Algorithmic trading: Flash crash or flashback?
19 February 2018
Earlier this month, when the Dow Jones fell by almost 5%, what followed quickly overshadowed any blame on human error. Prices in the US stock market experienced free fall over a period of 15-20 minutes. The shocks were felt sharply in the S&P 500, Japan’s Nikkei and, closer to home in the UK, the FTSE 100.
Flash crash or flashback? While algorithmic trading (algos) has the potential to improve liquidity and thereby reduce investor trading costs they can also rapidly magnify downward, or upward trends. As was starkly demonstrated in 2010 in the US stock market and in 2016 in the UK, when the pound fell by 9% overnight.
The issue hasn’t escaped the attention of global regulators, who are increasingly exploring ways to minimise the potential impact of algos on markets across all major asset classes. In Europe, ESMA’s MiFID II regulation, which came into force in January 2018, is a wholesale example of just that.
The algo provisions under MiFID II seek to prevent a repetition of past experiences; however, not without challenges. Algo firms may be forgiven for finding ambiguity in the practical implications of the regulatory rule book, particularly the pre-trade monitoring requirements. What other pre-trade requirements are there beyond those prescribed, which many firms can and already have in place? Is there a distinction to be made between investment and execution only algos in identifying and implementing pre- (and post-) trade controls or should they be aligned to the types of instruments traded? How sensitive should the parameters of the controls be to disorderly trends? These are just a few examples of questions being asked, which can be summarised as: How can the extents to which an algo might react to disorderly trends be anticipated, including malicious attempts by human traders to “bait” it? If the answer to this is known, the right controls can be deployed with adequate depth and sensitivity to disorderly trends to reduce “runaway” risk.
Advances in cost-efficient data processing now means firms can more comprehensively and precisely simulate investment or execution reactions to disorderly trends. Many algo firms already conduct varying levels of stress tests, particularly in relation to platform conformance and resilience. Yet, not many perform reverse stress tests to determine the impact of unanticipated scenarios and outcomes on markets. Such tests could also capture situations where human traders may use sophisticated strategies to lure and exploit an algo’s trend-following programming. Algos lack human emotion and judgment, which can make them more susceptible to falling prey to predatory trading strategies. The good news is that reverse stress tests, using data analytics, can inform control suitability, depth and sensitivity, to minimise risks to market impact.
Reverse stress tests can be delivered through machine learning and artificial intelligence (AI). Today’s technology can analyse huge volumes of data to automatically identify calibration requirements for pre- and post-trade controls, and detect control gaps as new risks are flagged. Applying machine learning and AI to the process of informing and managing controls requires sound management of the data inputs, human judgement and transparent validation, in other words, “responsible AI.”
Regulators in the UK have warned algo firms to tighten testing and monitoring, along with supervision and documentation. The PRA issued proposals to introduce a new code of conduct in June this year. The FCA also released its own expectations, echoing the PRA, and reiterating MiFID II requirements for mandatory “kill” switches, among other things, as a failsafe mechanism to prevent market impact. The stronger UK standards will sit closely with the Senior Managers Regime (SMR).
Sceptical or not, it seems now more than ever is the time to take a new approach and embrace innovative and emerging technologies as one way to safeguard against recent market events.