Experiments on the effects of trading algorithms on financial markets

算法交易，即计算机而不是人类在收到信息后的毫秒内执行交易，多年来一直在增长。一些作者估计，美国70%的交易是由算法执行的(Swinburne，2010)，其他人高达85%(Glantz&Kissel，2013)。因此，算法交易对所有发达金融市场产生了重大影响，但其影响尚未得到很好的认识；对这种相对较新且不断变化的现象有更好的理解，对于政策辩论至关重要，比如算法交易是否需要遵守更严格的规则，以防止过度波动或资产泡沫的形成。这个项目的目的是回答有关算法交易对金融市场的影响的基本问题；具体地说，由最先进的机器学习技术训练的算法是增加还是降低资产泡沫的可能性，它们是增加还是减少市场的波动性和风险，以及这些算法最初是什么样子的。后一个问题很简单，但不容易回答，因为成功的交易算法很有价值，因此会保密，以免被复制、利用或监管。另一个困难是，众所周知，利用金融市场的现场数据很难建立因果关系，部分原因是，通常甚至无法确定匿名市场中的算法执行了哪些交易。出于这些原因，我将通过实验室实验来回答这些问题，我让人类交易员和算法交易员在小型实验室市场进行互动。这种方法的优势在于，可以令人信服地建立因果关系：可以将只有人类交易员的对照组与根据各种结果指标(如泡沫频率、波动性/风险或流动性/交易速度)进行算法交易员的处理组进行比较，所有其他因素都相同。此外，每一个动作都可以在实验室中观察到，因此特定的交易可以被归类为由算法发起或由人类发起。它还允许我调查在与人类交易员或市场上的其他算法进行学习时，哪些类型的算法会进化。实验室中使用的一类市场将允许受试者/算法交易金融资产的多个单位。这项资产在每15轮结束时向持有者支付股息。由于在后几轮中剩余的股息收益较少，资产的基本价值正在随着时间的推移而减少。然而，在之前的人类交易员实验中，在资产价值下降的情况下，价格却大幅上涨的情况下，经常会出现金融泡沫，这也是在股市或房地产市场中观察到的泡沫模式。因此，这种实验市场环境对研究泡沫的形成和泡沫的原因是有用的，但到目前为止还没有关于算法交易在这一背景下的影响的研究。在另一类实验市场中，我将介绍关于金融市场交易中资产价值的新闻，并观察交易员和算法如何应对此类新闻冲击。特别是，主要的问题是，足够数量的趋势跟踪算法是否会导致价格大幅下跌或上涨，从而增加市场的波动性和风险，或者算法是否真的可以稳定市场。研究结果将为政策制定提供参考，并具有重大影响潜力。确定算法在金融泡沫和“闪电崩盘”等市场事件中扮演的角色，将有助于金融和监管部门决定是否需要创新或监管改革。在这些事件中，价格在没有明显原因或消息的情况下迅速下跌，但很快就会反弹。这些实验也是检验新政策措施或法规有效性的一个很好且廉价的工具。

Algorithmic trading, where computers rather than humans execute trades within milliseconds of receiving information, has been on the rise for years. Some authors estimate that 70% of trades in the US are executed by algorithms (Swinburne 2010), with others going up to 85% (Glantz & Kissel 2013). Algorithmic trading has thus had a significant impact on all developed financial markets, but its effects are not yet well recognised; an improved understanding of this relatively new and shifting phenomenon is essential to inform policy debates, such as whether algorithmic trading needs to comply with stricter rules to prevent excess volatility or the formation of asset bubbles. The aim of this project is to answer fundamental questions on the effects of algorithmic trading on financial markets; specifically, whether algorithms trained by state-of-the-art machine learning techniques increase or decrease the likelihood of asset bubbles, whether they increase or decrease volatility and risk in markets, and indeed how such algorithms look like in the first place. That latter question is simple yet not easy to answer, because successful trading algorithms are valuable and therefore kept confidential for fear of them being copied, exploited, or regulated. Another difficulty is that causal effects are notoriously difficult to establish with field data from financial markets, in part because it is typically not even determinable which trades are executed by algorithms in anonymous markets.For these reasons, I will answer these questions with laboratory experiments, where I let human traders and algorithmic traders interact in small lab markets. The advantage of the method is that causal effects can be convincingly established: A control group with only human traders can be compared to a treatment group with algorithmic traders along various outcome measures such as bubble frequency, volatility/risk, or liquidity/trading speed, all else equal. Moreover, every action is observable in the lab, so specific trades can be classified as being initiated by algorithms or humans. It also allows me to investigate which kinds of algorithms evolve when learning against human traders, or against other algorithms in the market.One class of market used in the lab will let subjects/algorithms trade multiple units of a finanical asset. This asset pays out a dividend to the holder at the end of each of 15 rounds. Because there are fewer dividend payoffs remaining in later rounds, the fundamental value of the asset is decreasing over time. Yet in previous experiments with human traders, financial bubbles frequently arose where prices increased dramatically despite the decreasing asset value, which is the bubble pattern also observed in stock or housing markets. This experimental market setting is therefore useful to study bubble formation and causes of bubbles, but so far there is no research on the effect of algorithmic trading in this context.In another class of experimental market, I will introduce news about the value of assets during financial market trading, and observe how traders and algorithms respond to such news shocks. In particular, the main question is whether a sufficient number of trend-following algorithms can lead to drastic drops or increases in prices, thus increasing volatility and risk in markets, or whether algorithms may actually stabilise markets.The research findings will inform policy formulation and have significant impact potential. Determining the role that algorithms play in market events such as financial bubbles and 'flash crashes', where prices drop swiftly without apparent cause or news, only to bounce back shortly afterwards, will assist the finance and regulatory sector in deciding whether invention or regulatory changes are required. The experiments are also a good and inexpensive tool to test the effectiveness of new policy measures or regulations.