In partnership with:
CME Group

87 An Engineer Uses Machine Learning in Investment Management with Andrew Baxter of Cambridge Capital Managment – 1of2

"I’m an engineer. I always have been." - Andrew Baxter (Tweet)

Andrew Baxter worked at British Aerospace as an engineer before joining the investment management world. He still considers himself an engineer.

Listen in to learn how he uses machine learning, why he keeps innovating, and how he started his own firm.

Thanks for listening and please welcome our guest Andrew Baxter.

Subscribe on:

Subscribe in iTunes Subscribe on Stitcher Subscribe on Soundcloud

In This Episode, You'll Learn:

  • Why Andrew considers himself an engineer
  • How he went from engineer to investment management
  • How he started his firm

    "We launched and started trading 3 years ago." - Andrew Baxter (Tweet)

  • His philosophy on the investment management process
  • Why he harvests risk premiums
  • What he likes to do when he is not working
  • How he works with machine learning

    "There is no single right solution, particularly in finance." - Andrew Baxter (Tweet)

  • The things they learned from their career history that helped them build Cambridge Capital Management
  • Why it is important to stay ahead of the pack and keep innovating

    "Data is often very very noisy. If you don’t know what the underlying process is it can look just like noise." - Andrew Baxter (Tweet)

  • What they do in-house versus outsourcing
  • The culture that he tries to cultivate at his company
  • How he talks about track record to his investors
  • How he evolved as the managed futures industry changed over the years

    "The world of investment management has become far easier for people to get started." - Andrew Baxter (Tweet)

  • What he does to manage risk
  • The objective of his program from a top-down view
  • The different markets that his program trades

    "We are a small firm but our systems and processes are institutional." - Andrew Baxter (Tweet)

Resources & Links Mentioned in this Episode:

This episode was sponsored by Eurex Exchange:


Connect with Cambridge Capital Managment:

Visit the Website:

Call Cambridge Capital: +44 (0) 1223 851 001

E-Mail Cambridge Capital:

Follow Andrew Baxter on Linkedin

"Performance is a blend of skill and luck." - Andrew Baxter (Tweet)

Full Transcript

The following is a full detailed transcript of this conversion. Click here to subscribe to our mailing list, and get full access to our library of downloadable eBook transcripts!


Welcome to Top Traders Unplugged, where my goal is to give you the clarity, confidence and courage you need to invest like or invest with one of the top traders in the world. It is the stories that you never get to hear set out as the most honest and transparent account that I can make of what goes on inside the minds of some of the best investors in the world. Today you're listening to episode 87. If this is your first episode you've heard, you might want to go back and listen to all the earlier conversations. 

But before we find out who’s on today’s show, I wanted to mention that today’s podcast is brought to you by the Eurex Exchange. Given all the market talk at the moment about rate hikes, you may find some useful ways of hedging your portfolio risk if you visit the Eurex website. 


This is Andrew Baxter. I’m the CIO and co-founder of Cambridge Capital Management and you’re listening to Top Traders Unplugged. 


Thanks for doing that Andrew, and by the way, if you want to read the full transcript of today’s episode, just visit the TOPTRADERSUNPLUGGED.COM website and sign up to receive access to all of them. Now let’s get started with part one of my conversation. I hope you will enjoy it. 

Andrew, thank you so much for being with us today. I really appreciate your time. 


You’re very welcome, Niels. It’s a pleasure. 


Good. Now today’s conversation I think will be really interesting for a large part of my audience who are either professional fund managers, or allocators to hedge funds, or perhaps on the other side of the spectrum, I believe we have a large group of people listening who would like to become a fund manager themselves, or managing their own investments. What’s really exciting about our conversation is that you can speak to both parts of the spectrum. 

For those who may not have come across you before, you have in fact been running a billion dollar plus portfolio before founding CCM where at the moment at least, you are managing a somewhat smaller portfolio. We will, of course, get into all of this. I just wanted you the listener to know that it will be worthwhile sticking around and learning from Andrew’s wealth of experience. 

Before we jump into all of this I just have a simple question that I try to ask all my guests in order to appreciate the many different answers that you get to this question. Basically it goes, if you meet someone who doesn’t know you and they ask what you do, how do you respond to them? How do you explain what you do? 


Good question. Well, I’m an engineer. I use my skills and experience and apply that in the world of investment management. I build portfolios of assets to generate an investment return: when to buy, when to sell, how much of each asset to buy or sell is based on the forecast that we generate of the future movement of those asset prices. Those forecasts are generated using quantitative models written in computer code. 


There we are. Thanks very much. Now before we go on to the usual questions, I want to try something a bit unconventional with you because I asked you, as I do with everyone, if you could send me a brief bio. Usually what I receive is to be frank, pretty standard. It is very factual, but they don’t show a lot of creativity, if I can call it like that. However, what you sent me was very different. It was like reading the beginning of a book. So if you don’t mind, I would love for you to share this creative bio and read it to our audience as if you were reading it to your children with a passion and excitement that comes with that. As you know, people don’t remember PowerPoint presentations very well, but they do remember stories so are you up for this Andrew? 


 Yes, thank you for the opportunity, Niels. It wasn’t with this intent when I wrote these words. It was to give you an opportunity to get inside my head and understand my journey, but if you think that’s useful to your listeners, I’m very happy to do that. 


Absolutely, let’s open up your mind to everyone. So I’m going to give the floor to you and why don’t you just read what you sent me. 


Thank you. So I’m an engineer. I always have been I still consider myself to be as such. I happen to apply those skills in the world of investment management. I majored in engineering at the University of Cambridge and spent five years at the British Aerospace Military Aircraft, 10 years in complex derivatives market making and risk management, and the last 10 years in quantitative investment management. 

At CCM we built our firm from first principles and this is my story: from engineering to investment management. I’ve always had a passion for problems, how things work, building things, and when I was younger I was always taking things apart to understand how they worked and hopefully putting them back together correctly. Before the world of computers, I was building electronic circuits, soldering irons. Then computers came along and I was probing those early computers, I guess when they hit the home market in 1980 in the UK. I built remote control aircraft from plans, paper plans and pieces of balsa wood. I had a passion for flight and would have loved to be able to fly fast jets in the military, but short sightedness put pain to that. 

