At the Profit & Loss conference in Singapore Damien Loh, CIO of Ensemble Capital, talked how AI tools can be applied to FX trading.
Profit & Loss: Artificial intelligence has become a big buzzword in finance. What does this term mean to you and how are you actually using this technology?
Damien Loh: So it's a very buzzy term, and that's why we make a distinction by saying that we're using deep learning. AI can just be a general definition where any process that requires thinking is automated in some way and there are people using the AI catchall buzzword that just have a linear regression in an excel spreadsheet.. And you can tenuously call that AI, but it's nothing really game changing.
Within the subset of AI, there's machine learning and then a subset of machine learning itself is deep learning. And deep learning is really where the advances that have marveled the world are really coming from, by which I mean things like automated driving and the computer, AlphaGo, that beat the human Go player.
Deep learning architectures are fundamentally different from the technology available before because the cooperation and open source frameworks that are used in these architectures mean that, rather than just having a single layer model to forecast markets, they can build upon multiple layers and get from simple concepts to very abstract ones as a result. An analogy for this could be machine vision, where the first layer would be to learn to identify specific concepts like edges and curves, the next layer would look for something higher up in terms of abstraction, like eyes and noses and then the subsequent layers look at the juxtaposition between the different parts of the face to recognise who it is.
P&L: So what part of the investment and trading process do you use these deep learning tools for?
DL: Where we found the most value for it is being able to look at all the different features of different things happening in the market and being able to feed all that data into our models and then forecast on the back of that. Humans have traditionally been staring at the Bloomberg screens and trying to aggregate all the information on there, to do this consistently in a non-linear way is possible with a team of people but it’s fairly hard and this technology is something we can use there.
For portfolio construction the technology is applicable but there it’s actually a subset of deep learning called reinforcement learning where you tell the computer what factors you care about and provide it with certain constraints. This can be tricky and requires a lot of work because if you say, for example, to just make as much money as possible then the machine might just put many times leverage in dollar-Turkey and put all your AUM into that, which is clearly not what you want to do.
So you need to be subtle but also be very explicit with the parameters you set. So you say that I don’t want to have too big of a drawdown, I want to have a Sharpe ratio that’s at least above two and a Sortino ratio of this. You have to be very explicit and not take for granted that the machine understands what you want.
P&L: When we talk about AI there often seems to be something of a misconception out there that people can just feed data into a black box and it will spit out trading strategies. So talk to me about the human element involved in using AI to trade.
DL: Yes, because of the buzz around AI some people think that it can do your homework for you and you can just sit back and relax on a beach somewhere. Maybe that will be true fifty years from now, but we’re not there yet.
Even though some of the concepts around AI have been around since the 70s it’s still a very new field and we’re still learning about the capabilities and limitations of this technology. Where it doesn’t do well is when there’s some form of episodic event, by which I mean there’s an earthquake or a major US election, something that is hard for it to learn from because it only occurs once or occurs too infrequently for it to be reinforced in a consistent and continuous manner.
So in those cases it’s important for a human to be able to come in and make a judgement. The other thing that humans are good for is reducing the amount of data that the machines need to process. If you had infinite resources then you could just input as much data as possible into the machines, but we have finite resources and time so having a human look at the inputs and whittle them down from, say, 600 to 300 will speed things up. Are the futures on the CME really related to Singaporean stocks? Probably not, a human would know that but a machine wouldn’t necessarily.
P&L: So then the human works to guide the computer by telling what to focus on so that it isn’t analysing irrelevant data?
DL: That's right. And it's also more of an art than a science because you don't want to be too heavy handed in that aspect because there are subtle patterns that the machine will be able to figure out that you might not have ever conceived of yourself. And that's the whole point of deep learning, it’s able to think in a very different way than humans and that's where your edge comes from.
P&L: So let’s talk about the challenges associated with sourcing the data needed for these AI and deep learning tools. Do we have a big data problem? What I mean by that is that the sheer volume data that exists in the world and online is growing exponentially, does this make it harder for you to find and feed the machines the data they need?
DL: So again, this is where the art comes of being able to find a happy medium in terms of putting in a lot of data in but not everything and the kitchen sink at the machines. The other part where the human element can be useful is by not just throwing in the data but also guiding what to look for. So the human can say look at the yield curve and just put in the two year yield and 10 year yield separately, but what’s also important is the two year/ten year yield as a spread, so they should probably put that in as a separate feature as well. It’s not always the case that a deep learning model can figure out to combine various features so it can be good to spoon feed it a bit.
P&L: Are there any data challenges unique to the FX market? Given that FX trades largely OTC bilateral prices vary based on the counterparties trading so even if you were to source this data some of the prices might not be relevant for your firm.
DL: Yes, but I think this is something that a lot of quantitative funds have to grapple with. There’s always a question of whether the data is applicable and because it’s a largely OTC market you’re never going to be able to get crucial data like volume unless you spend a lot of money. So yes, there are constraints and challenges but it’s not really a problem that’s specific to using AI.
P&L: Is another challenge that a lot of the market data isn’t static? You gave the example of using deep learning to recognise a human, but the data points there aren’t constantly changing the way that price data in the market is.
DL: It can be a challenge and it’s why I think that AI might not be that applicable for more high-frequency trading right now. We use it for trying to forecast for a week or a month from now so the minutiae of the price movements on a minute-by-minute or second-by-second basis are not as relevant given our time horizon.
P&L: We touched on the importance of humans helping to guide machines and help them sift through all the data out there, but how do you see AI changing the role of the human trader more broadly?
DL: I think what will happen is we will see a bit of revenge of the nerds. There’s going to be a lot more quant-type people on the trading floors and a lot less execution-type people. There will still be a requirement for traders with experience to be able to look at a situation, see different markets and be able to make adjustments to trading based on risk management perspectives but I think you will see quants move from being a fringe part of the floor to maybe 50% of the overall traders.