The World's Biggest Neuron Network
Niels: Historically, at least, the role price of a market has been the only input in systematic models, certainly in the trend following space. The universe of markets have also been very well defined, being highly liquid, exchange traded, like futures on CME. Tell me, how have you evolved when it comes to the data you use and the markets you trade?
David: We trade a lot of equities and we use a lot of other data sources, basically.
Niels: What could they be?
David: You got me there. (Laughter)
Most of the risk is on endogenous variables like price, intra-relationships between markets, and various convolutions of price, sectors, and this sort of thing. Obviously, we have all the balance sheet data, all the fundamental data, all the weather data, there are all sorts of different types of data. We have a lot of experimental systems with small amounts of money on them. I expect we have one or two bigger allocations with key data inputs, but those I'm keeping to myself.
Niels: What about you Marty? Are you looking in new directions when it comes to data and markets? Maybe I can follow-up because that's my next point I want to ask, is about are you also moving off the exchange? What's the motivation for doing that and what are the risks you have to take into account if indeed you are?

Marty: Well, so the first question is data and the evolution of the trading programs. Of course, we have an appetite for new ideas, new influences on markets, new effects. As David says, "If we knew what the next big thing was we wouldn't tell you and it wouldn't be research." I think there's a lot of hype these days with machine learning techniques and all this just explosion of new data sources that surely the answer is in there somewhere. If you just leave it to the folks at Google the answer will become immediately apparent. My view is it's a little bit harder than that. There's plenty of work to be done and there's plenty of opportunities. So, I'm not going to claim that we've got some fantastic new system that employs satellite data and engages recursive neural net and presto we know what's happening tomorrow and next week.

I think all three of us have a healthy paranoia around operating in the markets, and that comes from the real experiences of thinking it was safe when it wasn't safe.
—Michael AdamSo, no, it is overhyped. On the other hand, it is there. That data exists and there's more information out there than there's ever been, ever. You need to work out how to assimilate, how to digest and how to use that stuff. David: One of the things that experience has taught all of us is danger of hindsight bias or over fitting to data sets. You saw this recently in a rich data set, or maybe 5 or 6 years ago. This Google Trends is a huge and rich new data set, obviously, a vast amount of data using Google's algorithm which forecasts when there are going to be flu epidemics. It made the front page of all the newspapers. It made the BBC News.
This is somewhere between irritating and intimidating when your entire career has been based around time series analysis and you see these claims being made. Obviously what we think is, "I wonder if they've really tested that?" Of course, it fell apart. It didn't work at all. It didn't work at all, because of over fitting. There's an example of Google, a company which is renowned for its engineering ability, mining a rich new data source, getting a massive amount of publicity, and then it completely failed. That's not a mistake that...
Mike: I think all three of us have a healthy paranoia around operating in markets that comes from real experience of, when I gave the example of even in an area where I thought it was safe it turned out that it wasn't safe. If I needed a reminder that one's counterparts in markets are not one's friends then that was the sharp reminder. I think those things have always applied and will always apply. You can't research those away.
David: There's a big difference between counterparty and client. This is something that the investment banks got themselves confused about back in the pre-2008 era. Counterparties are not clients.
Mike: I think all three of us have been good and learned on that score. I don't think that that is going to change anytime soon and I think a lot of the apparently new emerging science in the space is naive with respect to that.
So the number of times all of us would have heard this: A fresh faced researcher comes in the most amazing systematic trading strategy, each trade has a 66% probability of being right and you literally can't lose money. The great thing is I'm going to be trading in the FX markets so it's infinitely liquid and there's absolutely no risk. Then the question I would always observe is, "So how much money can you manage?" Let's say the answer was, "Oh, we can easily manage $200,000,000," or whatever the answer would be. The reply would be, "Do you think it's worthwhile to the people with whom you trade to steal $200,000,000 from you?" To which the answer is, "Yes, I think it is."

