Clearly there are going to be areas of the capital markets where AI and ML will deliver more benefits than others, both in terms of business use and in different asset classes. For instance, they are already starting to make inroads in back office and compliance, while (for now at least) equity markets lend themselves more to these innovations than other assets. There are also many examples of institutions introducing “robo advisers”, as they are being dubbed, across wealth and fund management.
The FX journey
But what about foreign exchange? The FX market has never been slow to adopt new technologies to improve the speed, accuracy and efficiency of currency trading. Back in 1973, soon after many countries abandoned fixed exchange rates, electronic quotations of major currency pairs started to appear on Reuters screens – bringing transparency to the market and enabling banks around the world to execute trades with new confidence.
It wasn’t long before this evolved into electronic trading, expanding further as banks launched their own (EBS) platform to counter what was in danger of becoming a Reuters monopoly. But this still required traders to execute actual trades, taking and filling orders mostly by phone, but only gradually electronically. Nevertheless, this accelerated and the next quantum leap occurred in the late 1990s when black box algorithmic trading began over EBS as banks developed APIs to execute transactions automatically. Within a decade this accounted for around 60% of EBS turnover.
But individuals were still broadly in control. However, although volumes boomed (now in excess of $5 trillion a day, five times those seen 25 years ago and double that of little more than a decade ago) so did competition. Along with the greater transparency, these developments compressed spreads and drove down margins. Electronic trading is expected to account for over 75% of FX trades within the next year, from 66% in 2013 and 20% in 2001.
But costs still have to be tackled (the fines levied against banks for FX market rigging will further hasten the demise of the traders). The answer will be greater automation. Expensive headcount still needs to be replaced, but without threatening volumes. After all, for the big banking machines which still account for the lion’s share of the market, flow is everything.
At the same time, following the 2008 banking crisis, new regulations were heaped upon the banks which compounded cost pressures and drove up capital requirements, both of which further eroded margins. It was the perfect storm for the industry, creating just the right environment to accelerate its flirtation with automation in general and in particular its clever cousins, AI and ML. These had always threatened to change the face of banking but had never quite seemed able to fulfil their promise.
Enter AI and ML
Both AI and ML have been around for many years, but were mostly confined to niches because of cost and complexity. It might be that it is still too early for these capabilities to achieve widespread adoption, but evidence now points to wider traction across the banking and trading communities.
The sharp falls in the cost of data management and storage, and the wider availability of more sophisticated analytics, have started to make them more accessible. This has been complement by a growing pool of people educated in the higher levels of computer and data science.
Banks are therefore beginning to use AI and ML to tackle the proverbial “low-hanging fruit” where costs are highest and automation can be applied most successfully. Most notably this has occurred in back and middle offices, where for years manual activity has held sway. These tasks have multiplied due to the requirements of regulatory compliance. But now not only are functions being automated, both AI and ML are being successfully deployed to help deliver the hundreds of new regulations banks have to comply with each year.
In our own case we have just launched an AI/ML tool which can be applied alongside trade capture to anticipate potential costly errors and protect against both the need for subsequent corrections as well as potential reputational risk damage.
But the main challenge is the sheer volume of data now being produced and how to use it for the purposes required (to answer regulator expectations) as well as to deliver it to those who have to make more informed business decisions. According to IBM, 90% of the data in the world today has been created in the last two years, and is rising at 2.5 quintillion bytes of data per day. Banks do not have a great record when it comes to managing data, partly due to the siloed nature of their businesses, both in terms of departments but also asset classes. This has created fractured views of the business, where a holistic perspective is now required. To deliver that requires not only better technology infrastructure and analytics, but changes to business culture. Only once this fundamental hurdle has been cleared can more advanced steps be taken. Bank IT platforms need to be simplified to manage data analytics capabilities that can be used by IT personnel and lesser IT-skilled people, as well as being able to build high-performing applications across the enterprise. Once this is achieved, algorithms can be deployed with more confidence on the basis that their outcomes will be more reliable.
IT is why the attractions of AI are getting more support. Some would argue that these algos, along with machine learning, have already changed the financial markets given their apparent success in certain automated trading strategies. There is some evidence that they have been able to capture complex, non-linear relationships between factors that humans are unable to detect – as well as at much faster speeds.
