券和外汇现货主要通过交易商市场进行交易。纳斯达克是股票交易商市场的典型例子。我们感兴趣的是做市商将人工智能应用于金融工具的算法定价的案例，特别关注这些算法的相互作用及其对价格的影响，无论它们是否会导致默契的共谋。这些算法允许做市商通过反复试验从市场数据中学习来自动定价。随着人工智能算法越来越多地被纳入金融行业，监管机构和市场参与者越来越担心算法的这种相互作用是否会导致不良结果，即使算法在设计上并不打算这样做。本研究项目的目标是解决这些问题，并为市场监管机构提供在经销商市场使用人工智能算法的参考。首先，我们建立了一个数学框架，为做市商在经销商市场，然后引入多智能体强化学习算法，每个做市商通过定价游戏来设定价格。在离散时间设置下取得了一些初步的实验结果，我们正在努力将模型扩展到连续时间，其中随机控制模型和多智能体强化学习将相互干扰。我们的目标是开发理论框架，以解决人工智能算法在做市中的相互作用。有放大的文献算法交易从随机控制的角度。算法定价和人工智能算法的研究起步较晚，近年来的研究主要集中在商品市场中的默契合谋。然而，尽管监管机构越来越关注金融交易商市场的人工智能定价算法，但仍然没有现有的研究。然而，没有理论解释人工智能定价算法可能的相互作用。因此，我们正在进行的研究是新颖的。该研究项目福尔斯以下EPSRC研究领域：人工智能技术，非线性系统，数值分析，运筹学，统计学和应用概率。本研究项目与汇丰银行的外汇eRisk团队合作。

We study the interactions of artificial intelligence pricing algorithms and their impact on market making prices in dealer market. In a dealer market, dealers, also called 'market makers', post prices at which they are willing to buy or sell assets. Bonds and FX spots are primarily traded via dealer market. Nasdaq is a typical example of equity dealer market. We are interested in the case where market makers apply artificial intelligence in algorithmic pricing of financial instruments, with a specific focus on the interactions of these algorithms and their impact on prices whether they could lead to tacit collusion. These algorithms allow market makers to automate pricing by learning from market data through trial and error. As artificial intelligence algorithms are increasingly incorporated in financial industry, there are growing concerns from regulators and market participants whether such interactions of algorithms could result in undesirable outcomes, even though the algorithms are not intended to do so by design. The objective of this research project is to tackle these concerns and provide references for market regulators on the usage of AI algorithms in dealer markets. We start by setting a mathematical framework for market makers in dealer market, and then introducing multi-agent reinforcement learning algorithms for each market maker to set prices through a pricing game. Some preliminary experimental results are achieved under discrete-time settings, and we are working at extending the model to continuous-time, where stochastic control models and multi-agent reinforcement learning will intertwine. We aim at developing theoretical framework to tackle the interactions of AI algorithms in market making. There is amplified literature on algorithmic trading from stochastic control perspective. The topic of algorithmic pricing and AI algorithms have been started only a few years ago, in which recent research focuses on tacit collusion in goods market. However, there is still no existing research on AI pricing algorithms in financial dealer market despite growing focus from regulators. Yet there is no theory explaining the possible interactions of AI pricing algorithms. The research we are working on is therefore novel. This research project falls within the following EPSRC research areas: Artificial intelligence technologies, Non-linear systems, Numerical analysis, Operational Research, Statistics and applied probability. This research project is in collaboration with FX eRisk team in HSBC.