But I pursued this passion for engineering, and went to Cambridge University where I read Engineering, and followed my fascination of flight by joining British Aerospace in their military aircraft division here in the UK. At that time that was probably the nearest thing we had, in the UK, to what NASA is in the U.S. It was technologically fantastically advanced engineering opportunity. 

I worked at British Aerospace for five years and was involved in some really amazing research work. We worked on complex computer models, in those days it was main frames, but we were simulating combat scenarios. What is the optimal specification for a future aircraft? We worked on an aircraft designed to be delivered some twenty-five years hence. Is it small, nimble, agile, difficult to detect on the radar combat aircraft; or is it a bigger aircraft that can carry bigger payloads that has bigger engines, bigger fuel burn? These were the questions we were playing with. 

It was an amazing place to work. We had post cold war interceptor aircraft. We had lightning aircraft flying where I was based, was the UK flight test center. We had aircraft flying that could go from zero to 50,000 feet, 50% higher than what we fly in our 747 in less than 60 seconds. I did have the opportunity to briefly get hands-on with the aircraft we were building and there are bits of my work flying around somewhere in a cockpit in a military aircraft somewhere in the world. 

That was the 1990s. 1991 on was a period of rapid change. The Berlin wall came down. After five years at British Aerospace, I was asking myself why are we building these aircraft? Who’s our enemy? The timescales were very long. It was very difficult to have a real sense of the day to day contribution one was making. I was drawn to that urgency and excitement and engineering challenge, and I guess that had disappeared somewhat. 

I looked round and I saw innovation in the world of finance, and derivatives in particular, and it was at that time that I joined JP Morgan back in 1994. I spent the first ten years in currency derivatives at JP Morgan, Bank of New York, where I worked with the Susquehanna Investment group. One of the very first three groups to come off the exchanges into the OTC market and bring with them the derivatives trading skills. I joined Deutsche Bank, where I traded complex or exotic currency options. Options where the final value is dependent on the path the underlying instrument takes between trade date and exercise date. 

I used my computing skills in building pricing models for the complex derivatives. At the time the theoretical price from black controls was well understood, but came with a lot of assumptions, assumptions which were not practical in the real world. We had to transfer the theory to the practicalities of how do you price these complex derivatives? How do you incorporate the cost of hedging and risk management and make them usable in the real world? I wrote pricing algorithms that allowed the trading desk to deliver consistent tradable prices across a wide range of instruments, through some twenty traders globally and 200 sales people globally. It was an implementation of the theory problem. It was an engineering problem. 

I ran a complex books and 10,000 positions: everything from one day out to 10 years in this currency derivatives book. It was a real challenge to manage this risk. Derivatives is a solution. It’s a solution for the customer. Some people would say that derivatives are evil. Well, they’re not. We all have derivatives. 

You have insurance on your house, or your car. It’s a derivative that pays out if something happens it’s derived from some other event. So we were delivering solutions to customers and our biggest market was the U.S. and I moved to the U.S. and spent two years in New York with the North American clients helping them access this world of currency derivatives and it was there where I got involved with the investment management business. 

I spent two years with proprietary trading strategies that we were building on the desk and made them available to the ultra-high net worth space in North America. That got me thinking that this is a really interesting business. Derivatives were starting to mature. It sewed the seed of an idea for me to build an investment management business. When I returned to the UK, at the end of my overseas placement, I had the opportunity to do that at HSBC. 

At HSBC we started with nothing. The first guy through the door - I built trading models. I built a team. I built the business, the legal, the infrastructure, the whole framework to deliver our activity. Liquidity aggregation from the exchanges, algo execution and a sales and distribution channel. We were trading currencies, rates, equities, futures, over the counter foreign exchange. We traded equity start-up models - some 6,000 cash equities in the portfolio. 

At over five years and 1.8 billion dollars we’d raised and managed over that time, we were incubated. The business was built. We had great performance through 2007 and 2008 and we proved the quality of our approach within challenging market conditions. We built the business and incubated in global markets. We moved across to the investment management division. For many people that’s hard. There are many people in the investment management business who have a proprietary trading strategy and it works very well, but that’s not an investment strategy. 

For many people they may have an investment strategy, but that’s not a business, or they don’t have experience of how to build and run a business like that. We, the thirteen of us, we’d proven we could build and run a business and my baby had grown up, graduated investment management, and I left HSBC. To get together with my CTO from HSBC business, Ian Brown, we co-founded Cambridge Capital Management. That was 2010 and it’s been a long a steady build since then. We launched and started trading three years ago, just this month. The numbers have been very good. 

Our philosophy, the one we’ve pursued, I’ve pursued throughout my career - applying engineering techniques and the control and understanding of risk to the investment management process. We seek the best proven approaches from all areas of investment management. We’re not wedded to any particular idea or a particular style of a previous investment firm. 

As engineers we expect the solution to be good enough, but not right. There’s no right solution, particularly in finance where the answer we’re trying to solve is continuously, continuously adapting. We’re not held back by having a solution that’s less than perfect and many people, from the hard sciences in particular, find it difficult to move into the world of finance where there is no black or white. There’s no right or wrong answer. I think it’s particularly appropriate or conducive to an engineer that is willing to accept something that is less than perfect. 

We took investment technology from the large scale equity portfolio world, from the start-up techniques and techniques from engineering, signal processing and data analysis. We’ve been working closely with the University of Cambridge, here and the engineering department for the last three or four years to transfer over some of the techniques which may be well understood and relatively mature in other problem domains, but are not known in finance. This is a major edge for us. Not the techniques themselves, but the willingness, the humility to appreciate that the techniques when used today are not perfect. One needs to be continuously improve, revisit, and remain open minded. 

What do we do? We harvest risk premier, the same risk premier that many CTAs and other managers harvest - risk premier that are driven by macro factors. They’re plentiful, they’re persistent in the long run, and there’s a wide variety of techniques to access those. 

