There has actually never been a true ‘one-size-fits-all’ solution in FX execution, claims Dr John Quayle, Head of Client Algo Execution at NatWest Markets. This is because, at its core, every execution decision involves a trade-off between minimising transaction costs on one hand and minimising execution time and the associated market risk on the other, he explains. “While 100 per cent certainty can be achieved through an immediate risk-transfer trade, that certainty comes at quite a high cost,” Quayle adds. “Conversely, a well-designed passive execution algo can typically achieve better average pricing but may also introduce exposure to market movements while the order is being worked. The most common customisation discussion we have with clients at NatWest therefore revolves around the speed of the algo.”
“The most common customisation discussion we have with clients at NatWest revolves around the speed of the algo.”
Dr John Quayle
In practice, he argues that often this will translate as a discussion about where a client wishes to position themselves on the cost-versus-risk spectrum. So although some clients may be primarily focused on minimising market impact and are comfortable taking additional execution risk to achieve lower overall costs, other clients may tend to be more focused on speed, according to Quayle. He adds: “What has changed over recent years is that clients are increasingly seeking execution strategies that can be calibrated to their specific objectives and risk appetite, which of course varies depending on the currency pair and time-zone. We do not see this demand for customisation being limited to a particular segment of the buyside; whether the client is an asset manager, hedge fund or a corporate, the ability to tailor the balance between execution cost and market risk remains one of the most important aspects of FX algo design. Therefore, as liquidity providers continue to invest in analytics and gain a deeper understanding of their execution outcomes, we expect this trend towards greater flexibility and customisation to continue.”
Attitudes and expectations
In addition, client attitudes and expectations around FX algo customisation have evolved significantly in recent years. Buyside firms may well have become significantly more sophisticated, but the direction of travel is not simply towards more buttons and parameters, says Vittorio Nuti, Head of Algorithmic Trading, Deutsche Bank. He adds that clients are now increasingly asking for customisations based around improving execution outcomes, workflows and evidence, rather than seeking bespoke code for every trading style. “Historically, the discussion was often about basic controls such as speed, urgency, participation and benchmark tracking. The more recent client dialogue is about pre-trade analytics, real-time monitoring, post-trade TCA, API integration, governance and the ability to adjust execution dynamically as market conditions evolve,” Nuti says. He believes that the main drivers behind this shift are factors such as best-execution scrutiny, more sophisticated execution desks, independent TCA from providers such as BestX, Curex and Tradefeedr, broader adoption of electronic workflows and an overall increase in client desire to explain outcomes internally.
In addition, Nuti shares that the most common requests from clients tend to be focus on key aspects of FX algo execution, such as aggressiveness, urgency, benchmark objective, timing, limit constraints, passive versus liquidity-seeking behaviour, use of internal liquidity, and the ability to intervene during the order. He adds that clients also increasingly request customisations relating to their workflows and analytics: how orders are submitted, how live execution is monitored, how reports are delivered, and how performance can be compared against risk-transfer or independent TCA benchmarks. “In practice, many clients want to tailor how they interact with the algo more than they want to redesign the underlying execution engine,” Nuti says.
“In practice, many clients want to tailor how they interact with the algo more than they want to redesign the underlying execution engine,”
Vittorio Nuti
Simplifying the process
Meanwhile Pieter Oudshoorn, FX Sales for European Financial Institutions at NatWest Markets, adds that setting the most appropriate algo speed is often the most common focus for NatWest’s algo clients. The parameters which will primarily affect this are firstly, the placement of child orders in the order book (joining the passive side of the market or going in at mid, for example) and secondly, the choice of liquidity pools, he explains. “In practice, however, most users specifically want to be ‘passive’ ie try to capture spread and join the passive side of the market,” Oudshoorn says. “This in turn limits the flexibility from the order placement level, which leaves the choice of liquidity pools as the most impactful avenue for customisation.” He explains that this is overall an especially rich topic and one which is evolving continually as new venues such as OneChronos enter the fray, and existing venues add new features and change their approach to curation. Oudshoorn adds: “At NatWest we are able to support the real-time management of liquidity pools, therefore this is a particular focus for client discussions and where we see clients regularly testing new combinations. Knowledge of which liquidity source performs well in which pair and which time-zone is crucial and is a key area where we can add value to the relationship.”
