Future trading robots will be able to adapt and learn with little human involvement in their design Far fewer human
2 The impact of technology developments
2.4 Technology advances likely in the next ten years
2.4.3 Computer designed trading algorithms that adapt and learn
As De Luca et al. discuss at length in DR13, researchers have been studying and refining adaptive trading algorithms since the mid-1990s and, in 2001, IBM showed two such algorithms to be capable of outperforming human traders. Adaptive trading algorithms are automated trading systems that can learn from their experience of interacting with other traders in a market, improving their actions over time, and responding to changes in the market. Since the late 1990s, researchers have also studied the use of automated optimisation methods to design and improve adaptive trading algorithms. In automated optimisation, a vast space of possible designs is automatically explored by a computer program: different designs are evaluated and the best-performing design found by the computerised search process is the final output. Thus, new trading algorithms can be designed without the
involvement of a human designer20. The use of these techniques in the finance industry looks likely to
grow over the next decade. This is a development that is enabled and accelerated by the step change drop in cost of HPC offered by cloud computing service providers, and by the huge speed increases offered by custom silicon.
Because these next generation trading algorithms may have had little or no human involvement in their design and refinement, and because they operate at truly superhuman speeds, the behaviour of any one such automated trader may be extremely difficult to understand or explain, and the dynamics of markets populated by such traders could be extremely difficult to predict or control. The concept of adaptive trading algorithms, whose behaviour may be difficult to understand, and whose dynamics in conjunction with other traders, whether automated or not, may be difficult to predict or control, will be quite disturbing. One natural instinct is to try to ban all such algorithms from one’s own market. But that will not work, partly because the definition of what constitutes an adaptive algorithm is itself ill defined, and partly because markets will either exist, or can easily be set up, which will allow such trading practices. Once such markets exist, and they will, any market crashes or other disturbances in prices will reverberate directly back to the markets in which such algorithms are banned. Rather the need is to develop a culture and understanding which serves to limit such dangers.
Other studies, such as those discussed in DR25 and DR27, are needed to increase our understanding of market dynamics in the presence of automated trading systems that are adapting over time
and whose designs are the result of computerised optimisation processes. Furthermore, as Cliff & Northrop describe at length in DR421, there are likely to be valuable lessons to be learned from
fields other than finance or economics. In particular, sociological studies of how people interact with technology that is ‘risky’, in the sense that its safe operating limits are not fully understood in advance, have revealed that it is important to be mindful of the dangers of falling into a pattern of behaviour, a process called ‘normalisation of deviance’. When a group of people engage in this process, they gradually come to see events or situations that they had originally thought to be dangerous or likely to cause accidents as less and less deviant or dangerous, and more and more normal or routine. This greatly increases the risk of major accidents. Normalisation of deviance is discussed in more depth in Chapter 4.
Normalisation of deviance can be avoided by adopting practices from what are known as high reliability organisations (HROs). Examples of HROs that have been studied include surgical teams, teams of firefighters and aircraft-carrier flight deck operations crews. In all such groups, safety is a critical concern, and it has been found that HROs have often independently developed common deviance monitoring processes involving careful post-mortem analyses of all procedures, even those in which no problems were apparent, conducted confidentially but with an internal atmosphere of openness and absence of blame. Adoption of HRO working practices and other approaches developed in safety- critical engineering (such as the use of predictive computer simulation models, as discussed in more depth in DR4, DR14, and DR1722) would help to mitigate possible ill effects of increased reliance on
risky technology in the financial markets (for further discussion see Chapter 4).
20 Cliff (2009).
21 DR4 (Annex D refers).
2.5 Conclusions
It is reasonable to speculate that the number of human traders involved in the financial markets could fall dramatically over the next ten years. While unlikely, it is not impossible that human traders will simply no longer be required at all in some market roles. The simple fact is that we humans are made from hardware that is just too ‘bandwidth-limited’, and too slow, to compete with new developments in computer technology.
