Artificial intelligence and, to a lesser extent, machine learning, have reached buzzword status in the financial sector – they sound interesting in marketing announcements and are good for SEO, but technologists remain suspicious as to their practical applications.
Things like chatbots and RPA are now widely used to automate mundane tasks such as text and voice recognition, natural language processing and data processing. Elsewhere, quantitative funds have a deep history with machine learning, as applied to data analysis and trade signals.
For a discretionary fund, however, particularly one with fewer resources to invest in data scientists, the applications of AI and machine learning are in doubt.
The concerns around AI
Nowhere else is the scepticism more pronounced than in the applications of AI for trade execution.
Technologists have good reasons to be wary. As terms like AI have gained traction, companies have wised up to using them in their product description – even if actual machine learning is non-central, or even non-existent, in their proposition.
According to a highly publicised survey by London-based market research firm MMC, 40% of European start-ups purporting to use AI, do not have any technology that can be described as such, or use it in a way that is not central to their product offering.
These technologies have been criticised for lacking the personalisation, emotion and instinct required to decipher market developments, particularly with challenging situations; hedge funds voiced a very similar complaint when quizzed by HFMInsight in 2017 about AI and machine learning more generally.
The difficulty in separating the wheat from the chaff means that smaller IT and dev shops – the natural first adopters of AI-infused tools – often prefer to steer clear of them altogether.
Meanwhile, funds heavily reliant on in-house development for their trading applications, favour using developer time in better tested or more urgent solutions. Machine learning isn’t typically in the DNA of their dev teams and so naturally comes lower on the list of priorities.
“I think artificial intelligence is still in the early stages,” said the credit fund CTO. “You definitely come across a lot of vendors which have machine learning written all over them, but if you really drill down into the core concepts, the functionality is not that different from what a classic algo would do.”
The RoI is also not significant enough to warrant that kind of shift. The feeling among the fund technologists HFM Technology spoke to is that artificial intelligence – in its currently available form – is not exactly the messiah they’ve been promised.
“And you many times find they can quite precisely explain how the algorithm works, because with a true AI, you wouldn’t really know how it comes to make the decision. So many times the decision would be rules-driven, rather than AI-driven,” says the technologist.
The technologist cites explainability – a controversial concept related to artificial intelligence. It refers to the idea that the more easily explainable an AI algorithm is, the less sophisticated and the farther from true AI it is- or so the argument goes.
General AI vs narrow AI
Such perceptions of machine learning are not uncommon. In fact, they underpin one side of longstanding industry debate regarding the scope of AI advancement.
Existing iterations of artificial intelligence available off-the-shelf largely fall into the category of “narrow” or “weak AI” – a self-correcting algorithm applied to a single task. Meanwhile, the technology required to meaningfully automate execution is more akin to “artificial general intelligence” or strong “AI” – cognitive processes similar in number and complexity to human thought.
Take conversational assistants as an example – an area that has made arguably the biggest strides in recent years.
Conversational assistants – aka chatbots – cannot “yet meet user expectations related to sensing and responding with emotion,” according to David Konopnicki, an IBM research manager who studies affective computing.
“These limitations exist because computers have not yet made the great strides in natural language understanding and dialogue that they’ve achieved in NLP. Without this, most computer responses are painstakingly scripted by engineers using if-then rules,” Konopnicki commented in a recent IBM article.
To carry out a “natural” conversation, a bot would need to understand and contextualise a human’s prompt, interpret the available information, formulate an appropriate response and repeat this process for all subsequent commands.
A similarly complex process is required for a trade execution robot, for example, to be fully autonomous – the robot would need to interpret the conditions, select the execution broker, interpret the execution result and then update its model accordingly for every trade.
Few, if any, execution management vendors today would claim their tools have this level of sophistication.
AI for trade execution
While general AI might be a long way off, the execution market is moving in the direction of more intelligent automation.
In the past year alone, multiple firms looking to capitalise on the hype have released ML-based tools. The applications – though strictly narrow AI – are creative.
US-based Imperative Execution bills itself as an “operator of trading venues that help reduce execution costs.”
“A venue can pair up buyers and sellers in such a way that the market impact is as small as possible,” explains CEO Roman Ginis. “This is what our solution, Intelligent Cross was built to do. AI helps in the matching process, to pair up certain kinds of buyers and sellers to help achieve the outcome we are looking for – i.e. the price to remain as stable as possible during the execution.”
“This has only been possible with AI,” claims Ginis. “Historically, exchanges and MTFs [multilateral trading facilities] have matched orders as soon as they arrive. When you go to scheduled mapping, you can only schedule certain types of orders to pair up – those who have enough ‘patience’ to rest on the book. The schedule controls the urgency of interaction and the patience of the orders.”
Ginis, whose company uses AI on the venue side to achieve cheaper execution, believes that the main hindrance to accepting AI as commonplace is more practical – execution data is particularly difficult to collect, especially in the large quantities needed to train an AI algorithm.
“Applying AI to trading has always been one of those things that people believe is a decade away. This is because in order to make AI possible, you need to have a large number of data points. And once you apply it, you need to have a measurable outcome to be able to judge the efficiency of the application,” explains Ginis.
The execution data landscape is changing
However, a year on from the enforcement of Mifid II reporting, data availability is less of an issue than it would have been previously. This may help to explain the proliferation of AI-based best execution tools flooding the market.
A recent HFM Technology poll showed that the Mifid II landscape was positive – technologists were both aware and generally satisfied with the tech service providers out there. A generally smooth shift towards more comprehensive trade reporting means that, a year on, the execution data exists – and is only becoming easier to access.
The conditions are there for the wider adoption of artificial intelligence in trade execution. Reaching the tipping point, however, will require a much more precise understanding of the applications and benefits of intelligent algos – both from the buyside and from the vendors promoting the technology.