Machine Learning Could Make FX More Efficient
The Bank for International Settlements, the global regulator, said the increased use of machine learning in high-frequency foreign exchange trading could lead to more efficient markets, particularly the timely incorporation of diverse sources of data in market pricing.
The BIS said in a report, Monitoring of fast-paced electronic markets, that a good understanding of artificial intelligence and machine learning is increasingly important for the effective monitoring of fast-paced markets and will require a change in risk management techniques.
Central banks adapt the way they approach market monitoring as #ForeignExchange and other fast-paced electronic markets undergo structural changes: Markets Committee https://t.co/5oM4KZbydO pic.twitter.com/1QkrU5Qf4q
— Bank for Intl Settl. (@BIS_org) September 17, 2018
The study highlighted the sterling flash event of 7 October 2016 and the flash rally in US Treasuries on 15 October 2014 as demonstrating the importance of close monitoring and timely analysis of high-frequency markets, particularly foreign exchange, by central banks.
For example in spot FX, the share of trading volume executed electronically has almost doubled over the last decade and has become increasingly fragmented across a range of new venues. The report said that more than 70% of spot trading is executed electronically since 2013, while an estimated 70% of orders on EBS, a primary central limit order book and a major inter-dealer platform for spot FX, are now submitted by algorithms, and not manually.
“Greater electronification has led to the creation and commoditisation of large quantities of high-frequency data,” added the BIS. “The use of artificial intelligence and machine learning in trading algorithms, while nascent, also has the potential to introduce new market dynamics and increase complexity.”
Most of the machine learning applications in high-frequency trading have been focused on equity markets, some on fixed income, while applications in foreign exchange are at an early stage. “Nevertheless, research and experimentation is afoot, and lessons from equities will carry over to FX markets,” said the BIS.
For example, one bank has used decision tree algorithms on central limit order book data to forecast order flow direction over the next 20 or more ticks and reinforcement learning has been used to train a robot to identify trading strategies that will reduce market impact in given market conditions. Some market makers have also used electronic tools to review the profitability and volume of transactions with clients.
However the BIS also warned that use of machine learning could push lower-tier banks further towards an agency approach to risk management where they use third-party technologies without sufficient controls and governance. “Multiple participants using similar algorithms simultaneously could lead to herd-like behaviour,” added the report.
In addition, technical expertise in programming and analytics will be needed to validate machine learning algorithms. As a result some market participants have highlighted a need to shift to using segregated controls and the ability to limit market access.
“Market access controls can act as kill switches, but are usually applied on a graduated basis, imposing limits such as on price or aggregate volume, as well as the number of, duplication and size of orders being sent to trading venues,” said the BIS.
MIFID II data
The regulator continued that the introduction of the the MiFID II regulations in the European Union at the start of this year has created a range of new data which is being collected by authorities, platforms and data repositories.
Increased electronification and speed of activity in markets generate large amounts of high-frequency data, which can raise barriers to entry but also improve pricing and trade execution: Markets Committee https://t.co/5oM4KZbydO pic.twitter.com/rBP690NwyL
— Bank for Intl Settl. (@BIS_org) September 17, 2018
Although the use of this new data seems promising for the purposes of market monitoring the BIS said there are three challenges – spot FX markets are not included in the MiFID II legislation; the frequency of collection is not well defined and there are many exemptions; and the lack of centralisation and aggregation of the data.
“A major difference is that trade reporting needs to be sent in near real-time and is to be made public, while transaction-level data are highly sensitive and remain non-public,” said the BIS. “After being submitted at a t+1 requirement, transaction data are stored at a very high level of disaggregation.”
The study continued that private initiatives to store and make the post-trade data available centrally are emerging in the form of consolidated tape providers. “Such data will probably not be available free of charge and a 100% coverage is unlikely to be achieved,” added the BIS.
As a result of the acceleration in electronic trading and fragmentation in trading venues, large quantities of data are created by liquidity providers and trading platforms. The report cited one large bank’s e-trading desk producing around one billion FX quotes per day for clients globally.
“In order to deal with these data volumes, firms have begun using cloud services, which reduce the need for hardware and physical data storage,” said the BIS.
AiPEX with Watson simulates a team of analysts and traders to identify potential investments.
Machine learning models systematically scan newly arriving, anonymized data to identify anomalies.
The Cobalt programme launched in 2018 to help fintechs collaborate with the asset manager.
Users with different skill levels will be able to undertake machine learning and advanced analytics.
Fintech CEO discusses how AI applications have fared in highly volatile fixed income markets.