This is an edited version of a presentation which Scott delivered at the LBMA/LPPM conference in Rome on 30 September. He explains how systematic traders like AHL use mathematical models in order to try to predict the future path of prices, and why AHL uses algorithmic execution to increase speed and efficiency, with an ultimate objective of increasing investor returns by driving down trading costs.


Just by way of introduction, AHL/Man Systematic Strategies is the systematic trading arm of Man. We manage in the order of US$13.9 billion across a number of strategies and approximately a quarter of that is in physical commodities and commodity futures, with the vast majority being futures.

People often confuse ‘systematic’ with ‘high frequency’, so it’s worth clarifying: some of what we do might be called higher frequency, but I definitely would not group our strategies under the broader ‘High Frequency Trading’ label. It is important to distinguish between the really fast folks, who deal in fractions of seconds, and the ones like us, who tend to deal in time slices ranging from hours to months, as opposed to fractions of seconds, because our respective trading behaviours and potential market impact may be very different.

I am going to talk about two things today. One is just a brief introduction to what systematic traders do, because there seems to be an air of unintended mystery and more than a little confusion about what systematic investment shops actually do. I will then spend most of our time talking about execution, and algorithmic execution in particular, because depending who you ask, algorithms are either the answer to all our prayers, or they are the devil incarnate and spell the end of markets as we know and love them.

What Do Systematic Traders Do?

An Anecdotal History of Systematic Trading

To understand what systematic traders do, it’s useful to talk by way of an example and, to my way of thinking, the first properly systematic trader was actually a card counter by the name of Ed Thorp. To understand what systematic traders do, it’s useful to talk by way of an example and, to my way of thinking, the first properly systematic trader was actually a card counter by the name of Ed Thorp.

The history goes something like this: in 1954, IBM introduces the 704, the world’s first commercially available machine incorporating indexing and floating point arithmetic. In 1956, a PhD student in mathematics named Ed Thorp has a realisation that is now very well understood, but which was, at that time, nothing short of revolutionary: a deck of cards has memory. For example, suppose I have a deck of cards with four aces, the odds of getting an ace on the first card is 4/52. But if I draw an ace on the first card, then the odds of getting an ace on the next card is 3/51, and that is a different number from 4/52.This means that by observing the history of the cards that have been drawn, I know something about the likelihood of the future path of cards that are yet to be drawn. Thorp combines this insight with the computational power of the 704 and estimates a bunch of rules – a systematic approach – for placing bets in Black Jack, where the bet size is a function of the conditional distribution of the deck. Interestingly, and perhaps relevant for systematic trading as an investment style, he goes to Las Vegas and doubles his money in the first weekend. Four years later, he goes on to publish his system under the titillating title of Beat the Dealer and changes the rules of Black Jack at every casino in the world virtually overnight. Card counters, and Thorp personally, are universally banished from casinos, so he applies his insights to stock markets. He reports, according to non-audited Wikipedia data, a 20%-plus compound annual return for the next 28 years, so apparently he was on to something good.

“To understand what systematic traders do, it’s useful to talk by way of an example and, to my way of thinking, the first properly systematic trader was actually a card counter by the name of Ed Thorp.”

Estimating the Composition of ‘the Deck’

What does that mean for systematic trading? We are trying to predict the future path of prices based on their history. That is essentially the same problem Thorp needed to solve, with one key difference: we never know with certainty what the true composition of the deck– or rather the market – actually is. So, we start by observing market data instead of cards, and that could mean prices, fundamentals or pretty much any data source you can think of that I can extract, store and process with a computer. We then filter that data according to a bunch of rules – here is the systematic approach – and build a forecast of what we believe the next period’s return on a given contract may be. Precisely what this system does is not terribly important for this talk, but it is a live trading system that we are currently testing, because we think it has the potential to identify short- term fluctuations in prices. The system is called FASTCOMM, because it trades commodities and does so relatively quickly by our standards: it consumes data on energy markets on a time scale of minutes, and conditional upon observing certain patterns, it makes a forecast about what might happen.

’’So why are we so interested in algorithmic execution? I’ll show you some data on why ‘algos’ help in a minute, but for context, the primary objective of execution in the context of systematic investing is to drive down slippage.”