We model those premier, we build forecasts, and we build risk adjusted portfolios. We attempt in our approach to identify the failings in our approach, the risk, the shortcomings, the source of errors in the many layers of our systems. We measure that, we monitor, and we reduce where possible, and we manage that. 

We recognize the need for continuous improvement as the world around us changes, the techniques available improve, and our understanding of the problem improves. We iterate, iterate, iterate, all within a process. We use technology to deliver that process through a robust process we deliver investment performance. Better tools, process, understanding of the problem, and management of the risk and uncertainty - we will have better outcomes. It’s performance through process. That’s my story and I’m sure we’ll draw down into that in more detail. 


We certainly will. Thank you ever so much. Again, I hope people just take note that sometimes when you do something that’s a little bit different in the way you present yourself it draws the attention of the person that you want, so I thought that was a great way of doing a creative bio. So I said before that we’ll go back and revisit some of the things, but it is quite comprehensive in some ways. 

I just want to ask you more simply, when you look back at all the things that you’ve done and that you just talked about, to where you are today, where do you think you learned the most? Where do you think you took along most of the things that you’ve found really useful for where you wanted to go today? It’s such a broad background, from aircraft to finance and how do you frame that? 


 It’s a broad background but there’s some common threads. We have a problem. It happens to be an investment problem. We have a tool kit. We have a tool kit of models, we can build portfolios, we can measure signals, we can measure noise… In the derivatives world one is providing a hedge, trying to reduce the risk on a single or a series of cash flows for a corporate investor. One has a derivative risk manager or trader or structure. One has a tool kit. It just happens to be a derivative tool kit that builds cash flow that behave in certain ways. You can put all those together in a package. Your customer says, “Well that was a little expensive. How can I reduce that?” Well if you… is there any risk that you’re happy to take? I’m happy to take this risk, in which case we can reduce the cost of that package akin to insurance on your car, if you’re willing to pay the first X yourself as an excess, then your insurance is premier. 

It’s all building blocks, and the investment management process is the same. We have markets we need to try and understand what drives those markets. We then can try to model those markets. We need to recognize those models don’t really work very well. They are an approximation to reality. We need to understand why and in what way those models don’t work and then we can address that. We can diversify away the risks that it’s possible to diversify with. For those risks that we can’t, we try and control them. In the world of derivative risk management, the theory says the price is X but the reality is very different. The theory is conveniently phrased or constructed to ignore the fact that markets, for instance, don’t trade continuously, they have jumps. The returns are not normally distributed, and you're not hedging, and so on. 

So with investment management, I think the thing that runs through, to answer your question, is it’s an engineering problem. We have a tool kit. We recognize the shortcomings, we measure and monitor those risks and we control them. 


Sure, I appreciate that. Now I’d like to talk about a bigger theme that I would love to hear your opinion about, but before we do that, you’ve obviously told us a lot about your background, but I just want to bring it up to today and I want to give the audience a chance also to get to know you a little better. So before we dive into the technical part, just sort of very basically, clearly being the CIO of Cambridge Capital is a full time job, no doubt, but what do you like to do outside the world of finance? What do you like doing when you’re not working. 


For me, I’ve always loved the out-of-doors: mountaineering, rock climbing, diving, skiing, and I still love aviation. I have a pilot's license and I fly. I’ve recently discovered the world of cycling which I think, as I read recently, cycling is the new golf. It’s certainly in the UK. It’s a very popular sport and I’m sure that’s helped by the success of the Olympic squad in international cycling in recent years. It’s a great activity for someone whose knees aren’t as young as they used to be. There’s a theme there: outdoors, mountaineering, climbing and so forth. 



Now, I want to talk about machine learning, or as some people would say, artificial intelligence. It’s something that people nowadays it’s a term that they use more and more, and in different industries it means certain things and so on and so forth. Since we’re going to touch upon that later, I just want to find out, what does machine learning mean for you? 


We’ve been working for the last three or four years with the professor of the machine learning group here at the University of Cambridge. I guess, one of the great things with working with the gentleman whose name is ZoubinGhahraman … The great thing with working with Zoubin and his team is we have access to a wide variety of ideas that fall under the brand, if you will, of machine learning. There are lots of ideas and techniques. It’s not one single algorithm. There’s the right tool for a job and often the challenge is finding which is the right tool for the job you’re trying to do – for the problem you’re trying to solve. 

First of all, one needs to understand the problem itself. Then finding someone who understands the wide range or tools and their strengths and weaknesses and try and apply that tool. I think there’s many people who have some exposure to machine learning and are disappointed and there can be a variety of reasons for that: Often I think it’s because people have maybe hired a PhD. from a data science area, given them a problem to solve, and the result has been disappointing. Did they use the right tool? Did they model the data? Did they try and understand the problem? I think that’s one of the challenges. 

I can tell you a story. There’s a very interesting problem they have in their department and it’s a bit of a toy problem, but I think it’s a good illustration. In engineering we’re not limited by data because if you have a physical system, you’re trying to build a better suspension system for a car, every time you drive it around that racetrack it behaves the same way on the same corner. You can hone the suspension; the electronics may be able to help that suspension work. In the same way if you have any physical system that is repeatable, the thing is we’re limited in finance to the data set that we have, and each day that goes by we get a little bit more data. 

Also, the finance data is changing over time. We only have to look back ten, fifteen years ago we had voice broking, now we have machines making markets. So how far can you look back in the data? Also you’re actually in the market by trading influences potentially tomorrow's data and our competitors do the same. 

So in the lab they have a trolley that goes back and forward on a track and beneath this trolley hangs a pendulum, and they have a machine and they say to the machine you know nothing about this and actually all you have is a camera and a camera watches this trolley move around. From the camera you can see this pendulum, it has to recognize the image. It sees the pendulum and your challenge, machine, is to balance the pendulum, just like you would balance an umbrella on the palm of your hand by moving left or right, watching the tip of the umbrella, trying to keep it vertical and it does the same with the trolley moving left or right. This runs for hours, upon hour, upon hour learning to initially swing the pendulum, get it past horizontal, get it vertically, and then keep it there. When it’s learned that, they add another pendulum to the bottom of the first pendulum and repeat, and given enough time it will balance these pendulums. It has learned, the machine has learned, but it’s not limited in data and we can’t do that in finance. 

