There is a new breed of hedge funds that are using artificial intelligence (AI) tools to trade the currency markets. Galen Stops takes a look at a few of these emerging funds.
“AI has become a catch-all phrase, everybody and their grandma wants to use it now because it’s a buzzword,” says Damien Loh, the CIO at Ensemble Capital, a Singapore-based hedge fund.
With an academic background in computer science, Loh spent 15 years at JP Morgan before launching Ensemble Capital in 2017 alongside Atsuo Ogaki, the former head of FX at Nomura in Tokyo and 22-year veteran of JP Morgan. Given that many funds utilising a wide range of technologies are indeed claiming to be AIdriven these days, one way that Loh’s firm has tried to distinguish itself is by taking part in an incubator programme run by the American technology company, Nvidia, that works with AI and data science startups.
“This shows that Nvidia has looked at the people we employ and the methodologies that we use and recognised us as a bona fide AI company – it shows that we’re not just another firm trying to use the term in our marketing,” he says.
Specifically, Ensemble Capital uses deep learning tools to run four different trading models – seasonality, economic, momentum and value (asset relationships) – across 20 currency pairs, with the weighting towards each currency and model automatically readjusting based on performance in recent market conditions.
Loh says that in his experience, deep learning (a subset of machine learning, which is itself a subset of AI) has been the most applicable AI technology to FX trading because it is able to look at different layers of data and subsequently find more subtle patterns than what he terms as “single-layer models”.
“If you think about face recognition technology, the first layer that the machine might look at would be something a little more abstract, a higher-level concept, like a curve. The next layer might then look at individual features, like the eyes or the nose. Then the next layer would start looking at whether the eyes are juxtaposed to the nose, etc, and each layer builds up on the preceding one, increasing the probability that the computer will be able to recognise that face.
“Bringing this back to trading, by compiling these layers, the computer is able to understand subtle nonlinear relationships. For example, if India is very sensitive to oil, if there’s a 10% move in oil but oil is moving from $10 to $11, that’s not a really significant move on an absolute level. But if it’s moving from $100 to $110 then, yes, you should perhaps pay attention to oil as one of the factors in your strategy. Whereas if you were using something more simple, like linear regression, it might just indicate that because oil is higher, USD/INR is going to go higher – it doesn’t necessarily take into account the subtleties of where the base price for oil was or how the broader economy is doing,” explains Loh.
Playing Like an Orchestra
“There are a number of ways that we apply artificial intelligence, but when it comes to our trading algorithms, they operate very much like an orchestra,” explains Niklas Höjman, CEO and cofounder of Century Analytics in Stockholm and a former Goldman Sachs analyst.
“Generally speaking, in an orchestra you have two layers: the musicians and the conductor. Whereas the musicians are specialised in specific instruments, the conductor’s role is to steer these musicians so that they output music together rather than just sound by themselves. Our algorithmic infrastructure is set up in very much the same way. So in the first layer we have ‘agents’, which are algorithms that specialise in a specific task and on top of that we have the second layer, the ‘selectors’, which is a reinforcement learning algorithm that takes the outputs from the agents and takes decisions based on these,” he says.
The complexity of these agents, adds Höjman, can vary widely given that some of them are sub-trained with machine learning tools and others are not. The key thing in this setup is the selector algo which, because it is a model-based reinforcement learning algorithm (another specific niche within machine learning), is pre-trained on the outputs of the agents and not the underlying data itself.
“So, given the information from the agents, the decision the selector can take is whether or not to trade, optimised odds trade, ie the risk/reward ratio, and then it’s going to continuously evaluate whether or not to stay in that trade based on market dynamics and the expected returns given those dynamics,” says Höjman.
Underpinning the foundation of Century Analytics’ algorithms is the belief that market prices are driven by emotions and therefore, there exist inefficiencies that can be exploited, and that they can be exploited particularly effectively using AI tools.
“Using AI and machine learning to conduct signal discovery in order to identify irrational behaviours in the market and create the basis of the trading strategies for the agents, and then the policies of how these agents function is all thereafter continually refined through an iterative process between humans and the AI,” adds Höjman.
Looking for Patterns
For Isaac Lieberman, the CEO of New York-based hedge fund, Aston Capital Management, AI in relation to trade really means machine learning, and this he defines as the process of working with big data to understand relationships, patterns and organisational frameworks and then conducting analysis on these to achieve a specific outcome.
“In today’s world we have two things that really benefit us. The first is that we have a tremendous capability to organise and process data, which is in itself the first big pillar of AI and machine learning. The second is that we have the ability to train machines to know what to look for within all this data,” says Lieberman.