Flexibility is also ultimately vital to the execution process – market conditions change, risk appetite evolves, and so the tools available do need to be able to easily accommodate this need, Oudshoorn explains. He argues that the pitfalls ultimately arise not from the scale or scope of flexibility offered, but rather from the challenge of gathering enough data for a given set of parameters to determine if they actually benefit the user or not. “The fact of taking market risk, that each order has a variable outcome as a result, means that getting a statistically significant set of data for a given set of parameters takes time,” says Oudshoorn. “If you always use the same settings, you will get there quickly, if you use five different settings on a regular basis, you need to be able to log the performance for each run (in itself not a given, depending on the analytics being used) and it will take you five times as long to get there.” He adds that this is where third party services such as Tradefeedr can add value. While peer universe-style aggregations can help, Oudshoorn believes that these again tend to cover only a relatively small number of the customisation options available, which puts the onus back on the user to limit the variations used to improve the data. He adds: “The ability to quickly identify a small number of key default settings is key to successfully navigating this complexity, which is why we work very closely with our clients to thoroughly understand their execution requirements, leveraging our data and TCA services to map the most suitable parameter configurations.”
In addition, customisation requirements can differ between different client segments, with asset managers tending to care most about arrival-price performance, as well as other factors such as their portfolio workflows, benchmark tracking, governance and TCA evidence, according to Nuti. “Official-sector clients, on the other hand, typically place more emphasis on auditability, information leakage, transparency and clear suitability guidance, while quantitative and highly automated users are more likely to focus on factors such as API integration, parameter control and systematic workflow,” he adds. Furthermore, Nuti adds that regional banks tend to value workflow simplification, hedging automation and spread capture, as well as requiring access to institutional-grade execution technology but without the need to build everything in-house. He explains that corporate and less frequent users will generally prefer to opt for simplicity, confidence and clear defaults rather than requiring a wide parameter set.
However, Nuti warns that such flexibility can become counterproductive if clients are asked to choose parameters they cannot reliably optimise, or when small parameter changes create complexity that is not justified by measurable improvement in outcomes. “For example, urgency, limits and timing choices can materially affect execution quality, but clients may not always have the market context to choose them optimally in real time,” he explains. “Too many choices can create decision fatigue, inconsistent usage and lower-quality execution decisions. A better model would be guided flexibility: keeping meaningful client controls but framing them around execution objectives, while also protecting the user with analytics, defaults and safeguards instead.”
Customisation requirements can differ between different client segments
Finding the right balance
The most useful division of labour is instead when clients define the objective, and the provider helps determine the route, claims Nuti. He believes that clients should specify whether they care most about speed, market impact, benchmark tracking, discretion, completion certainty or leakage reduction, while providers should increasingly translate those objectives into settings using market intelligence, liquidity data, execution analytics and pre-trade modelling. “This reduces the burden on the client and should make execution behaviour more consistent,” Nuti adds. “That does not mean removing client control, instead it means giving clients the ability to override or guide execution, while using provider expertise to recommend sensible defaults and dynamic adjustments.”
Quayle notes that clients now also need to choose the level of market risk that they are willing to take, and that is an aspect of the algo execution set-up which “simply cannot be automated”, he adds. However, he explains providers can instead make it easier for the user to select the combination of parameters that best suits their needs, which means that their expertise is vital. “The client determines the desired balance between execution cost and market risk, while the provider uses its expertise, analytics and market intelligence to optimise the underlying execution logic within those boundaries,” Quayle adds. “This highlights one aspect of customisation which is often overlooked – the need to easily set up customisable defaults.”
He explains that this is due to modern algos being able to support a very large number of parameters and potential combinations, but in reality, most clients only require a relatively small subset of those configurations to meet their day-to-day execution needs. In turn, the ability to create and save these customised defaults can significantly improve usability, operational efficiency and consistency of execution outcomes, Quayle says. “The industry is unlikely to be moving towards a fully automated model where providers make all execution decisions,” he adds. “Instead, the trend is more towards smarter automation within client-defined parameters; clients will set the risk appetite and execution objectives, while providers make it easier to implement those preferences and use their expertise to deliver the best possible outcome for the client.”
Nuti observes that leading providers also now appear to be moving toward a controlled-choice model: intelligent defaults, pre-trade guidance, real-time feedback, ability to intervene, and post-trade analytics to assess the result. “Deutsche Bank’s pre-trade cone-and- outcome box tool is a good example of this direction,” he adds. “It helps clients understand not only expected execution duration, but also historical market-move ranges, forecast arrival-price performance, risk-transfer comparison, and the impact of switching between styles before or during execution. This allows the client to remain in control, but without the need to perform all the execution analytics manually.”
Use case for AI
Furthermore, AI and machine learning are also beginning to support the parts of algo execution where prediction and adaptation matter most – namely in the functions of liquidity forecasting, short-term alpha, market-move estimation, impact modelling, timing, and dynamic order placement, Nuti explains. He adds: “Investment in proprietary short-term forecasting is a large area for Deutsche Bank research. The objective is to use data and prediction to improve execution decisions, not simply to automate a static schedule.” A more advanced use case is adaptive execution that responds to changing liquidity and volatility, while still remaining inside client-defined objectives and safeguards, says Nuti.