With a 64% probability trading signal, that is exactly what they will do because that's their job, which is to remove anything with such a high edge that the signal flow itself is valuable. The markets are brilliant at doing that and that's good because it makes them more efficient. So, when looking for systematic trading strategies what you need to do is find something that is fantastically mediocre. Because it's not useful as an individual signal to your counterparts. That means computers are really bad at finding those things because it's a real self-discipline to understand that that's the challenge you face in systematic trading. It's not to be really, really good. It's not to be useful to markets.
Niels: It sounds to me like you're a little bit sceptical about your machine learning.
Mike: That's why it's a difficult discipline.
Marty: Sceptical is putting it too strongly. It's cautious. Interested....
Niels: Are any of you using machine learning today?
David: Not in the trades actually, but in research. Niels: In research. OK.
Marty: In some of the peripheral areas, you know how we allocate to different markets, like David, not in divining what's going to happen tomorrow.
David: I think I could make a plausible claim without stretching it too much that we were using machine learning 30 years ago. The research work we were doing in the late 80s is describable as machine learning.
Machine learning is a sub discipline of statistics and data science, isn't it? Machine learning is neural networks, it's definitely deep neural nets. It's a branch of science and what we were doing is a subset of that branch of science. If you go into the machine learning courses at university, today, they have lots of stuff on neural networks and financial markets and this sort of thing. But there are lots of other algorithms you can use in machine learning.
Mike: I think it's really interesting what's happening with Google Translate in terms of neural networks and that's a perfect example of where neural networks are incredibly powerful. There are no truly catastrophic outcomes, on the mistranslation of a piece of text. Admittedly you could say what if it was actually used by a machine to then fly a plane into a... Yes, you know, you could invent one, but broadly speaking there are no catastrophic outcomes.
But if you apply the same logic to financial markets and don't take into account the fact that human beings and greed are involved. I neither mean greed is bad nor greed is good when I say that, but people are highly motivated to find a way to make money from their trading counterparts. In that world a machine that learns how to do something in a theoretical world and then does it in a practical world is almost certainly going to have its head handed to it on a plate. That's what markets are brilliant at.
Markets are actually neural networks. They are hundreds of thousands of people motivated to make money, deploying capital and taking risk with a view to playing a game against each other in which they hope to be on the winning side. That's a neural network. That's a neural network operated by human beings and that has proved, through history, to be unbelievably efficient... unbelievably efficient at taking money
David: The stock markets are vastly powerful.

What we're trying to do is trade as many diversifying opportunities as we can in as broad a set of opportunities. If there were no liquidity constraints, we'd trade everything we possibly could, almost in equal quantities.
—Marty LueckMike: Maybe Google can replicate that number of actors motivated, but the way around I would put it is the world's biggest neural networks are already the markets and they are unbelievably good at what they do, so beware.
Niels: Yeah, absolutely. Very fascinating.
Mike: Do you agree David? That's what they are. As a bunch of brains - that's neural. (Laughter) With competitive algorithm where the survivors get rewarded with more money and capital. Which is exactly what a neural network does, it's exactly how it works. Human beings, it turns out, have quite powerful brains.
Niels: Good. Well, let's jump to another topic that I think our listeners will learn a great deal from. It relates to the importance of asset allocation. I think Ray Dalio, who runs the largest hedge fund in the world, describes asset allocation as the secret to his success.
How would you describe it? Also, how would you explain the asset allocation process that is built into your investment strategies as well as the benefit that investors have by putting a portion of their investments into strategies employed by your firms? Marty, why don't we start with you on this one,
Marty: Goodness, I guess the starting place is that the way we think about asset allocation is more about creating opportunities. So there's no inherent prediction of which assets are going to be the hot areas. What we're trying to do is trade as many diversifying opportunities as we can in as broad a set of opportunities. If there were no liquidity constraints we'd trade everything we possibly could almost in equal quantities.
Then you have to take into account liquidity constraints and the correlation between those different instruments. Then, what the model does (this is one of the beauties of the trend following approach) is it systematically identifies the opportunity set and does the asset allocation for you, ostensibly. So, I don't view asset allocation, actually, as a separate component of the model. The model dynamically identifies opportunities and moves risk in and out of those opportunities.
Niels: Sure. How would you describe it, David?

David: I agree with Dalio. What we do is asset allocation, our systems are long or short the world's major asset classes and they profit or lose thereby. The difference between Winton and Bridgewater is that Bridgewater, I think, is philosophically based on economics and econometrics. Whereas Winton - I can't speak for Aspect - but AHL and Winton, are more based on mathematics and statistics, I would say. We have never, I can say, had any economics in our models. That may be to our advantage or to our detriment, but I just mention that because that is the difference. Otherwise, we are an identical firm to Bridgewater in terms of we do asset allocation.
Niels: Yeah, absolutely.
Mike: It depends whether you see asset allocation as something that you do before you start trading or something that follows from the way that you trade. I think that's a real misunderstanding about, certainly in what AHL through the years have done, and I know that Aspect has done, which is as Marty says, asset allocation is a product of a systematic approach to trading as opposed as an input into it. The whole point is, you're taking as far as you can, an equal risk allocation to markets but you instantaneously look at where your capital is deployed, it shifts like the shifting sands. That's the point.
It's moving money around very, very efficiently in a very evenhanded way without needing an analyst to make some call that the next big thing is going to be whatever the next big thing is going to be. That's a product or output of systematic trading, not an input into it. I think confusing those two things is very challenging. So, allocating between different systematic trading strategies is an extremely delicate and difficult thing to get right.
Niels: Is there any point where diversification, which is as you mentioned, it's the only free lunch - at least that's what we're being told in finance - is there any point in time where diversification becomes de-worsification where you cannot add more markets or models and get an advantage out of it. David: Well, there's a mathematical answer to that question. What you're looking for when you're building a highly profitable portfolio are things which have a positive expected return and low correlation with the other things in your portfolio. But you never know what something's expected return is. There's always an uncertainty associated with that expected return. That uncertainty may be greater than the expected return. The expected return, the forecast return might be1and the uncertainty might be 10. In other words, over 10 years you have no certainty there's a new thing tied to the portfolio that is necessarily is going to make money, even in the next 10 years.
So, where all the quantities you are estimating - the correlations, the expected returns, the uncertainties - there's a lot of uncertainty in building. In knowing whether a new thing added to your model. Whether you're putting it in with the right. If somebody tells you what the return properties of the new assets are, and the return properties of your portfolio, then there's only one answer. If the return is between 0 and 1 then that will always make your portfolio incrementally better. It doesn't mean you should always do it. I think, mathematically, that's probably true.
Niels: Sure. I want to shift gears and I want to address the low return period that our industry has been in - sort of a drought of 5 or 6 years in terms of returns. David, you've studied market history going back hundreds of years. Can I ask you whether you can put this kind of market environment that we're in, in historical perspective? What do you think is happening in the market in this area right now?
David: I think, as Mike said at the beginning, we were trading quite fast and we didn't think that the opportunity would persist for a particularly long time. It's proven remarkably persistent. But over the years, the faster trend following systems that we used to use are not profitable anymore. I never believe it to be static. In my view, these trading systems do get worse with time. I don't believe that our forecast Sharpe ratio from trend following or forecast quality of trend following returns is (given by the last 30 years simulation) I believe it is worse than that, which is why I believe you have to innovate and you have to struggle to innovate. I think it's getting worse. I think that what we're seeing is consistent with it getting worse.


