Machine Learning will Create New Sales Bots on Trading Desks

Trading is moving to a point where there is almost no human involvement, as market participants employ algorithms that generate – and respond to – indications of interest (IOIs), according to Giuseppe Nuti, head of machine learning at UBS.

“I imagine that in the next few years, we’re going to see a lot more approaches coming up. So much so that it will even be algorithmic-to-algorithmic IOIs, almost removing the human component,” Nuti said.

Market-makers issue IOIs, or ‘axes’, to signal interest in buying or selling a security. IOIs allow counterparties to gauge available liquidity in the market without creating visible orders.

Nuti said that while it has taken around three decades to establish fully automated algorithmic execution, conversely, a lot of sales work has seen little transformation, especially for illiquid instruments.

More recently that has started to change though, Nuti said: “We’re working on software that allows us to help the salesperson connect with a client immediately. As soon as the trading desk has an axe, it pops up on the salesperson’s desktop saying, ‘look, we’ve analyzed all the data, we think these three clients are most likely to match up with our axe, would you mind calling them?’”

“The next iteration of that is sending it out directly,” he explained. “Obviously, indications of interest and axes have existed in the equity and credit spaces for a number of years. But what has changed is our ability, not just at UBS but our collective ability, to pinpoint where the interest lies and to minimize information leakage. I think the journey has really just started.”

Nuti was speaking in a pre-recorded debate for the FX Markets Europe 2020 event on December 10.

Speaking at the same event, Cetin Karakus, global head of quantitative and analytical solutions at oil and gas major BP, agreed that generating sales suggestions represents a sweet-spot for machine learning.

“It’s a perfect place for machine learning, to easily generate a lot of [sales] ideas,” said Karakus. “The fact that it hasn’t really happened so far is probably due more to the organization of human elements, rather than the technology part.”

Nuti said machine learning can be applied to order routing to increase fill ratios and reduce pre-trade market impact. The UBS foreign exchange, rates and credit trading team has found using supervised machine learning to route client orders through a warren of venues can cut the bank’s execution costs by a fifth.

Notoriously difficult to mitigate, market impact, or information leakage, is a common fear in request-for-quote (RFQ) trading and can be felt when rivals to a winning bidder on a trade then raise the cost of subsequently hedging that trade.

However, Nuti said that despite more than 20 years of speculation that the RFQ trading protocol used in less liquid instruments might be replaced by equity-like central limit order books, he doesn’t think technology will lead to diminished use of RFQ.

Exploration Versus Exploitation

Where machine learning will have a big impact is helping to understand more explicitly the exploration-versus-exploitation process, Nuti believes. A dilemma for systematic investors is knowing when to switch to new strategies in search of greater payoffs (exploration) or to stick with tried-and-trusted investments (exploitation).

Within order routing the problem occurs when better execution achieved by a particular route encourages more trades to be sent that way, but at the expense of further exploration. Hedge fund AHL has addressed the problem with a self-learning algorithm that directs trading based on past experience, rebalancing as it goes.

“Right now, we have a very large menu to choose from in terms of routing and sending orders, not only in terms of venues, but also in terms of protocols,” Nuti explained. “Should I do a request-for-quote for the whole thing or should I split the routing? What has been overlooked until recently is how we model our process of exploring the space. Because if an algorithm always sends orders to venue A and never tries anything else, we may think we’re doing the optimal thing, but we literally have no way of exploring that space.”

“It is an area which has received very little attention,” said Nuti, “and we need to focus explicitly on managing the exploration-exploitation trade-off.”

BP’s Karakus asked when Nuti might decide to risk exploring a space with a large order. “It’s a safe play to explore with a small size or chunk of order, but what is explored might be sub-optimal,” Karakus said.

“Absolutely,” said Nuti. “The approach of the past has been to explore with small orders or with house orders.” He added that a balance needs to be struck between “taking a crazy decision for the sake of exploring the space and proactively managing the [exploration-exploitation] trade-off”.

Editing by Alex Krohn