Just as real physical robots revolutionised manufacturing engineering, most notably in automobile production, in the latter years of the 20th century, so the early years of the 21st seem likely to be a period in which a similar revolution (involving software robot traders) occurs in the global financial markets. The decline in the number of front line traders employed by major financial institutions in the past few years as automated systems have been introduced is likely to continue over the next few years.
Nevertheless, there may be increased demand for developers of automated trading systems, and for designers of customised computer hardware that runs at high speed and with low latencies. It is most likely that the skills needed in these designers and developers will be those learned in advanced undergraduate and postgraduate degree courses. From a UK perspective, serious investment in coordinated programmes of research and development (for example, funded by the UK research councils or the UK Technology Strategy Board)23 could help to secure the future ‘talent base’ (i.e.
the pool of trained scientists and engineers that have skills appropriate for work on advanced trading algorithms and hardware).
The increased reliance on CBT is certainly not without its risks. Sudden market crashes (or sudden bubbles) in the prices of financial instruments can occur at greater speed, with chains of events proceeding at a much faster pace than humans are naturally suited to deal with. Furthermore, the globally interconnected network of market computer systems arguably means that an adverse event in one market now has greater potential to trigger a wave of contagion that could affect markets around the world (see Chapter 4 for further discussion of systemic risks of CBT). Equally, natural hazards in the form of floods or electrical storms can incapacitate data centre telecommunications networks; an interplanetary coronal mass ejection (a huge burst of plasma and electromagnetic energy from the sun, causing a geomagnetic storm when the ejection reaches the Earth) could seriously disrupt or disable the electrical power systems and radio communications of an entire city or larger area. Ensuring the resilience of critical computer and communications systems in the face of such major natural hazards, and in the face of attack from cyber criminals, must always remain a priority.
Even in the absence of such exogenous hazards, there are serious issues to be addressed in dealing with major endogenous risks, such as destabilising systemic internal feedback loops, and the pernicious effect of normalisation of deviance in risky technology systems, both of which are discussed further in Chapter 4. In addressing issues of systemic risk and financial stability, it is particularly important for regulators to develop and maintain the capacity to analyse extremely large data-sets. The challenge of dealing with ‘big data’ is already severe enough for any single major financial trading venue (for example, an exchange or dark pool): the number of active HFT systems and the rate at which they generate (and cancel) orders means that the venue’s order book showing the best bid and offer for a particular stock may need to be updated thousands of times per second. A slight shift in the price of that stock can then cause the prices of many tens or hundreds of derivative contracts, such as options or exchange traded funds, to need to be recalculated. For regulators, the issue is made even more difficult as they are required to deal with aggregate market data generated simultaneously by multiple trading venues, as part of their regulatory role to oversee fair and orderly market systems, to identify policy violations and to monitor the need for policy revisions. Well designed mechanisms to manage the risk of volatility are one means of reducing the effects of risk in today’s and future markets.
23 The UK does already have significant investment in relevant PhD-level research and training, with doctoral students from various initiatives such as the UK Doctoral Training Centre in Financial Computing and the UK Large-Scale Complex IT Systems Initiative (www.financialcomputing.org and www.lscits.org, respectively) working on issues of direct relevance to the topics addressed in this chapter. Despite this, there has been comparatively little investment in coordinated programmes of relevant academic research among teams of post-doctoral level research workers; that is, major research projects, requiring more effort than a single PhD student working for three years do not seem to have been prioritised.
On the basis of the evidence reviewed in this Chapter, it is clear that both the pace of development of technology innovations in financial markets and the speed of their adoption look set to continue or increase in the future. One important implication of these developments is that trading systems can today exist anywhere. Emerging economies, such as those of Brazil, Russia, India and China, may capitalise on the opportunities offered by the new technologies and in doing so may possibly, within only a few decades, come to rival the historical dominance of major European and US cities as global hubs for financial trading.
The next chapter builds on the information provided here to assess the impact of CBT on liquidity, price efficiency/discovery and transaction costs.