MACHINE LEARNING IN PRACTICE.
James Hilton, Head, AES Sales, EMEA at Credit Suisse
“No man is better than a machine, and no machine is better than a man with a machine” (Paul Tudor Jones, Founder of Tudor Investment Corporation)
Competition between brokers in the algorithmic trading space has never been greater. When MiFID II was implemented in January 2018, it reinforced two very key points: that the consumption of execution and research should be independent and that there should be even more focus on best execution. Brokers should not expect to be compensated for investment research via execution, but to be compensated if it can be proved they have better trading performance. In that sense, MiFID II provides a platform for success.
Since the start of last year, we have observed a number of trends from our buyside clients. First, there is much more structure around the selection of algorithmic trading providers. Our clients want to get a deeper understanding of our algorithms, what liquidity we access and how we do it. Comparing brokers side by side in this way provides more of a platform to see who doesn’t stack up. It’s subjective, but it’s also obvious if a broker hasn’t invested in connectivity to new liquidity sources or low latency routing technology. We have seen a clear concentration of buyside wallet amongst the top brokers, as sub-standard or white-label offerings don’t pass the relevant tests.
The second trend we have observed is a proliferation of both algo wheels and a more dynamic allocation of order flow based purely on performance. This is where there is a real opportunity for brokers. Quantitatively measuring performance in an unbiased manner is a great objective, but there are many pitfalls. The largest by far is measuring on insignificant data sets, which generally occurs because flow has been segmented across too many categories or too many brokers. From a broker’s perspective, properly understanding how clients measure us is critical to delivering enhanced performance.
Where clients have significant data sets and their metrics are well defined, brokers can make incremental improvements. Many clients have been able to demonstrate that their total cost of trading has declined by implementing these types of systematic frameworks. Since MiFID II, we don’t often hear the adage of algorithms being commoditised. The infrastructure required to compete has become more expensive, which raises the barriers to entry, and polarises capabilities. But for those brokers who have properly invested, where does the edge come from now? We certainly see a large benefit by finding innovative ways of limiting trading on lit markets, but the most recent and impactful advance for Credit Suisse has been machine learning.
Over the years we have worked with clients to develop customisations to our algorithms to better suit their requirements. Some key factors have included stock volatility, momentum, and size. Measuring the performance before and after a customisation can be difficult due to different market environments – an algorithm might work well in a mean-reverting market but not in one which is strongly trending. We’ve put frameworks in place for some clients to route to two or more different variants of a strategy and randomise on an order by order basis, thereby reducing bias. The next step beyond this is to use a machine learning assisted approach to determine the best variant of the strategy for each situation.
Machine learning brings various benefits. Whilst operating in a very controlled environment, the system identifies which variant of a strategy is likely to work best for a particular order based on its characteristics, including the stock, size, market conditions and the client. Rather than pre-defining that variant A works better than variant B in liquid stocks with a low projected participation rate, the machine learning framework can dynamically adjust what counts as ‘liquid’, and finesse at what participation rate variant B starts working better.
The ability to continuously learn and adapt as the market environment changes is extremely powerful. Rather than waiting for a review at some random time period, the framework allows for adaptation as soon as significant outcomes are observed. Rather than only trying to find significance in the outcomes of a particular client’s order-flow, the framework also learns across our whole platform. It is everything that we’ve previously been doing, but better and quicker.
Myth or reality
There is a lot of noise and frequent headlines about machine learning, and you’ll be hard pushed to find a conference that doesn’t dedicate some time to AI, machine learning or big data, and the potential it might have on your business. The challenge is to make it a reality. At Credit Suisse, the machine learning framework is live and ready to be used by our clients. Over the last six months, on a randomised basis for internal VWAP orders, we have been routing to both the adaptive strategy and our standard strategy. The results have been extremely encouraging: the adaptive strategy powered by machine learning has outperformed by more than 30% (vs VWAP in spread terms).
Best execution will continue to be a huge focus for buyside clients, partly because it’s mandated by our regulators, but mainly because it’s a competitive advantage. The use of machine learning technology will continue to evolve at Credit Suisse, becoming more entwined in every aspect of our business. It will enable us to continue to deliver on our clients’ demands for ever improved performance, adapt quickly to changing market environments and stay ahead of our competition.
©Best Execution 2019