Machines can save 20% on routing orders, UBS says

Machine learning is navigating the labyrinth of venues, figuring out how best to land the trade

The UBS forex, rates and credit trading team has been using supervised machine learning (ML) to route client orders through a virtual maze of venues, shaving as much as a fifth off the bank’s execution costs.

Giuseppe Nuti, a managing director with the Strategic Development Lab in UBS’s foreign exchange, rates and credit business, says the group has trained its machines how to optimally route and trade so-called child orders, or ones that have been split off an initial ‘parent’ order. The Lab is made up of technologists, scientists and programmers looking for ways to leverage data and technology.

The routing project is in contrast to trying to schedule trades. That involves calling the market, something machines can’t really do.

Routing, therefore, “is where we have been most active, with the basic thesis that it is very hard to predict the direction of the market with any kind of accuracy”, says Nuti. “Conversely, really focusing on optimal routing can allow savings of easily 10% to 20% of the mid to trade cost, which is ultimately going to be translated in 10% to 20% saving for the parent execution.”

The proliferation of trading venues over the past decade makes order-routing an ideal problem for the machines to crunch, Nuti says. Today, there are exchanges, electronic communication networks (ECNs) and more, each with a unique microstructure that governs trading protocols. The jumble makes it difficult to compare prices from two different venues, he explains.

In the forex market, for instance, some of the newer execution venues focus on streaming prices, where market-makers tailor a price stream for one or more specific market-takers. These venues tend to follow ‘last-look’ protocols where the market-maker can hold a trade request for a few milliseconds before deciding whether to accept it.

There are also traditional all-to-all venues that resemble exchanges, he says, where bids and offers are stored in a central limit order book to match with other orders. He adds that these venues – EBS and Reuters form the bulk of the primary markets in spot forex – also have a matching approach that oddly resembles last-look, giving anyone canceling an order priority over other instructions, like placing an order.

Comparing a firm exchange to an ECN with last-look, the firm venue should ordinarily get the order.

“But the more interesting question is, when the difference in price is quite limited, is that sufficient to compensate the routing to a non-firm venue?” he says. “Those are the sorts of interesting questions on the micro level of routing that we have focused a lot of work on.”  

Giuseppe Nuti

Nuti tries to avoid what he calls the “kitchen-sink approach” to ML, where loads of variables are dumped into an equation, possibly overfitting to the data. Most ML techniques are based on ‘input space partitioning’ – essentially, learning how to group data together – he explains. But as the number of variables grows, that can lead models down blind alleys. 

“Each new feature grows the potential partitioning available exponentially, and thus may highlight spurious patterns that are not necessarily reflective of the true (albeit unknown) underlying data-generating process,” he says.

As the number of features that go into clinching a trade increases – such as the exchange, the size of the trade, volatility and others – the number of relevant observations also becomes smaller.

“We specifically look for patterns that are corroborated by numerous examples, trying to ensure that our assessment of the probability of executing a trade – and, more importantly, of what happens to the market after an execution or a miss – is based on a sufficient number of observations,” says Nuti.

“The fact that we send one order and miss it doesn’t immediately lead us to believe that we are never going to execute anything in that particular exchange,” he adds.

Predicting the market’s direction might be the “ultimate limitation” for intelligent machines working in execution, he says. Another conundrum is explainability.

Nuti says the lack of insight limits the use of more complex techniques such as neural networks, or at the very least requires the firm to alter its models so their results can be understood.

“We need to be able to explain to ourselves, to our clients and to our regulators why we have taken particular decisions,” he says. “This is a challenge of machine learning in general, where a lot of the very interesting techniques are fundamentally a black box.”