One of the biggest challenges is taking techniques that require vast amounts of data. Data can often be very, very noisy in fact… is very, very noisy in fact and we’re limited in quantity. The failing is to learn the data. One should learn the system that generates the data. So if you don’t know what the system is, it can just look like noise, and a lot of people are putting data into machine learning techniques and they’re just trying to fit noise. One of the big challenges is not to over fit, think you have a solution and have actually just learned the noise of the system itself rather than underlying drivers. 


Sure, sure. 


 So that is the challenge. 


Yeah. I’m also trying to apply this to the investors that you and other firms in the same field are trying to attract. I’m just thinking back on my own career. It stems predominantly from trend following, and I don’t find that a particular difficult strategy to understand. You buy the highs and you sell the lows which is opposite of what most people think you do in order to make money, but essentially that’s what trend followers do. 

But I’m just thinking how do you bridge the gap when it comes to machine learning to the investor? Because we know that investors are very good at saying, “No thank you,” if it’s something they don’t understand, rather than admitting that they don’t understand it. I can imagine that when you go into your world, which is a bit more complex than just basic trend following that this is quite a challenge. How do you deal with that? 


 I think as an industry, I think we are guilty of not making enough effort to help our investors understand what we do. What is the saying? We should be able to explain what we do to a six-year-old grade student. Call it what you will, but if you can’t explain it simply enough, then you don’t truly understand it. I think sometimes there’s an element of that in this industry and I think there’s an element of feeling one doesn’t want to disclose what it is that we do, because if investors knew that it could be written down on a piece of paper they wouldn’t invest because they wouldn’t believe it was worth paying for. 

I challenge any manager not to be able to write down what he does on one side of a fold. Getting that to an investor doesn’t destroy the purpose of that business. I think it’s quite the opposite. It makes the investor understand that the manager can articulate what he does. We could write down… we’re all using the same ingredients, but the difference is the implementation and how that is managed. I could have the recipe for a Michelin starred restaurant and I couldn’t produce their food, nor would I want to. 

So machine learning is putting… for us at least, it means having an amount of adaptability within the quantities and relative quantities of the ingredients that we use within our recipe. So we model, I don’t know if we’re getting ahead of ourselves here, Niels. 


No, no, that’s fine. Go ahead. 


 For us, we believe that there are… we forecast asset prices. We don’t forecast the prices directly. I think they’re very noisy. We know that markets have different regimes, and some regimes they’re… some drivers of markets have more influence, in other regimes different drivers have more influence. So we try and model the underlying drivers of price returns. We model a number of different drivers and we recognize that in some markets you need to weight some drivers more highly than others. 

It’s the adaptability – the weighting of those drivers for us, where machine learning comes into it. Let me give you an example: prices are very noisy. It’s hard sometimes to see what’s going on, but if we knew the structure. If we knew exactly how prices were driven, life would be very easy. 

So I was reminded of mobile phones. Mobile phones have a really weak signal and you’re trying to transmit your phone call in a very noisy environment. How do they deal with that? Well what they do, they say, well I’m not going to transmit… if you think about dialing your old radio in the old days, you got crackles and hiss on some channels. Well the same with mobile phones. So what they do is they transmit across many, many channels and they will move from channel A… channel 1, to 5, to 10, to 3, to 7, to 20 and so on. They will skip between channels just for a few milliseconds at a time, skipping between. 

So they only expose to noise on any one channel for a certain period. In addition, anybody listening on just one channel wouldn’t hear the mobile phone call. It’s skipping around. Only there for a very brief instant. That technology came out of covert security – covert transmission in the military where they were trying to move quickly between channels in a pseudo random way to avoid eavesdroppers. But the effect is that at any one channel the noise has very little influence. So if you listened on channel 1, you really wouldn’t hear this mobile phone call. It just sounds like noise in the background. But if you know the structure, if you know the sequence that it’s going to hop between channels you can reconstruct that phone call and that’s what the base station and handset do. They know the model, the structure of this what looks like random data. 

If we know the structure of what drives the price returns of markets, then that’s what we should be modeling and that’s what will have much more success at forecasting. We know markets move over time. They evolve, they change, we have different regimes, risk on, risk off if you want to call it that. There are different regimes for different point in the economic cycle, markets behave differently when rates are high then when they are low – when we have high inflation or low inflation. So we want to have the ability to adapt to those different regimes all within the system. That’s machine learning for us. 


Sure. Well I certainly learned a lot more about mobile phones just from listening to that, so I appreciate that. I do wish, however, that our Skype connection today would have a few more lines that it could jump through because it’s a bit static, so I apologize for that to the listeners, but we’re doing our best to clean it up when we have finished the recording today. 

Now, Andrew, we’re going to talk about what you have is called the Systematic Global Macro Program, but I just want to take the opportunity to congratulate you because I know you just finished your first three years of live trading and that’s a good milestone. I think many investors will take things a bit more serious once you’ve put the first three years of actual trading behind you, even though, of course, as I have mentioned before, and as people will have realized by now, you are and your team are very experienced in this business. 

Now let’s jump to the next topic because it relates a little bit to that and it’s your organization. I would like for you talk about the challenges you face in trading the things you do, but also how you obviously have taken all the knowledge of managing a very large portfolio with a big team, and how you actually were able to (and I’m putting words in your mouth) but I’m imagining you’re pretty much capable of doing exactly the same today in a different and smaller entrepreneurial environment. Talk to me a little bit about how you've organized yourself to be able to do that. 