Allocating between different Systematic Trading strategies is an extremely delicate and difficult thing to get right.
—Fname LnameDavid: I agree with Dalio. What we do is asset allocation, our systems are long or short the world's major asset classes and they profit or lose thereby. The difference between Winton and Bridgewater is that Bridgewater, I think, is philosophically based on economics and econometrics. Whereas Winton - I can't speak for Aspect - but AHL and Winton, are more based on mathematics and statistics, I would say. We have never, I can say, had any economics in our models. That may be to our advantage or to our detriment, but I just mention that because that is the difference. Otherwise, we are an identical firm to Bridgewater in terms of we do asset allocation.
Niels: Yeah, absolutely.
Mike: It depends whether you see asset allocation as something that you do before you start trading or something that follows from the way that you trade. I think that's a real misunderstanding about, certainly in what AHL through the years have done, and I know that Aspect has done, which is as Marty says, asset allocation is a product of a systematic approach to trading as opposed as an input into it. The whole point is, you're taking as far as you can, an equal risk allocation to markets but you instantaneously look at where your capital is deployed, it shifts like the shifting sands. That's the point.
It's moving money around very, very efficiently in a very evenhanded way without needing an analyst to make some call that the next big thing is going to be whatever the next big thing is going to be. That's a product or output of systematic trading, not an input into it. I think confusing those two things is very challenging. So, allocating between different systematic trading strategies is an extremely delicate and difficult thing to get right.
Niels: Is there any point where diversification, which is as you mentioned, it's the only free lunch - at least that's what we're being told in finance - is there any point in time where diversification becomes de-worsification where you cannot add more markets or models and get an advantage out of it. David: Well, there's a mathematical answer to that question. What you're looking for when you're building a highly profitable portfolio are things which have a positive expected return and low correlation with the other things in your portfolio. But you never know what something's expected return is. There's always an uncertainty associated with that expected return. That uncertainty may be greater than the expected return. The expected return, the forecast return might be1and the uncertainty might be 10. In other words, over 10 years you have no certainty there's a new thing tied to the portfolio that is necessarily is going to make money, even in the next 10 years.
So, where all the quantities you are estimating - the correlations, the expected returns, the uncertainties - there's a lot of uncertainty in building. In knowing whether a new thing added to your model. Whether you're putting it in with the right. If somebody tells you what the return properties of the new assets are, and the return properties of your portfolio, then there's only one answer. If the return is between 0 and 1 then that will always make your portfolio incrementally better. It doesn't mean you should always do it. I think, mathematically, that's probably true.
Niels: Sure. I want to shift gears and I want to address the low return period that our industry has been in - sort of a drought of 5 or 6 years in terms of returns. David, you've studied market history going back hundreds of years. Can I ask you whether you can put this kind of market environment that we're in, in historical perspective? What do you think is happening in the market in this area right now?
David: I think, as Mike said at the beginning, we were trading quite fast and we didn't think that the opportunity would persist for a particularly long time. It's proven remarkably persistent. But over the years, the faster trend following systems that we used to use are not profitable anymore. I never believe it to be static. In my view, these trading systems do get worse with time. I don't believe that our forecast Sharpe ratio from trend following or forecast quality of trend following returns is (given by the last 30 years simulation) I believe it is worse than that, which is why I believe you have to innovate and you have to struggle to innovate. I think it's getting worse. I think that what we're seeing is consistent with it getting worse.
Download "Trend Following" Ebook
Listen to the Full Podcast Episode
More from AHL
View All podcast episodes