In equities in particular, high-frequency trading (HFT) uses algorithms to execute trades in micro seconds, based on analysis and identification of pre-programmed conditions. Some believe there are days where this accounts for more than two-thirds of market volumes. Clearly this also has a significant impact on cost reduction, but it has also raised a number of practical and ethical issues due to the evolution of “dark pools” and the ability of the robots to potential “front run” other institutional orders in the market. But whether they consistently generate Alpha remains uncertain.
There are also questions asked about whether the algo-led computer trades feed off volatility or actually cause it. There is still no satisfactory answer. It is, however, expected that the newer breed of intelligent (ML-powered) computers will learn from past success and failures – much like humans.
Turning to foreign exchange, there is still scepticism that there is not yet the power required to run relevant and successful AI/ML initiatives in real-time, and that it is still too costly. This would suggest that there is unlikely to be a rush to use these techniques in trading. Nevertheless, there is no shortage of people willing to experiment, due to perceived longer-term competitive advantages.
Many of these initiatives focus on a combined AI system that integrates with neural networks and expert systems to support FX trading decisions. These are designed to try to simulate the traders’ knowledge with qualitative input from the expert system, along with quantitative input from a neural network that predicts currency prices based on its analysis of historic movements.
This is easier said than done, due to the enormous number of factors involved in moving currency prices, particularly macro-economic factors, geo-political developments and other news which can produce unexpected and short-term volatility. The decision by the SNB to stop artificially suppressing the value of the Swiss Franc in 2015 (which resulted in an almost immediate 15% appreciation) was a case in point – which would have wiped out most AI/ML-based positions.
In these instances, a neural network does not often work as expected because the market itself does not behave in the way the neural network has been trained to analyse the market. As has been said often, “neural networks are very good at finding patterns where the human eye can’t, but there is just too much noise in the FX market and too many surprises and no pattern to follow.”
In addition, while a neural network can have a relative degree of success trading around support and resistance lines, or perhaps in the immediate aftermath of a significant market-moving event, it cannot identify these levels by itself, nor can it interpret market-moving news and respond to it. Not yet, at least!
But other factors are playing into the hands of those in favour of greater FX trading automation, in particular new regulations like MiFID II. This is set to come into force next January and will drive the transaction of foreign-exchange derivatives onto exchange platforms and out of the OTC market, in order to increase transparency and ensure customer value. This, and the new Basel rules on market risk which will increase bank capital allocation against riskier transactions, will combine to make FX trading even more expensive.
Some believe this will drive more banks out of the mainstream FX market and encourage those that remain to invest more in technologies like AI and ML to maximise the shrinking margins that will be on offer. But initially this will probably mean more electronification and direct client connectivity rather than pure AI.
The real traction for AI in FX will arrive when there is more co-ordination between the FX traders and those writing the codes for the systems and the algorithms. There is still great expectation (hope?) that this combination will eventually deliver the computational capabilities of the machines with the strategic judgement and experience of the traders. There is no sign we are there yet.
On the asset management side, however, there are signs that funds are embracing AI and machine learning, where AI systems can not only analyse large amounts of data at speed, but continue to improve themselves through ongoing analysis. Hedge funds have been increasingly showing some successes in this regard, but although they rely on quants with algorithms examining complex models, many of these are static and a far cry from the dynamic volatility seen in FX.
There is also the risk that, once one investment bank or FX trading fund successfully delivers AI-led trading, that it would not be long before it was widely copied. These proprietary secrets rarely remain under wraps for long. As traders are known to remark, this is an area where you benefit not just from being smart, but being smart in a different way from the rest.
If a large enough segment of the market behaves in the same way, it becomes self-fulfilling in the way it then changes the market. The opportunity to arbitrage the difference will only last for a small window. Nevertheless, if the rewards are high enough – and with shrinking margins across so much elsewhere in the capital markets, even small opportunities are snapped up – there is no doubt that some investors will give it support. So while the elusive search for Alpha in foreign exchange trading will be likely to include an element of AI in the not-too-distant future, it might not deliver a sustainable point of difference. The challenge will be for these tools to demonstrate that they are robust enough to deliver meaningful and reliable benefits in the volatile world of FX. But as JM Keynes said with some precision nearly a century ago: “Markets can remain irrational longer than you can remain solvent.” And even AI might not be smart or rational enough to outsmart that logic.