 Thank you. When we built CCM we certainly learned from our experience in the integrated model trading team at HSBC. As I said earlier, we built a team and a business from zero - all the complexity, all of the business issues, the sales channel, development and all the technology, everything from collecting data, running models, building portfolios, sending orders to market, dealing - we would trade over a billion dollars a day of currency, and aggregating liquidity from multiple sources to minimize market impact, and all the technology that we had to build around that. We brought with us that experience and it’s certainly the second time or third time around it’s much easier. It took us two years of hard work from the day we started CCM to the day we started trading. Certainly we had to wait for regulatory licenses to go through, but there was a lot of technology and infrastructure and so forth to build. 

We’ve built a much more comprehensive system here. The approach is much more complex and sophisticated than what we used previously. The complexity is managed through technology, so complexity doesn’t necessarily translate into increased headcount or difficulty of applying that. The world of investment management has become far simpler. If I look back, there’s been a lot of change over the last ten or fifteen years. I think some techniques have become much more broadly known. Sophisticated libraries have become more widely available. There’s less need for people to hand roll solutions. That’s made it far easier for people to get started. I think it’s also made it far more challenging for managers who are not adapting and continuously improving their process. We can go and buy a book on Amazon this afternoon on trend following and I’m sure that will produce a model which is relatively good at trend following in markets that are conducive to that. 

It is important to continuously be moving up the sophistication spectrum to maintain… to use the latest techniques to stay ahead of the pack. So it is a lot easier today than it was ten years ago to build a business. It’s still hard to build a business that delivers high quality risk adjusted returns, however. 


 Sure, very true. So which functions do you have in house, if I can call it that, and is there anything in being efficient today and having set up CCM, is there anything that you decided to outsource of what you do? How do you manage that? 


Yeah, we spent… so we’re three years old by track record. As we approach the end of our second year we made a big push to make ourselves accessible by a wider audience of investors. We brought on-board a COO, a very experienced COO, to lead the development of all the business processes that are necessary to support institutional investors. We also brought on-board a dedicated sales person and received earlier this year our U.S. regulatory license. 

So there are some things that have to be done in house: all the research, developments, and software and systems. All the business management is done in house: legal, and compliance; we have external advisors, obviously. In terms of hardware and infrastructure, that is a business risk for us. It’s hard to run technology well and it’s something that we believe can be delegated to others. 

So everything that we run here, runs in one primary and a secondary failover data center. We have dual redundant systems. Within the office her we have PCs. We have technology for software development and email and so forth - anything related to production exists in a robust environment. I think that’s important. It’s what institutional investors expect, so we have a fully distributed and production quality infrastructure just as we would have had in the bank. 

Fifteen years ago that was hard for an early stage manager to produce. It is much easier these days. Likewise, one of the great benefits of working with Ian, my co-founder and CTO, he came by a similar route: A Cambridge engineer, military aircraft. He’s a software guy through and through. You could buy games he wrote on the High Street. When he was a teenager he was writing games for the big software companies. This is his background, working in real time software systems, so visual recognition for autonomous aircraft when he was at British Aerospace. 

He’s been working in the investment space since 2007 when he joined the HSBC. He brings with him commercial real time software development experience. That means that we’re a small firm but our systems and processes are top notch, whether that be continuous testing, automatic deployment of code, database design, and so forth and so forth. 

So one can be small and it’s the way one approaches the problem. We have built the business with a foundation for growth from the beginning. It’s easy to build a strategy, find investor interest and realize that actually one then needs to start building the infrastructure for growth to support those investors, to manage multiple accounts - to have the safety around the implementation of the business that investors expect. We built that from the beginning, for growth, as we saw the benefits of doing that at HSBC. 


 Sure, Sure. A slightly different question but also related to this just to round off the thing about organization. You’ve built a couple of teams before. Do you have a particular culture that you try and create with the teams and businesses that you build? How do you build a strong culture? I can see, as you mentioned Ian, that there’s some continuity there, that you’ve been working together for a number of years and it takes work to… you can put people together but actually building a strong culture between people is not necessarily so easy, so how do you do that? 


We’ve always believed in a collegiate environment – free exchange of ideas and an open sharing of knowledge. It’s very easy to exist in silos, particularly once a team gets beyond a small size. I think it’s often the case that teams coming from an investment banking culture, where there is a lot of benefit in being siloed, find it hard to integrate into a collegiate environment. The value is greater than the sum of the parts. 

The sharing of ideas… one of the great things about the academic environment and working with the team here is it’s a given that people will challenge and thrive off the dialogue that goes on when one’s trying to explore and develop new ideas. That’s what we try to do here. 

We are relatively small, we have four full time members of staff and we have three people in the University currently – a professor and two of his research team that are formally engaged, and we have other members at University doing project work. I would expect that as we grow we will continue to bring people in who have strong technical skills but most importantly have that sharing of knowledge at the core of how they operate. 

I don’t believe it’s necessary to bring people from a finance area. We’re in Cambridge. We’re 45 minutes from the center of London, here. Cambridge is not well known for its finance community and I don’t see that as a disadvantage. We need to use a different approach than our competitors, otherwise we’ll get the same outcome. We want to use new ideas to think about problems. I think that’s one of our strengths that we, from first principles, have tried to build an investment management business without creating a sort of “me too” or copy of someone else's approach, otherwise we’ll get the same outcomes. 


Sure, sure. That’s true. 

Speaking of outcomes, the next area I’d like to talk about is track record, and I don’t mean numbers month by month, but I just mean how do you show investors enough data, given the fact that now you have three years of track? How do you help investors get an in-depth understanding of the program and how it’s going to cope in different environments given that fact that, actually, the last three years have been quite interesting? We’ve had… 2013 was a pretty bad year for CTAs or systematic traders. 2014 was probably one of the best years we’ve had in a very long time. 2015 is shaping up to be somewhat tricky I think for most people. So how do you best manage that and give investors whatever comfort they need with a relatively short live track record, at least at this stage? 


