A.I. impacting buy side’s trading counterparties
The lesson from this massive buy side: not every broker’s algos compete when machines make more decisions.
The introduction of artificial intelligence into the equities trading desk at J.P. Morgan Asset Management is influencing the sell sides it trades with, says Lee Bray, head of trading for Asia Pacific in Hong Kong.
“It’s impacting our use of counterparties significantly,” he told DigFin.
This doesn’t necessarily mean the same sell-side brokers are destined to win all flows as more buy sides add A.I. to their trading decisions, because only so many firms deal at J.P. Morgan A.M.’s scale.
“Globally, we’re doing 1% of global flow daily through our asset-management function,” he said.
He declined to name the sell sides benefiting more from interacting with his team, but says the firm is increasingly sending trades to those firms whose algorithms are deemed best by J.P. Morgan A.M.’s proprietary A.I. tools.
Going long the machine
Bray joined the firm in London in 1999, and moved to Hong Kong in 2013 to manage all trading responsibilities. For the past two years, the business in Asia has developed machine learning to automate and measure trading in the most liquid stocks, and recently put this into effect.
The heart of machine-learning tools (developed by the firm in Asia for global deployment) is to assess the many algorithms that brokers will suggest to implement a given trade.
By learning from a history of outcomes, the machine directs J.P. Morgan A.M.’s flow to the algo with the highest probability of whatever outcome the firm wants. The need for lots of data points from many trades, which feed information to the machine, makes the technique suited to larger asset managers. “It’s not about relying on a few trades to make a decision,” Bray said. “You need the right kind of inputs.”
This also works best for equities, as opposed to bonds, because equities are already traded electronically, while most bond deals are conducted by human traders over the phone. The algo parameters are already stored where the machine can access them. Similarly, machine learning isn’t being applied to traditional sales trading.
“We use it for liquid orders,” Bray said, suggesting this meant trading up to 15% of a given stock’s average daily volume without impacting its price.
Made by Mifid
Fundamental to honing machines to picking the best algos is regulation, as well as tools developed for the first wave of electronification of equities.
The regulation in question is Mifid 2, the European directive on financial instruments, which in 2014 required market participants to execute client orders for the best price, regardless of other commercial relationships. Initially large buy sides reacted by “unbundling” their trades from sell-side perks like research, but it has also freed quant desks to focus purely on trading.
“Mifid put the onus on asset managers to decide which counterpart provides best execution,” Bray said.
They are developing machine-learning models based on methodologies that have been in play for a number of years. One is transaction-cost analysis, or TCA, which rates the spreads between two potential trades to figure out how efficiently a desk is trading. Another is implementation shortfall: the difference between the decision price sought by the buy side, and the actual execution price. I.S. continues to serve as the gauge of a trader’s performance, but the new wave of technology is changing how it is measured.
Bray says the firm continues to develop its A.I. tools. The next focus is indications of interest, or IoIs, when sell sides advertize business they have at their firms. IoIs can be very helpful to asset managers looking for counterparties to take the other side of a trade. They can also suffer if the market perceives an asset manager is is trying to unload a position – but that occurs when brokers act as principals rather than on behalf of other clients, and J.P. Morgan is developing IoI analytics for agent-related activities, which account for the majority of IoIs.
Therefore using A.I. techniques to analyze data can help the asset manager make smarter decisions about how and when to respond, based on track records of brokers’ usage of IoIs. These can help an asset manager decide which counterparts have the best flow, and who have a track record of acting in their best interests. “We’re still reviewing models for this,” Bray said. “We get one million IoIs every six months…By looking at this concept, we think we can extend what we learn to other parts of the business.”