Electronic trading has increased to around two thirds of global spot FX transaction volumes (BIS survey, 2016; spot is the most mature electronic FX market) and this proportion will be considerably higher when measured by the number of transactions. Most interaction with wholesale FX market venues will involve the use of algorithms of some sort.
These may be explicitly deployed by the user, via automated trading and execution systems, or be embedded in dealing interfaces used to access the venues (these interfaces will include order retry algorithms, as well as typically a range of order types). Thus all participants in wholesale FX markets need to consider the safe and compliant use of algorithms and their controls in their FX execution path.
Definition of Algorithmic Trading – Signal vs. Order Placement
A standard definition from MiFID II (adopted in other regulations) defines algorithmic trading as “trading in financial instruments where a computer algorithm automatically determines individual parameters of orders such as whether to initiate the order, the timing, price or quantity of the order or how to manage the order after its submission, with limited or no human intervention”.
When considering algorithms used in trading, it is helpful to make the distinction between a “decision-making” algorithm (or “signal”) and a market-facing execution (“order placement”) algorithm, which relays these (or human) decisions to the market venues. MiFID II uses the terminology: “investment decision algorithms” and “order execution algorithms”, specifically: “Investment decision algorithms make automated trading decisions by determining which assets to purchase or sell. Order execution algorithms optimise order-execution processes by automatic generation and submission of orders or quotes, to one or several trading venues once the investment decision has been taken.”
MiFID II’s definition of algorithmic trading only excludes “those pure investment decision algorithms which generate orders that are only to be executed by non-automated means and with human intervention”. However, most “manual” interfaces to market venues will have some embedded “order placement” functionality and controls, and these should be reviewed as algorithms.
Regulations: “Regulatory Technical Standards” and Principles-based
Any trading algorithms must comply with both the explicit algorithmic trading requirements (broadly covered by MiFID II, specifically the “Regulatory Technical Standards” (RTS) defined by ESMA), and the relevant wider market behaviour regulations (that apply to any type of trading activity), much of which is principles-based. For the FX market, this latter element is consolidated within the FX Global Code (FXGC), which has recently been recognised by the FCA in the UK.
ESMA RTS for algorithmic trading define, among other requirements, explicit pre-trade controls on price collars, order value, order volume and message limit, as well as continuous monitoring of position limits and outstanding exposure. These standards build on widely adopted guidelines, including the ESMA Automated Trading Guidelines from 2012 (which cover all financial trading instruments). The FXGC contains many aspects relevant to algorithmic trading – either explicitly, or by defining compliant modes of behaviour in FX markets, applicable to all types of trading activity.
Separating signal generation and order placement
Broadly speaking, “order placement” algorithms are where the explicit pre-trade controls from the RTS apply, while “signal generation” algorithms need to be compliant with principles-based market behaviour rules (assuming that the signal is sent to the market via a suitable order placement algorithm).
In terms of algorithmic systems design, it is sensible to separate the order placement function and controls from the signal generation element. This ensures that explicit pre-trade controls are always enforced, including when a range of different signal generation algorithms are being used, and these controls cannot be over-ridden without appropriate compliance oversight (an explicit requirement of the RTS). All market-facing algorithms at a single institution need to go through the same pre-trade controls (to accumulate the full message limit and to monitor outstanding exposure for the institution).
In terms of algorithmic systems design, it is sensible to separate the order placement function and controls from the signal generation element.
Some aspects of acceptable market behaviour can also be encoded in the order placement function, provided suitable information is available (for instance on the overall position). As an example of this, consider Principle 12 of FXGC, which addresses “spoofing” and other practices which intend to create a false impression of the market price, depth or liquidity. Each submitting “signal” algorithm is required by the RTS to have a unique identifier, so the objective (for example, risk reduction in market-making or risk increasing for an alpha-seeking strategy) is known (internally).
If an order request is submitted that would move the position in the wrong direction, then this may either have been generated on stale information or be an indication of a spoofing type of behaviour within the signal generation. In either case the request should be blocked and not submitted to the market.
Pre-trade controls – data integrity
The example above illustrates that in order to apply pre-trade controls effectively, the information used by these controls needs to be reliable and preferably independent of that used in signal generation. These pre-trade validation data should be subject to a high degree of monitoring, including checks on latency from source, validity and consistency (for instance checking for excessive short-term price moves). It is essential that final validity checks on outgoing orders can reference independent incoming data for the specific trading venues to which each order is being submitted.
Principles-based regulation – constraints
Signal generation algorithms must comply with principles-based regulations. This has consequences for the range of algorithms that may be appropriate to consider – for instance, highly flexible machine-learning algorithms, with no constraints, might find an “optimal” behaviour (if optimising only for execution cost for example) that is inconsistent with principles-based regulation. This can be a particular problem where the model/algorithm is “adaptive”, since unforeseen market states may lead to unexpected behaviour, without a degree of determinism in the model specification. Algorithmic methods originating from less constrained problem areas should be applied to the trading context with caution.
Constraints on algorithmic behaviour ensure that compliance with the principles-based regulations can be demonstrated before execution, rather than relying on post-trade monitoring. Review of algorithmic systems is made systematic if the separate functions of signal generation and order placement are considered and the RTS and principles-based regulations form part of both the systems design process and the choice of algorithms.