It’s a good question and I think, actually, what I believe investors are trying to buy is skill. Their challenge is how to measure skill when all they can observe is performance. Performance is a blend of skill and luck. Markets, as we’ve said, change and a skillful team will continue to demonstrate, over time, it’s ability to make the necessary changes and the innovation to adapt to those changing conditions and continue to deliver good performance. So how can an investor invest in skill? Well, he can look at the numbers. He can look at the numbers and see whether they are consistent with his understanding of what the manager says he does. He should understand from the manager what he does, and more importantly why he does it, and to see whether it makes sense. 

Why has he chosen to do what he does? Too many… I think many investors will ask what you do, but not why. If they can understand what you do, then they can understand whether it will work in market conditions A or B and C. The track record, it’s a very poor instrument to use to judge the quality of that manager. 

So as the institutional investors do, many just spend time with the manager to understand the manager, understand the team, the process, and whether there’s evidence in the past (not just that three-year track record), evidence in the past of the ability of that manager to demonstrate skill. Does he bring the necessary skills to deliver what he claims he can deliver, and has he demonstrated in the past that he’s done that effectively? 


 Sure. I think that makes sense. Just out of curiosity, I’m sure you’ve probably traded the program in some shape or form while it’s developing it and certainly you obviously have a lot of experience in the approach or the process that you are applying, but because the markets have been so different since we had the financial crisis, is your program or is the process behaving as you expect? Are you delivering the return risk profile that you would expect? 


 Yes, within what we expect, but it is different than it has been. So the risk premier that are being harvested by managers in the managed futures/CTA space can be harvested in a variety of ways. They can be harvested… let’s be clear here, we’re not trading some transient effect - some end of the day market closed timing abnormality. We’re not using some Mis-pricing algorithm that is going to be arbitraged away. If there is a large macro factor which is driving prices, our involvement in that market, our removing of the premier by taking a position doesn’t cause that risk premium to go away. 

What has changed though, as I look back over… when I look back to New York, when we were starting to offer some of our proprietary trading strategies out to the high net-worth space, and we raised ¾ of a billion dollars in about nine months selling packaged, early stage quantitative systematic products, when I was in New York. If I look back, then the techniques that worked then still work now. But the risk adjusted return expectation today is far, far lower. 

I think that’s a reflection of a few things, but primarily the space has become much more crowded. It is much easier for relatively quantitative numerate, individual even, to buy a book, learn about the market and implement some relatively basic techniques. With enough people doing that on scale, then that starts to become a crowded market and where market conditions change and everyone runs for the door, then the prices adjust accordingly and that delivers a poor risk outcome. So the techniques that need to be used today to harvest those premium need to be more sophisticated, need to focus more on controlling risk and perhaps maybe a second reconsideration than a number of years ago. 

So what has happened, if I look back over fifteen years? The risk premier have been constantly there, but the method used to extract that is grown up. I’ll give you a hard example. In currencies one can borrow in Yen at zero and convert that into Australian dollars and lend that at 10% 15 years ago. Put that in the bottom drawer and come back in a year and you’ve made 10% interest rate differential and something on the currency move as well. It may have been slightly hairy a couple of days of the year when the currency might have moved against you. In the long run that cross currency carry trade was very profitable. 

Those days… that is a short gamma trade. It’s a trade that earns a small risk premier every day and that has shocks every so often. The frequency of those shocks and the size of those shocks has increased over time as people are running to the door to liquidate – passive carry trades like that. Whether they be passive carry trades in the Australian dollar versus the Yen, or whether they be passive carry trades in curve rolled hours and interest rates, or commodity trades, passive trades relatively easy to access trades, simply trend following trades, have become a much less satisfactory technique then they once were. 


 Given all of that and given the fact that you obviously observe clearly the changes in structure of markets and these different risk premier, what would you say just from a sort of top level point of view, what would you say along your journey has been the main discoveries in dealing with these changes that you’ve been able then to apply in your strategy, if I can put it that way? 


The early days of moving from derivative risk management into the systematic or quantitative investment management space were, I think, difficult for me. I came from a world of… there were maths, there were equations, this is how these markets behave, these are how these derivative instruments can be priced and it’s very complicated. There’s lots of degrees of freedom of risk in a portfolio, but we sort of understood how to risk manage complex derivative portfolios. 

Then when I started in 2000 looking at what people were doing in the managed futures space, and I was in the U.S. at the time, certainly many people saw the world as a signal, a trade, I go long, I go short, maybe I’m flat. I have a hundred dollars, I have a hundred markets, one dollar in each market, that’s my diversification. So it wasn’t a continuous space where prices move continuously. I have a discrete signal – long or short. Risk management was I put a dollar in each of my baskets. 

That didn’t sit comfortably with my experience, my view of the world. I think it was when we started to trade cash equities at HSBC, when we started to trade equities at our portfolios, we had some 6,000 stocks we would trade and we built long/short portfolios. When we were forecasting future moves in these equity markets, it was short term reversion moves and we were looking to build a portfolio of long equities against short equities, maximize the reversion effect in that portfolio, but wait and choose the constituents which equities we held to try and reduce the market risk. The market risk we didn’t want beater, but not just the first driver of returns in the market, but the first end risk factors, the first end factors that were driving equity market returns. We insulated ourselves from those by the way we constructed this portfolio. So we left behind the risk we wanted, the expected returns that we’d forecast and we removed all the sources of risk. 

It was that technique, that technology we’ve pulled across to the world of managed futures. So we don’t generate signals, we generate forecasts. Asset A goes up 10 basis points, asset B goes down 5. With those forecasts we can build a portfolio which means we have a portfolio which has the risks we want, the holdings that give us the risk we want, and we remove the risks we don’t. 

So instead of having 100 markets, 100 dollars exposed and thinking we’re diversified but actually we’re not because currencies - all the risk is the dollar; commodities - well it’s energy isn’t it? Then you’ve got U.S. rates and so forth. There are relatively few macro drivers of risk but they appear everywhere and it looks like you might have a reasonably diversified basket of holdings until there’s a shock, correlations go up, it’s not as diversified as much as you thought. 