Put into a trading context, what this technology enables firms to do is look for recurring patterns within the data and, eventually, teach the machine to identify these patterns on its own. In the case of a quantitative trading firm like Aston, this means using the technology to build algorithms and trading strategies to notice regime patterns that occur throughout the day in the FX market.
Lieberman makes the case that this technology is in some ways particularly applicable to FX trading because of the amount of data that is produced in this market – pointing out that it trades 24 hours per day from Sunday through Friday with data updates that are often published in micro or milliseconds.
“Then think about how many currencies there are, including the cross-rates, and what happens throughout the day, you have all of these patterns and regimes, ebbs and flows where the market is going through these oscillations and then you have breakouts and congestion areas, places where liquidity actually has collisions and places where there’s gaps in liquidity in the market. These exist and they’re constantly recurring all the time. So machine learning and AI is about taking these recurring patterns and organising them in a way that a machine can take real-time data and identify as a leading indicator that a pattern is either pre- to mid-formation or to certify that there was an occurrence, in which case, depending on your strategy, there should be a reactionary type move,” he adds.
Not Just a Black Box
Perhaps surprisingly, all three of these hedge funds are constantly at pains to emphasise the role and the importance of humans when it comes to using AI in FX trading.
For starters, they’re all quick to nip in the bud any suggestions this involves feeding a load of data into a black box. Höjman insists that the iterative process he describes between humans and machine to build the models deployed by his fund is absolutely essential to develop sustainable trading strategies, and that the staff at the fund are always well aware of just what inefficiency in the market these AI driven strategies are looking to profit from.
Meanwhile, Loh states that AI is really about just enhancing the capabilities of the human trader, not replacing them. “The goal is to distill all of your knowledge and your way of thinking about the market into the computer so that it’s almost like a better version of yourself,” says Loh.
Lieberman echoes this idea, stating that his firm aims to use the expertise and instincts of individuals with experience in prop trading and taking risk in the FX market to feed the machine thousands upon thousands of scenarios to help it understand what to look for in the future.
“AI then supplements the intelligence of the human person and can start directing them to take various trading actions. But then, more importantly, whereas traditionally a human trader might use a stop-loss to decide the maximum risk they want on a trade, the machine can monitor market patterns in real-time and decide when to disengage from risk or to engage it further,” he says.
The Human Touch
Seeking to enhance the abilities of the human trader only makes it more important to have people with the right experience within the fund, say all three. This is made more acute by the amount of information available to hedge funds and other trading firms today; on the one hand, having more data to feed into machine learning tools is beneficial to the end result – on the other hand, finding the relevant information in the avalanche of data that exists out in the market can be challenging. Teaching the machine what data is more important than other pieces in the market is where this human experience comes in.
“If you ask a data scientist to produce an AI model for trading, they could probably spit something out without too much hassle,” says Höjman. “But it’s likely to be very over-tuned, over-trained and overfitted without a strong understanding of the underlying markets. The crucial part of this process, and really the value-add, is the feature selection, the weight definition and the framework that needs to be put in place to make that model sustainable.”
Source: BarclayHedge July 2018 Hedge Fund Sentiment Survey
Loh comments: “The challenge is trying to separate the useful data from the white noise out there, but what we’ve done is use our more than 40 years of experience trading the FX markets to highlight for the machine which features are important for a given currency. So rather than having the machine look at 50,000 different data points to get the model, we might be able to give it hundreds instead, which makes it more effective and saves on hours and hours of computing power and number crunching.”
FX also presents unique challenges when it comes to trading using AI tools, however, because it is a largely OTC and bilateral market, in addition to being highly fragmented. This means that, firstly, it can be hard to gather the necessary data from across the market and, secondly, many of the prices quoted are based on the other counterparty’s trading style, flow toxicity and credit profile, meaning that these prices might not actually exist for the hedge fund collecting the data.
Lieberman says that, in actual fact, the first of these challenges can help overcome the second. This is because the fragmentation of the market enables machines using interpolation techniques to establish with a high probability what information it should be using in its models. Giving an example, he points out that a machine could establish what the mid-rate of the market should be for a given currency pair by harvesting a wide range of price data and market signals from a range of different sources, normalising the data and then using interpolation techniques to predict the midrate for that currency pair. While this data might not be as accurate as what Lieberman terms “true, specific organised data”, it uses the “guide rails” embedded in the marketplace to determine what the price should be, something that it wouldn’t be able to do if the data was organised and provided via a central limit order book.
“AI tools aren’t only just as good as the data that you’re feeding into them, they’re only as good as the people that are sifting through them and organising the data,” he says. “Make no mistake, trading using AI is a very labour intensive process. It starts with the actual team, you need people with a very technical skill set that are good at working with the data. Then you have to be able to store all that data, then to call it back and put it into other machinery that can run scenario analyses, you have to be able to tag all the data, then put it into backtesting, then put that backtesting into production and create strategies that will source offline data and connect to the realtime data and manage the risk around that.”