He also argues that a natural next step is for execution systems to learn from a client’s historical trading behaviour. This would include areas such as preferred urgency, typical intervention patterns, tolerance for timing risk, benchmark choice, execution horizon and which outcomes the client actually rewards, says Nuti, while also allowing the system to recommend or apply client-specific execution profiles, instead of asking the user to choose between generic settings on every order. “The barrier is less about the technology and more about the explainability, governance and client consent elements. Any adaptive profile would need to be transparent, testable and easy for the client to challenge or override,” he adds.
This means that transparency also becomes more important as customisation and automation increase, according to Nuti. “If an algo is making more decisions on the client’s behalf, the client will increasingly expect to understand why those decisions were made and whether they were consistent with the stated objective.” He explains that algo clients increasingly want information on strategy logic, expected performance, limitations, liquidity sourcing, slicing, internalisation, execution path and to be able to compare this data with the pre-trade expectations. Nuti adds: “The FX Global Code framing is helpful here: algorithmic execution is a form of delegated execution, so providers need clear communication, documentation, governance and accountability.”
Oudshoorn agrees, adding that transparency is absolutely critical in building trust for the user when it comes to customisation. “The value of customisation is only fully realised if clients understand not just what controls are available, but also how those controls influence the algorithm’s behaviour in practice,” he adds. “Without that transparency, it becomes difficult for clients to predict outcomes, evaluate performance, or align an execution strategy with their objectives and risk appetite. At NatWest, we see strong demand for transparency around both the available customisation options and the practical implications of those settings.” For example, Oudshoorn notes that if a client were to choose a more passive execution profile, they would want to understand how aggressively the algo will seek liquidity, how completion risk may change, and how the strategy is likely to respond if market conditions deteriorate or liquidity becomes scarce. He adds: “Many clients will expect the algo provider to provide nothing less than 100 per cent of the information on execution decisions as and when required.”
“The value of customisation is only fully realised if clients understand not just what controls are available, but also how those controls influence the algorithm’s behaviour in practice.”
Pieter Oudshoorn
For most people the ‘better outcome’ is reduced market impact and this is the most meaningful execution metric for the algo user, according to Oudshoorn. This is measured by looking at the all-in fill vs the spot mid at inception (implementation shortfall) and represents the actual cost of the execution in simple monetary terms, he adds. “A customised strategy is only genuinely better if it improves the overall execution frontier: either it delivers lower cost for the same risk, or lower risk for the same cost – this is the ‘speed vs impact’ equation again,” Oudshoorn adds. “Whether you are using a customised setting or a standard one doesn’t change this, and so the objective then is simply to be able to gather enough data to be able to make an informed decision. This comes back to the challenge of restricting the range of different settings used in order not to dilute the quality of the data too much – a common pitfall for those new to algos.”
Direction of travel
The use of analytics and TCA is also definitely helping users improve outcomes, Oudshoorn argues, with the realisation that the benefits of algo execution can now be measured, and that the relative performance of different providers can be compared is now widely accepted. He adds: “One of the biggest changes is that for many of the more active users meaningful segmentation of performance analysis can be achieved i.e. determining that a given setting which works well in G5 pairs, is not necessarily the same as that which gives the best outcome for CE4.” For example, a given venue is likely to have different levels of information leakage depending on the pair, Oudshoorn claims. “Therefore identifying this and setting up the appropriate customised defaults for the different use cases is one of the goals of high-quality TCA analysis,” he adds.
“Customisation should also be judged against the objective it was designed to improve,” says Nuti. Relevant metrics include arrival-price performance, for example, in addition to other factors such as implementation shortfall, risk-transfer comparison, market impact, fill quality, completion rate, internalisation rate and execution consistency. “The key is statistical evidence over a meaningful sample, not a single-trade anecdote,” he adds. “We always stress that algos should be evaluated on average and in aggregation because individual orders can vary due to market risk.” As a result, Nuti claims that a customised strategy is genuinely valuable only if it improves the relevant metric without introducing offsetting costs such as slower completion, higher operational burden or poorer consistency.
Furthermore, TCA is making execution customisation more evidence-based, with clients increasingly using independent and internal analytics to compare algo results against risk-transfer benchmarks, peers, historical performance and alternative execution styles. “However, many risks and data gaps do exist which makes these types of evaluations complex,” Nuti warns. As buyside execution analytics mature, firms are also likely to develop increasingly distinctive house execution styles, just as they already have distinctive investment philosophies and governance frameworks, Nuti believes. He adds: “Some firms will favour market-impact minimisation, others benchmark tracking, others internalisation capture, speed, certainty or reduced information leakage.” These preferences can be encoded into preferred execution profiles, workflow rules and governance-approved defaults, Nuti argues, and adds: “The likely future is not a unique algo engine for every client, but a common core execution platform with client-specific overlays that reflect trading philosophy, data feedback, operational workflow and governance expectations.”