So to take an active approach to managing risk, building a risk model and controlling that portfolio to have the risks you want, that’s for us when we saw the way to marry managed futures and the continuous world of risk management and the derivative risk management background I come from. There are a lot of challenges in doing that, but that’s sort of the nexus of those technologies and where we started. 


Sure. I want to move on to the heart of the strategy, namely the program itself, but I do want to pick up on something that you just said because I think it’s worth touching upon. I agree with what you put forward, that in the U.S. back 15 years ago, it probably was a relatively simple type of trend following predominantly that people were applying. I think that Europeans, in general, probably took a slightly different approach and made it more “sophisticated” by doing more things. On the other hand, if you were trying to be objective, as I will try to be now, I have to still give credit to the simpler strategies because I still see managers doing something along those lines and being very successful with track records, of 20, 30, 40 years with very few down years. It may be a more volatile ride, I can’t, from memory I can’t judge that, but I don’t think that the first… it’s not the first iteration of trend following that maybe is still applied but some managers have stayed pretty true to their origins of trend following and are doing it with still a lot of success. Is that a fair observation? 


 I think that’s fair. I have nothing, no reason to criticize trend following. It is a very effective strategy. I think, as you say, the trend following factor… if a strong factor exists it’s traded exclusively by some managers, the techniques that they use today will be different from the techniques they used in the past. It’s all in the implementation, isn’t it? As we said before, for us, we trade more than just trend, we… trend is a strategy that captures moves, so if there are no moves it will bleed: enter trade, leave; enter trade, leave; enter trade, leave – it bleeds. 

So as a strategy it’s a difficult strategy in low volatility markets, in markets which have false breaks, in markets which don’t have strong macro-themes. Traditionally it has been a very good strategy combined with equity portfolios which perform badly in those market conditions. I think a more interesting thing is to look at the return dynamics and say, “Well we have this thing that I pay a premium every day for the opportunity to capture a trend and it doesn’t happen and every so often the trend happens and it refunds me for the premium I’ve paid.” 

Well maybe we could identify another strategy where I earn the premium every day and every so often something happens and I give some of that back. Maybe they will complement each other quite well. So a carry type strategy compared with a trend type strategy, maybe that makes sense. That’s the kind of a thought approach for us. It is a different product. The trend product is one product, what we do is different. 


Sure, and that’s what I wanted to gently interrupt you and steer you in that direction because I really would like you to tell me what is the key objective of the program from 30,000 foot down? I think that you’re already heading in that direction. 


 To deliver high quality risk adjusted returns through a diversified method of models of risk management control and drawdown management; but risk adjusted returns. Below that, we as a manager use a variety of techniques to do that but that’s what the investor should expect. 


Sure, sure. Let’s talk about that. Let’s talk about how you have structured the program from top down, and why you designed it the way that you have in order to deliver those high risk adjusted returns. 


So, first of all, our philosophy is that there are drivers of market prices. We can’t observe them. We may have a sense of what they are. We know that some asset prices are more sensitive to certain factors than other assets. The fact is that it’s driving prices that vary through time, so we know that if you want to think of it as such, there’s a carry factor and carry is the “pass if” harvesting of a risk premium through taking of a spread or risk position. 

It collects a premium every day till it doesn’t and gives some back. We know those strategies work well in a low volatility, stable environment with low event risk. Other strategies work better when markets are more volatile and moving, like we talked about – trend. So on that basis we could build a model for carry. We could build a model for trend. We could build a model for value or reversion, perhaps. We could follow these models and use them to forecast an asset. 

So maybe the Euro might be a function of the forecast from carry, a forecast from trend, a forecast from value or reversion. In what ratio do we blend these different types? It depends on the market regime, doesn’t it? That’s something that needs to adapt and change over time. We build this model, but we can’t observe this thing that we’re trying to model. It is very noisy. The forecast coming out of our model doesn’t match up very well with the thing that we’re trying to forecast. It’s very noisy this signal. If you go back to engineering, if you’re trying to deal with noise in the signal, then there are some well understood ways to do that. 

If your system is stable, then maybe you can filter some out because you know everything outside some range is not information, it’s noise and we’ll filter that out. You may be able to use different types of models and combine them together. If you think of a model as perfect forecast plus some noise, that’s model A. Model B has a perfect forecast plus its own noise, and model C has a perfect forecast plus its own noise. When you combine them together, you have three sets of perfect forecasts and three sets of noise which are not correlated. So the noise tends to diversify or cancel out, and you improve your signal to noise ratio. 

Something that we were doing at British Aerospace was we were looking at machine vision for the auto landing of aircraft. You have a camera that sees the runway, which is great. It can see the runway. It can line up. It can land, but if it’s nighttime you need an infrared camera. What happens if there’s some bright… there’s a fire on the runway. Well that’s not going to work. It’s going to disrupt the camera. Maybe you have some radar system as well. So building up a series of different sensors using different techniques, all seeing the same thing – the runway, but with different noise that would interfere in that signal. Combine them all together and you have a composite which is a much higher quality forecast for the runway environmental markets. 

So we have many models and we build many forecasts using different techniques and across the many different risk factors. Then we combine all those together in an ensemble technique – a committee technique to give us the forecast for the asset we’re looking to trade. 

Now, we know market conditions change. We know that over time the techniques need to improve. We also know that over time the economic environment changes - yield curves will at some point they will go up. Volatility rise, it will fall. We have rapid changes - risk on, risk off, if you want to think of it as such. Some models work well in some environments and not in others. So we need to have a dynamic nature there selecting which models we place emphasis on and which ones we don’t. 

One of the great arguments for systematic approach to investment managements is its repeatability. The fact that you can run a backtest. You can talk about what people might expect. It’s not subject to the human implementation of discretionary management has a process it has implemented in a more loose way, perhaps, by a human. Where systematic is implemented, hopefully, in a rigorous way in code. That’s a real strength, but it’s also a massive weakness of the approach if it is not dynamic. 

