Author: Fabian Belsö – Director FinSide Consulting
European financial regulations have raised the bar for investment firms (both buy-side and sell-side firms) to comply with their Best Execution duties under MiFID II. Some firms already started applying advanced technologies like Machine Learning (ML) to improve execution quality. Others just comply with the bare minimum prescribed by the regulator and are at risk to lose competitiveness meeting the clients’ needs. This article outlines the problem of compliance with the regulatory requirement and discusses the key success factors that firms should look at when applying ML to Best Execution as part of their RegTech endeavours.
Best Execution requires firms to achieve the best result for clients when executing orders to buy or sell securities on their behalf. For less sophisticated retail clients, the primary objective is to achieve the best price at the lowest cost. For more sophisticated investors, further criteria such as the speed of execution or the likelihood of settlement may play an important role, too.
This is a difficult task for humans, given the myriad possibilities of executing a client’s order: Due to the fragmentation of markets that occurred in the last decade, there is a is a large number of execution venues with different characteristics and liquidity. A firm might route an order to lit markets (such as continuous limit orderbooks), their affiliated or a third-party dark pool. In the case of large orders, it might be beneficial to slice them into small child orders and execute them over time to reduce the market impact.
These are just a few considerations which might influence the quality of the execution. Clearly, there is no deterministic outcome and traders have to make a prediction at trade when deciding on the best way to execute orders. Could RegTech solutions that leverage Machine Learning help to solve this problem?
In the past, traders exercised large amounts of discretion based on their expert knowledge. This was possible in a low frequency environment with limited number of potential execution venues. Post MiFID and mainly in the most liquid equity markets, traders were supported by order management and execution systems. In these systems, fixed decision rules were programmed to determine the trajectory of the order execution. These rules would be static IF-THEN statements hard-coded into the system. E.g. a rule could say that an order in a small-cap domestic stock for a retail client up to a certain threshold would be executed at a given trading venue with a certain limit. These systems are inflexible, difficult and expensive to maintain over time.
Today, Machine Learning is everywhere, including in financial services. According to Andrew Ng (one of the most influential researchers of ML and Artificial Intelligence), Machine Learning is “the science of getting computers to act without being explicitly programmed”. In the case of Best Execution, maximising the execution quality is the statistical problem to be solved. Increasingly, there is “big data” from market data providers and proprietary transaction data from within the firm. This data needs to be applied to the appropriate mathematical or algorithmic models to extract knowledge and make predictions towards the best result for the client.
With (1) the availability of relevant data, (2) the implementations of ML algorithms, and the (3) computational power, it should become the industry standard “taking all sufficient steps” to leverage Machine Learning for Best Execution and endorse RegTech solutions.
What are the key success factors that firms need to focus on when embracing these new solutions?
Data, data, data
Capital markets and trading are driven by data which need to be processed and analysed for Best Execution purposes. Execution Venues produce tick-data up to nanoseconds which they are obliged to publish regularly (RTS 27 reports). Investment firms have to demonstrate how they use this data in the ongoing review and improvement of the execution quality. Clearly, getting the data right and making it available in a large organization in a non-redundant and normalized fashion is a challenge. Redundant data are often spread across in various data warehouses with ambiguous ownerships within the firm. Hence, the firm’s Chief Data Officer should be involved and incentivised to allow for sufficient data quality and further processing.
The art of machine learning
Machine learning algorithms range from simple statistical methods such as linear regression over supervised and unsupervised learning, to more advanced deep and reinforcement learning techniques. Leading firms in the field of algorithmic execution are experimenting with reinforcement learning.
RegTech talent
Trading floors and their associated IT don’t necessarily employ data scientists. However, the right talent might be found in-house working on other prediction problems, e.g. the compliance departments are using ML to detect rouge traders or fraudulent credit card transactions already for a longer time. Although these are different problems, there might be transferable skills and available talent. Pure data scientists lack the specialized market structure know-how that traders have acquired over years. Hence, teams with the right mixture of skills must be composed. It might also be worth exploring external RegTech providers who already offer solutions which can be integrated in the firm’s IT architecture.
Humans vs. machines
The market structure across different asset classes varies dramatically. E.g. the level of automation, liquidity, trade sizes, order types etc. can be very specific to markets. While e.g. equity and FX markets have high Straight-Through-Processing rates, other less liquid instruments like exotic derivatives are high-touch and less automated. Hence, there is no one-size-fits all for Best Execution, and equally Machine Learning cannot be used the same way across all asset classes. Likely, a right mix of ML algorithms and human judgment chosen for every asset class will lead to the best result.
Transparency
Investment firms’ execution policies need consent from a range of internal and external stakeholders including senior management, legal, compliance, regulators, and clients among others. Adding another level of complexity by applying Machine Learning to Best Execution might impede the creation of transparency which is a primary goal of financial regulations. This requires additional effort and considerations around documentation and communication.
Ready for the next step?
Firms have made large investments to comply with the MiFID II regulations and are now gaining experience. Now might be the right time to explore emerging technologies and understand how RegTech can help taking Best Execution to the next level.
