Machine learning can help to automate decision-making for efficiency, but also to optimise human decision-making. Dan Barnes reports.
The application of machine learning algorithms in the trading landscape extends from high-touch to low-touch trading. New use cases are coming to light, across asset classes, which are enabling trading desks to attain better execution with more measurable, quantifiable processes to deliver improvement.
Algorithms that use machine learning are typically split between supervised learning models, in which an algorithm is trained on a labelled dataset and conclusions are adjusted for accuracy, and unsupervised learning models which explore unlabelled datasets to find patterns.
At TradeTech 2022 in Paris, buy- and sellside firms explored many use cases in the front office. For any ML model, the quality of data is enormously important. In trading, this typically limits the use of ML to asset classes in which data is plentiful.
“Many challenges in quant finance can be formulated as a supervised learning problem, which you can back test and then continuously improve,” says Dr Peter Ho-Spoida, VP data strategy in Market Data + Services at Deutsche Börse. “But I would say that the level of noise, compared to other areas where AI is being applied is very high. So you need a very disciplined, robust approach, and a lot of data.”
Even for the most successful quantitative trading firms this creates a situation where “AI/ML are not applicable to all strategies,” says Antish Manna head of trading analytics, Man Group. “They are most useful if a firm has a huge sample – millions of orders, worth billions of dollars – because these technologies will be able to provide the rich data set insights that traders will need.”
By taking a rigorous scientific approach to the use of data and analytics, capital markets firms can begin to filter high volumes of information for viable signals, which can then be used to spot opportunities, or risks, at far higher speeds than human observers, allowing the front office to be more responsive.
Hasan Amjad, head of execution risk at Schonfeld, added that the evolution to using machine learning is often a result of trying other quantitative models and finding they do not yield results that are effective enough.
“You start by asking a trading desk for some hybrid parameters; and they just give you a number,” he says. “From there, you set up a linear regression and then you realise that this is too ambiguous, it’s not good enough. So you start using Monte Carlo methods* to build up the distributions and start looking at confidence intervals. Then you might realise that some of the distributions you’re working with are not even stationary.”
After trying other methods such as ‘random forests’, or gradient-boosting techniques it will often become clear that manual feature engineering being set up in machine learning is not fitting, leading to more advanced methods.
“You move on to using neural-network based methods like deep-learning, where finally you realise that you can get the network to do more of the work for you,” says Amjad. “Going into deep-learning-based reinforcement learning, using possibly things like restricted Boltzmann machines, which I know some shops are doing already, some techniques coming out of research teams are new symbolic methods, which combine machine learning with old school rules-based models to try and get the best of both worlds.”
The right starting point
It is important to understand where a firm is comfortable in order to begin on this journey. “It’s not necessarily about AI being very complicated to start with,” explained Ho-Spoida. “Basic stuff might typically be having a signal framework where you monitor volume curves which is quite dynamic and changes every day, particularly event-driven days. Liquidity can disappear from one minute to another, spreads can spike, volatility can spike. So, you need to have a dynamic system in place where you monitor this. Supervised learning is a natural fit in this space.”
Complexity is not the only aspect to be considered. For some investment strategies, automation of trading itself is not the goal, but the efficiencies that are to be found through optimising the human decision-making process.
“We’re not going to have a machine learning model sitting on top of our execution; that’s a dream scenario, of course, but it’s not realistic,” says Sindre Falkeid Kommedal, equity trader for Nordea Asset Management, at TradeTech 2022. “[We are looking at] how we can explain to someone that they should do one thing instead of another, even when that potentially contradicts their beliefs.” This model takes a quantitative approach to identifying human bias in trading behaviour, and seeks to eliminate that bias through guidance.
“From a trader’s perspective, looking at the execution of investments, a key thing is human behaviour. We tend to be creatures of habit, and do the same thing, especially when we’re stressed out. If you do the same thing over and over again, it might not be the best thing to do,” says Kommedal. “You can mitigate it by using what we’ve seen in the low touch equity world with algo wheels, but applied to high touch trading. [In low touch trading] instead of picking the exact algorithm, you pick the style you want to execute with. That also feeds into a more measurable process for human bias.”
The right closing point
Beyond the application of data science and ML technology, it is also crucial that a business has a viable use for the system it is developing.
Daniel Carpenter, CEO of analytics provider Meritsoft, is now seeing trade execution analytics being developed using ML, in order to improve effectiveness. This is beginning with cost analysis. “It’s the trade expenses, the categorisation that one needs to look at in order to understand what the total trading expense is including agent bank fees,” he says. “Once you have that you can start applying logic to assess if one deal is better than another deal, or one counterparty is better than another counterparty. At the moment, so much of that is siloed internally, either regionally, or by asset class.”
Joe Wald, managing director, and co-head of electronic trading at BMO Capital Markets, notes that another real-world sellside application is the capability to predict likely market activity, in order to pre-empt resource allocation and potentially necessary responses.
“There are different systematic internalisers (SIs) that have come to the marketplace now that are using AI in terms of the way they allow the [trade or order] match to happen,” he says. “There are different order types, even on some of the bigger exchanges, that are looking at ways of allowing you to interface with that to provide, for example, what the volume on close is going to look like, before it’s there. A lot of people aren’t using that simple market data to understand where that fits into their algorithmic puzzle.”
Starting simple allows the value to be tested, and could potentially prevent hitting dead ends, which Amjad notes are a real risk.
“Even when you have a machine learning model, there’s no guarantee that it’s going to be actionable,” he says. “In fact, I would hasten to say that, regardless of where you are on that frontier, new does not necessarily mean better, or profitable.”
*Monte Carlo methods are used to value and analyse complex instruments, portfolios and investments by simulating the various sources of uncertainty affecting their value, and then determining the distribution of their value over the range of resultant outcomes.
©Markets Media Europe 2022
TOP OF PAGE