A systematic manager can see his portfolio. He can see, sometimes it isn’t working. He can see a particular model is causing a problem, but what can he do? Does he pull the lever and turn it off? That’s discretionary intervention. Does he have an investment committee meeting and then pull the lever and turn it off, then it’s discretionary intervention by committee; or does he leave that model in the portfolio for later for when it starts working again, for when it’s the right market conditions and have the whole system recognize that that model is no longer… its marginal contribution is decreased and it starts to turn itself off. That is the learning aspect, the dynamic aspect to what we do. 


Sure, sure. 


 So we’re forecasting markets. 


 Yep, and we’ll get into that maybe in a little bit more detail. I wanted to ask you, how many markets? You clearly believe in diversification in terms of models, that’s for sure, and time frames I’m assuming, but how many markets do you trade and do you cover all sectors: financials, FX, commodities? What does that portfolio look like when it comes to the construction of the portfolio? 


 To a large extent we’re market agnostic in that the fact is that we’re trading a peer across all markets. The recent oil moves have appeared very strikingly in a variety of different currencies, whether that’s Canada and its exposure to Tar Sands, or in the Norwegian currency. So we’re largely agnostic to the assets that we trade… asset markets that we trade. We trade about 60 markets at the present time.  

We started out not trading the commodity markets and that was a decision we took based on the account size when we started. The feedback that we’ve had from investors, subsequently is that it’s unusual to see a commodity-less portfolio. For those investors that want a commodity-less portfolio, they gain a great deal of comfort that we’re not trying to carve commodities out of an existing track record for them. 

Today, 60 markets, and as we grow we would expect to extend that and we would trade a larger range of markets and spreads within those markets as well. 


Sure. You mentioned the type of strategies and model you use and then it is clear that the forecast and the selection and weighting of these different models into the position for each market becomes very important. Talk to me about the forecasts, how long do you try to forecast? I noticed the word you used “ensemble method” I don’t know whether that’s just a fancy word, because I’ve heard someone else use the same word and I’m not sure I fully understand its importance, if it is important, or whether it just means everybody has a vote. So maybe you can educate me and enlighten our audience a little bit on that as well. 


Sure. Ensemble methods in a classical machine learning sense has a certain strict definition. I’m not sure how other people see that. We generate some 5,000 forecasts a day across the markets we trade. 


 Wow! That’s a lot. 


Yeah, it’s a lot, but it’s a technology problem. So we have quite a lot of data. From that data we can look at the history of a model and look at its forecast history and what the realized returns of the market it was trying to forecast, and look at, “Did it do a good job?” 


 Can I just… sorry to interrupt, I just want to make sure I understand. So the 5,000 forecasts, is that because you have 5,000 different model market combinations and that each of them make a forecast? 


 Yes, but to be clear, in an ideal world every model, when forecasting market A, would have error or noise which was zero correlation to the error coming out of model B on that market, coming out of model C, so when you combine them together you have maximum diversification. In practice, they’re not truly independent because the drivers of the noise can appear in many different places. What is that noise? But, essentially, yes. 


 Ok, ok. Well go back. I interrupted you, so please continue with your explanation of the forecast method and how it all evolves in the process. 


 So we have some 5,000 data points a day, and what we’re trying to do is identify how well that model is doing. So if we asked a person what his forecast was for the S&P for the end of the year and he said, 10,000, then we probably suspect his judgement was quite poor. If his forecast would be the same price as we closed on Friday, we might think his judgement was better. So when we ask somebody a question, we make a decision about the quality of the answer. Sometimes that quality of the answer is based on how high quality of hit that individual’s answers have been previously. So really what we want, when we ask a model for a forecast is, “What is your forecast?” and “What level of confidence do I have in the answer you gave me?” 

So we don’t ask models for forecasts, we ask them for both: their answer, their forecast, plus the uncertainty or confidence that they have in that. Knowing that, we know how to combine all these forecasts together to create a single combined forecast for the asset. So let’s think of a practical real world example. 

You are the manager of a proprietary trading desk. You have ten people on the desk. They all trade 10 markets each. You have capital you need to allocate to members of that desk. Your most experience trader trading the most profitable market this year, US equities, gets the biggest allocation, and the least experienced trader, the new graduate on the desk, trading Japan, gets a very difficult market, gets the least allocation. You make a judgement based on the underlying competence of that model and the market on which it is trading, when you decide allocate capital. 

In a similar way we can do a similar thing here. So when we look at the forecast coming out of our model, we look at the performance of the model itself. We also look at the performance of that model applied to that market. So how might the system… what are the failure modes of the system? 

Well we know that carry models which we talked about before, carry models work, and then they don’t, for a while, and then they’ll work again. When they start working, they tend to start working across a wide variety of assets. When they start working, different types of carry model start working across a wide variety of assets. So you’ve got quite a lot of potential information there that you can use. There’s some groupings and structure, some knowledge of how markets behave that you can take into account when trying to learn whether this forecast is going to perform well tomorrow. 

If a carry mode on asset A, on day goes by and you have one more data point. You can’t do a lot with one data point. But if you look across the cross-section of carry models and the cross section of markets on which you apply them, you potentially have hundreds of data points every day that you can use to help learn whether carry models on those assets is going to perform well tomorrow. It’s that approach of understanding the drivers of the performance of the models, the drivers of the market, that is the main knowledge. That knowledge of the investment process using that as a starting point to apply dynamic or machine learning techniques is a key. It’s transformed the problem from being his 5,000 data points and ten years of history go and work that out, to let me give you some strong hints of what is the system underneath that’s driving this data. 

We talked before about the importance of, in finance, understanding what the system is that is driving the data because without that it’s just a very noisy data problem that machine learning techniques don’t work well on. 

So to answer your question, ensemble techniques is a dynamic mechanism for allocating across a committee or a pool of forecast supply to assets to generate a single combined forecast for each of the assets. 


 Sure. So I’m going to try and make it even more simple… 


 Ready to learn more about the world's top traders? Go to TOPTRADERSUNPLUGGED.COM and sign up to receive the full transcripts of the first 10 episodes of the show and visit the show notes where you can find useful links to other amazing resources. Thanks for listening and we'll see you on the next episode of Top Traders Unplugged.