Bilateral ESRC/FNR: Experimental Assessment of the Societal Impact of Algorithmic Traders in Asset Markets

The has been proliferation of computerized algorithms traders (ATs) in electronic markets. This has generated considerable policy debate on whether the presence of ATs promotes or obstructs healthy markets. Further policy concerns are what steps - if any - should be taken to regulate AT behaviour and what responsibilities should be assign to market exchange providers regarding disclosures of AT participation, or whether they should be compelled to provide AT-free alternatives. Our study generates evidence to inform such policy debates and policy recommendations.

We use the methodology of experimental finance to gather this information. We conduct asset markets with financially motivated human traders and introduce AT's and information about their presence in a controlled way. Unlike traditional asset markets we control, and therefore know what the underlying fundamental values of assets are and what information each traders knows about these. Also, we have clear identification of what markets actions are taken by the AT's and their underlying strategy. We get further clean measurements by exogenously controlling the types of exchanges that are available to the traders.

We first assess whether retail and institutional investors are averse to having alternative types of ATs participating in the same markets. Alongside this analysis, we also measure the market impacts and wealth distribution effects of introducing AT-free "safe haven" exchanges.

We then turn our attention to arbitrage, a fundamental principle that drives market efficiency. AT's that seek riskless opportunities for instantaneous profit monitor multiple exchanges which facilitate the trade of the same asset. They can either look for attempt to create mispricing opportunities in which they purchase the asset in one exchange and sell it at nearly the same time at a higher price in a different exchange. In our experiments, we examine the impact of liquidity consuming ATs who arbitrage through market orders and liquidity providing ATs who arbitrage through limit orders. We also assess the impact of high frequency trading by varying the speed of the ATs between that at which humans can take market actions to those who act at sub-human speed.

There is a strong belief in the artificial intelligence and computer science community that a key attribute for ATs to outperform human traders, and therefore be a viable alternative for investors to use, is that they trade fast. In an efficient market speed leads greater opportunity to trade with misprices limit orders, and it also allows on to maintain the proprietary position in the order book to capture market at an advantageous price. The existing evidence supporting this belief comes from hybrid experiments/simulations in which equal numbers of ATs and human traders interact in a market. We introduce unbalanced designs, with ATs and human traders placed on opposite sides of the market, to measure whether this impacts aggregate performance of the market and traders. We also, vary the relative market power of the sides of the markets.

Finally we conclude with an assessment of the impact of placing Dark Pool markets, where assets are traded bilaterally with delayed price announcements, alongside public markets. Then we assess the impact on allowing predatory ATs into either of these markets. This allows us to address contemporary regulatory problems in which private providers of Dark Pool markets have a moral hazard in preserving the ATs safe haven they promise to deliver.

The role of algorithms in asset markets is a current and ongoing policy question. The objective of a financial market is to process and clear information from subjective valuations. Algorithms can only process the information available to them from the history of prices and any specific pricing rules programmed in to them. Hence a purely algorithm driven market could have a far lower information content than one driven by human beings incorporating all available information resources. The major difference being the speed at which new information is processed. The predominant research paradigm for assessing the impact of high frequency algorithms in financial markets has been to look at historical patterns in quoting and transactions. Unfortunately, (this research has provided little in the way of definitive results, we propose to conduct as series of laboratory experiments to assess the precise effect of different market structures and typologies of algorithmic trading. This research should provide foundational results that can be used as basic evidence for the regulation of markets.

The main, non academic, beneficiaries of the proposed research schedule are regulators and exchange providers such as the financial conduct authority, Bank of England and the London Stock Exchange. In addition over the next three years (the lifecycle of the project) a number of national and supranational legislative mechanisms will be drafted such as the European Commission Markets in Financial Instruments Phase II (MiFID II) and any follow up. Another interesting legislative adjustment may occur in the United States if there is a refinement on the regulation on National Market Systems (reg-NMS) which enshrines the best execution. The reg-NMS "order protection rule" requires that brokers execute against best price and not speed. A further decision to restrict co-location to authorized brokers in the US has also been enacted. So there is an active legislative agenda both in Europe and the United States. Our results will provide helpful evidence for future market design and any implementation of restrictions on trading.

Our results will also be of interest to market participants. Market participants are often divided into two groups, liquidity demanders (those who seeks a specific position in a security, long or short) and liquidity providers (those who trade to provide positions to liquidity demanders and charge a small fee for this service). Liquidity providers, such as NASDAQ market makers provide near automated services at high speed, but generally are not trading to obtain a particular position. The difference between a high speed market maker and a high frequency trading algorithm can be quite small, both are designed to seek out arbitrage opportunities between quotes. Hence, results that allow the discrimination between typologies of algorithm will be useful in ensuring that regulation is sufficiently nuanced to ensure that transaction costs are not increased. For instance, the THOR trading algorithm (Royal Bank of Canada) has received some publicity as it forces stock quotes to be displayed at the exact same time, even thought this might be a fraction slower than the best execution time. Our results can be used to establish if this type of algorithm "hobbling" approach is the most appropriate in financial markets.

Finally, our results will also be of interest to the psychology community that looks at the interaction between human beings and algorithms. Algorithmic trading is one of the few clean experiments that can be designed where there is a clearly delineated interaction between different types of algorithms and humans.