UBS leverages machine learning to optimize site selection

UBS has developed a state-of-the-art machine learning (ML) technique to look for patterns in the data that skew the outcome of passive order placement on equity platforms towards higher fill rates, an information leak reduced, lower market impact and lower execution costs. .

When considering which venue to select to post passive liquidity, it is important to have a granular understanding of how your company’s trading profile interacts with each venue across a range of metrics. Marketing data received directly from sites regarding fulfillment performance generally reflects the average site-wide experience. This will inevitably be weighted by the performance of certain participants, depending on the context in which those participants are trading. This performance will not necessarily hold true for each company’s unique experience when trading on this venue.

“Properties in any given location can affect market participants differently depending on when and in what context they are traded,” says John Fruen, head of market structure and liquidity strategy EMEA, UBS. “Therefore, it is more important to rely on the experience of an individual company than to rely on the idea that a place has inherent properties that can be exploited, without knowing if they correspond fully to the In addition, understanding the context in which an order is placed creates the opportunity to use differences in site experience at the most opportune times.

Estimation of execution quality in passive trading has historically focused on metrics such as how quickly or how likely an order will be executed. While fill rates and margins are important factors, they ignore any potential information leaks and the opportunity cost of unfilled orders.

“Average daily volume [ADV] was often used in older models,” says Fruen. “The problem with this is that it creates a ‘chicken and egg’ problem – if you assume a venue is performing poorly in business because it has no volume, you’ll never send streams there. orders and you’ll never find out what this place is like.”

Using machine learning to improve trading

Not only is this use of ADV restrictive, but it also assumes that sites are static, which is far from true. To reflect the need for a more dynamic approach to intelligent order routing (SOR), UBS has developed a machine learning model to better support routing decisions – initially for passive order flow. He looks for patterns that correlate with better execution results – higher fill rates, reduced market impact and information leaks, and lower fees – on enlightened stock platforms. .

With UBS’s approach, the innovation lies in the use of the Bayesian decision tree model, which provides transparency on how exits are selected and matches the characteristics needed for stock trading. It has low latency and can consider both parent order and order book information when deciding on order placement, depending on the differences between the characteristics of each relevant place in different contexts.

A structural advantage of using machine learning tools is the adaptability they demonstrate to changing market circumstances, adapting in real time to reflect changing properties of a platform. form of negotiation.

“Because properties can change relatively quickly, depending on the arrival of new members or order flow decisions, it pays to continuously explore and adapt to these changes,” notes Fruen.

Traditional approaches can be limited by their reliance on a daily measure such as ADV, where the UBS model uses Thompson sampling to balance the need to explore sites to find the best execution experience versus the using historical data to determine which site is likely to perform best under current market conditions.

“The Thompson Sampling methodology offers us a way to optimally manage the trade-off between capitalizing our knowledge [drawn from historical experience] and exploring potentially exciting new investment strategies,” says Giuseppe Nuti, head of agency quantitative research at UBS.

Take the chance

In the United States, the A/B trading scenario on the use of UBS’s Machine Learning Intelligent Order Router produced some interesting results, confirming that the usual patterns of order routing can be improved.

One of the main advantages of this approach is its ability to explore new order types and quickly determine the most appropriate market and order context to take advantage of the new offer. For example, it interacted very positively with the type of D-Limit order offered by the Investor Exchange (IEX). The D-Limit will make an order one tick more passive if IEX thinks the market is likely to move towards it based on an order book signal.

“IEX tries to limit adverse selection, defined as trading only when the likelihood of the market trading through our limit price is high,” Nuti says. “Instinctively, the benefit of this type of order should be more pronounced when queues are high, as the signal of a collapsing quote, central to IEX’s offering, will be clearer to investors. larger queues. Yet, we found that the model learned to take greater advantage of the D IEX limit for small queues based on two interesting, and not necessarily obvious, observations. First, small queues will be subject to more frequent tick changes, i.e. take advantage of the insurance inherent in the D IEX limit when you use it more frequently.

Second, choosing IEX for a limit order joining a large queue by definition forgoes being on other sites, which may offer a higher rate of queue natural depletion ( and therefore a higher probability that a filling does not lead to adverse selection).

Another example of some of the interesting behaviors observed by this approach involves the use of reverse locations – those where the abuser is paid to take cash. Various studies have shown that information leaks for the release tend to be quite high, as the market interprets the order as an effective gap reduction.

Essentially, reverse sites offer enhanced fill probability, paid for in fees and information leakage. Overall, UBS believes the market has adapted to different pricing structures, offering little discernible benefit for posting to reverse sites under normal circumstances. Since the value of routing to a reverse exchange increases with the width of the spread, as market conditions present wider prices, it becomes increasingly valuable to leverage reverse sites. This is illustrated above (where the spread ratio represents the ratio of the spread at the insertion of the limit order to the average spread for this security for the day): the wider the security is compared to its average spread, the more we can take advantage of reverse sites.

The context-aware ML technique is embedded in UBS’s US and EMEA cash equities SOR, and the firm further seeks to use it in routing dark orders to find median liquidity and when crossing the gap.

To learn more about our in-house developed machine learning techniques, please contact the UBS trading desk.

Martin E. Berry