In a paper publisher earlier this year, Saeed Amen, founder of macro research firm Cuemacro, outlined how machine-readable news from Bloomberg could be used to create systematic FX trading strategies.
Unstructured text data was converted into structured data, which was then aggregated into sentiment indicators for currencies. According to Amen, the news-based FX trading strategy considerably outperformed a generic FX trend-following strategy over a similar period.
Bank of America Merrill Lynch recently made its first foray into FX research based on machine learning, using a combination of supervised and unsupervised learning (the latter providing no guidance to the algorithms on how to process the information) to analyse fundamental and survey data around EUR/USD.
Celent senior analyst, Josephine de Chazournes, refers to increased use of machine learning for tasks such as routing logic across disparate venues, analysis of technology and connectivity infrastructure and various analytics around risk and P&L as well as client and counterparty analysis.
According to Yaron Golgher, co-founder and CEO of algorithmic forecasting solutions provider I Know First, demand for his firm’s artificial intelligence-based FX predictive models has increased threefold over the past 12 months.
He says machine learning is capable of supporting the speeds demanded by high frequency FX markets. “After an initial training period, the algorithm can work online, receiving up-to-date feeds and returning the answer within milliseconds,” suggests Golgher. “This process does increase the extremely small trading window that is characteristic of high frequency trading, but the predictive power of the algorithm more than makes up for that time increase.”
However, Johannes Tynes, head of R&D at empirical analysis firm Inpirical, observes that some algorithms are inherently slow and cannot be usefully deployed for live predictions.
“In some scenarios, sets of linear approximations (factors by which to multiply and aggregate time series variables) and decision trees (maps of the possible outcomes of a series of related choices) can capture the main results of a machine learning model and be used on a live basis,” he says.
But even then it is often necessary to have in-memory implementations of both the data stream and the mode, holding them in random access memory, or RAM, so that they don’t need to access much slower disk storage, adds Tynes.
Amen notes that the training phase can be quite computation-intensive, so it really depends how often the trader needs to retrain their model. “Can you retrain your model overnight to find all the parameters and then use it with live market data the next day, for example? When using a trained model, you would have to ensure that the computation would be sufficiently fast if you wanted to trade at very high frequencies,” he says.
Unless you are a big fund willing to invest large amounts of money in advanced hardware and the best machine learning scientists to develop algorithms that work in the range of milliseconds — without any guarantee of profitability — applying machine learning to high frequency FX trading is not possible, according to David Lopez Onate, founder of artificial intelligence trading algorithm Forex Artilect.
Speed is a major limitation, especially if the system is supposed to learn during market hours, adds de Chazournes. “Current systems learn overnight on the day’s trading day, out of hours, and sometimes don’t even adjust the system until a few days later,” she says.
The cost and computational power required to run successful machine learning initiatives in real time (as well as the sheer number of factors that can produce unexpected and short-term gyrations across currency pairs) continues to deter many FX traders, agrees IC Markets managing director, Angus Walker.
He explains that the special challenges for machine learning presented by high frequency trading generally arise from the granularity of the data. Development starts by connecting the proposed model to the data that is to be traded, followed by validating if the system is viable to trade.
Lucena Research CEO, Erez Katz, reckons there is widespread misinformation and lack of understanding in the market as to how to evaluate and use alternative data effectively. “We hear time and time again from our hedge fund clients that they prefer to fail fast when dealing with alternative data so as to not waste precious quant research time,” he says.
In July, network data analytics company Corvil announced it was rolling out software that uses machine learning and big data analytics to help traders identify anomalies, triage areas of greatest concern and predict conditions for improved planning.
The firm’s chief marketing and business development officer, David Murray, says traders can ensure that the external data they use is reliable and timely by testing their data sources, examining the resulting noise factor and isolating and either adding or subtracting data sources to get the best fit.
Tynes observes that for conventional (structured and quantitative) time series data, timeliness and consistency checks are usually manageable because traders can analyse the quality of data by backtesting and investigating on a long time series from the source.
For new data types and unstructured data the due diligence process is much more hands-on. “Sample data with a decent length is still important, but you really have to look into methodology and definitions carefully,” he concludes.