Prediction is very difficult, especially if it's about the future
In the context of achieving best execution, there is a growing focus on the pre-trade stage of the process. Accessing the most appropriate information at the point of trade can help execution desks make informed decisions around their various execution choices, including timing and style of execution.
When making an informed decision at the point of trade, one key input is an understanding of the prevailing liquidity conditions, and if not trading via risk transfer, an indication of how those conditions will develop during the period that the order is worked in the market. This is not as straightforward as it sounds, even for a listed market where liquidity data is readily available. For a market such as FX, which has a complex, hybrid market structure, with liquidity fragmented over many venues, it is extremely difficult.
Hence the need to look for help outside of traditional financial theories and practices. At BestX we are always creatively looking to solve problems by thinking laterally and learning from other disciplines and industries.
We are aware that the statistical properties of trading activity in FX exhibit clustering behaviour, similar to volatility, in that periods of high trading activity are more likely to be followed by another period of high trading activity. This ‘memory’ effect is evident, for example, around key points in a trading day such as the release of key economic data, times of key benchmarks such as WMR etc. This behaviour led us to exploring a statistical framework that was first used to model the impact of earthquakes in terms of predicting the frequency and magnitude of subsequent tremors. This is analogous to the observed characteristics of trading activity and volumes, i.e. there is a ‘shock’, for example an unexpected interest rate rise from a central bank, which results in a significant spike in trading volume, which then proceeds to have ripple effects throughout the market over an ensuring period of time. Clearly, there would be significant value if it would be possible to have some form of estimate of how the ensuing liquidity conditions would evolve, both in terms of magnitude and frequency.
Before we explore such a framework further, lets just return to why this should be of any use whatsoever? Well, for example, if you are looking to execute a large block of risk, say 1bn USDJPY, and have discretion on the execution approach and timing, then having some idea of how volumes in USDJPY may evolve over the next hour could be a key input in your decision making. If you’re objective is to minimise spread cost, and you are therefore leaning towards selecting an algo designed to be passive in the market and ‘earn’ spread, key variable for the performance of such an algo will be the depth of liquidity available to the algo whilst it is working. If, from previous experience, you know that the algo you are planning to use performs particularly well in market states characterised by high volumes, then a prediction of volumes will help decide whether to press ahead with this strategy or perhaps select a different product. Or, you may be seeking to minimise market impact and wish to time the trade to coincide with a high volume period to help ‘hide’ the trade. Insights into when such periods may occur over the next few hours is clearly helpful in this decision making.
Back to earthquakes. A statistical framework called the Hawkes Process has been used for some time for modelling earthquakes. This framework relies on historical information but is ‘self exciting’, which is the technical term for describing the phenomena where when an event occurs, e.g. an earthquake, then the probability of more events occurring, i.e. tremors, increases. From a volume prediction perspective, this can be thought of as a spike in trading activity due to a data release, will generally increase in increased subsequent trading activity. Over time, assuming there aren’t any further events, then the trading activity will revert back to its normal state. The Hawkes Process attempts to capture this behaviour.
As a brief aside, just to further clarify a key point – we are not attempting to predict an actual earthquake, or market shock. We are trying to predict what the likely impact of such an event will be. We were once asked if BestX had a model for predicting ‘flash crashes’, the query was met with an incredulous pause. If we had a model for predicting flash crashes then it is highly likely that the BestX business model would be very different, and probably involve relocating to the Cayman Islands and making lots of money.
Going back to Hawkes. Importantly, the hierarchical nature of the framework allows for the capture of the effect of different volume distributions typically observed on different days of the week. Also, the amount of trading also varies, with the quietest day of the week usually seen on Mondays, and Wednesdays showing, on average, the largest volumes. The charts below display the average intraday volume profiles for USDCAD over the course of 2016, also highlighting how different the pattern is on days such Non-Farm Payroll days, where trading activity is concentrated around the time of the data release.
So, does it work? Well, still early days but the initial results look very encouraging. The charts below show the predicted volume profile compared to the actual profile (black line), and the shaded area represents the 95 percentile confidence interval for this prediction.
We are now going to put the research into a development environment which will allow a more systematic testing of the framework, across many more currency pairs and using a much deeper data set to ‘train’ the framework. In addition, we wish to model other key event days such as ECB, FOMC meetings days, and also month, quarter and year-ends. Assuming the framework continues to perform to our satisfaction, we will then look to incorporate this within the BestX Pre-Trade module at a later date.