Rather than being like a black box that gets fed data, Lieberman likens the process to a factory, and one that is still very much dependent on human activity.
Source: BarclayHedge July 2018 Hedge Fund Sentiment Survey
Democratising FX Trading
A broader question that hangs over these hedge funds is: will these AI tools allow them to effectively challenge the larger incumbents in the FX market? Or, put another way, will AI ultimately democratise FX trading and level the playing the field, or will it entrench the position of the bigger players in this market?
The argument in favour of the former is that, whether it’s via the rapidly growing use of cloud computing or open source tools or other methods, it has become cheaper than ever to access, store and analyse the data needed to drive the machine learning tools being used by these newer hedge funds. Again, this is why the human element is so important, because in many cases it seems that this is the differentiating factor.
The argument in favour of the latter is that the larger firms still have vastly superiour resources by which to hoover up the most talented individuals with FX and data science backgrounds and employ the most state-of-the-art technology. In addition, they have vastly larger internal data sets that stretch back many years with which to work with when developing AI tools.
The real answer is simply that in some ways AI will act as a democratising force and in others it won’t.
As Höjman puts it: “It depends on what kind of data your models are built on. We’re not in the game of finding new alternative data sources to exploit in order to gain an edge, but if you are in that game, then it’s probably going to lead to consolidation because the firms that have more capital to find and use these new data sources have an advantage. But if you’re more focused on using the data that’s already available in the market, then I think AI will prove to be a democratising force because many of the tools that you need to make use of this data are becoming increasing available.”
Loh also predicts that the answer to this question will depend on what type of approach the fund in question is taking.
“I think that what will happen is that there will be almost a bifurcation in the market. What I mean by that is there will be value generated by the bigger firms by just crunching terabytes and terabytes of data, and then there will be smaller firms that will be more technique focused and perhaps a bit more sophisticated that will be able to come up with value in a different way,’ he says.
As a case in point, he cites Google as an example of a large firm that has been able to take advantage of the troves of data at its disposal to develop some very interesting applications using AI and deep learning tools. On the flipside, Loh points out that DeepMind was able to create an AI “Go” player that was able to beat the top ranked human at the game by feeding the rules to the computer and then having it play against itself, in other words, with very little actual data required.
Is AI Actually Intelligent?
While currency-specific hedge funds using AI, or machine learning, tools appear to be few and far between, this technology has been gaining traction in the hedge fund space for a number of years. In a recent survey from BarclayHedge, 58% of hedge fund respondents said they have used AI for three or more years, while 37% claim to have used the technology for fiveplus years.
Hedge fund managers were among the earliest adopters of advanced algorithms and artificial intelligence techniques, which might help explain why a plurality of survey respondents said they have been using AI for more than five years.
Source: BarclayHedge July 2018 Hedge Fund Sentiment Survey
Notably, 68.8% of survey respondents said that they use AI on assets of less than $50 million, suggesting a disinclination to make big bets on the bots. Another possibility is that smaller funds are better able to execute their trades without giving away their techniques and strategies to competing funds.
“The 56% of respondents using AI/ML suggest we’ve passed the halfway point in the race to digitise alternative investment processes,” says Sol Waksman, founder and president of BarclayHedge. “But we can’t ignore that more than four out of 10 of the survey’s respondents still rely on conventional human thinking to guide their investment processes. Concerns about machines taking over the alternative investments landscape may be premature.”
He adds: “Most of the hedge fund managers surveyed are leveraging advanced algorithms and human judgment to deliver smarter investment decisions. The hedge fund pros we surveyed are not turning everything over to algorithms. Instead, they’re using them to formulate investment ideas and build portfolios informed by data analysis that the human brain could never hope to accomplish.”
The reason why machines won’t be putting hedge fund managers out of their jobs any time soon was rather succinctly encapsulated by the keynote speech delivered at the Profit & Loss Forex Network London conference by Cristóbal Conde, the former CEO of Sungard who is now is an activist investor and member of several boards of start-up financial technology firms. He argued that the development of “artificial intelligence” as it was originally conceived has been a complete failure because there is no real intelligence or understanding behind the AI machines that exist today. Instead, he claimed that all these machines can really do is optimise at a speed and scale that humans could never hope to replicate.
But in the increasingly data-driven world of financial markets, and FX trading in particular, the ability to optimise this data to find new patterns, relationships and inefficiencies becomes incredibly valuable. That is why it seems inevitable that the use of AI tools will only become more prevalent amongst hedge funds, and why firms like Ensemble Capital, Century Analytics and Aston Capital Management might be amongst the vanguard of a new type of FX fund.