The blue bars on this graph are the strength of the signal, so a plus two is two standard deviations positive, or what we would view as essentially max positive; a minus two is the converse. The green bars are what we observe over the next 500 minutes, so about the equivalent of a trading day. What the pattern, from the lower left to the upper right, of the green bars tells us is that stronger signals are more likely to be associated with more positive returns, and weaker signals are more likely to be associated with more negative returns. This signal is then transformed into a desired position in lots, a ticket is generated, and a trade is executed. Ultimately, this process is what systematic trading is all about.

Why Do We Use Algorithmic Execution?

The Role of Automated Execution

Talking about FASTCOMM provides a natural segue into our other main topic, algorithmic execution. If I have a system that consumes a large amount of data and that wants to place a large number of trades relatively quickly, how is it going to execute those trades? It would be impractical at best for a human to trade even a single system like this for a handful of markets and yet we might (and in fact do) build dozens or even hundreds of such systems, which trade across hundreds of markets. The answer is, humans don’t execute for FASTCOMM, machines do. In fact, the entire process, from observing the data to entering an order into the market, is done without any human involvement.

So why are we so interested in algorithmic execution? I’ll show you some data on why ‘algos’ help in a minute, but for context, the primary objective of execution in the context of systematic investing is to drive down slippage. For every dollar that we lower our execution costs, we raise the net return to our investors, and making money for our investors is what this business is ultimately about.

So, that’s the end game, who is going to win – the scary guy with the machine gun or the excitable humans in the pits? It’s not quite that simple, but it’s close.

The Rise of the Machines

However, you slice the numbers, there is simply no getting away from the fact that the machines are taking over: looking at equity markets, we reckon that between 30% and 60% of all stocks are traded by algos, and the figure is likely to rise further. If you look for commodity futures, we estimate that the algo share – so not just electronic platforms, but the share of executions done by algos – has doubled from about 20% to about 40% in the last two years alone (according to internal AHL estimates). We believe there is a reasonable chance you could see that share increase by a similar amount again over the next couple of years. At the same time, high-touch execution, as in you and I dealing via instant message or voice, seems to be converging rapidly to zero for liquid futures.

The AHL Approach

How do we use the machines in practice? We have a fascinating system, whereby we create explicit competition between humans and the machines. We have two trading desks, one staffed by humans and one run by algos. Very small orders all go to the algos. And what we consider large orders all go to the humans. But for everything else in between, in what we call the matched band, we randomly allocate them between the algos and the humans, we track their P&L, and they compete for who can achieve the smallest implementation cost. When we measure execution P&L, or cost, we mean the difference between the price on the screen at the instant at which the order is received and the price at which the order is executed. The difference between those is slippage and that is the measuring stick for the race that we’ve put our human and machine traders in.

AHL’s Evidence on Competitive Execution

In 27 of the 32 commodity markets that I looked at for this analysis, the machines beat the humans. That’s about 80% of the time. In three out of four precious metals markets, the machines beat the humans, and in the fourth one, palladium, it is basically a wash. That’s a relatively easy call to make (based on internal AHL estimates).

Alogorithmic execution - Who is going to win? Machines or humans?

Electronic Competition in the Dealer Market

We have a pretty deeply held belief that we want to create proper incentives and competition wherever we can, so we apply this basic principle in other ways as well. For example, in the same way that we pit our humans against our machines, we also use our machines to pit dealers against dealers in order to create what we view as a more efficient marketplace.

Our internal order book for spot gold consists of about half a dozen dealers who stream prices into our system. I pulled a random screen shot from 21 August and found that if I were unlucky, and just happened to look for a bid and got the lowest bid, or I looked for an offer and got the highest offer, I would see something like a 60-cent bid/offer spread. If I called someone for a voice quote, I might get a 40-cent bid/offer spread. However, what I see via the machine is an inside spread of the best bid to the best offer of only 24 cents, which is the natural outcome of creating competition and using a machine to compile the resulting information. While the precise numbers vary, we see remarkable consistency in this general pattern.

Competition between Marketplaces

I have talked a lot about how good the machines are, but to be fair to the flesh and blood traders on our desk and amongst all of you, there are certain things that the machines really are not good at. For instance, as I mentioned earlier, orders that we classify as ‘large’ are generally directed to the physical (as opposed to virtual) trading desk. This is particularly important for relatively illiquid markets, where we have observed that slippage tends to increase more rapidly with size and where the rate of increase tends to be worse for the machines than it does when the order is actively managed by a person. Some of our data on palladium would appear to support this general observation and demonstrate one of the ways in which both our human traders and our risk-taking counterparties have an opportunity to add value. When I’m only going to trade one, two, three lots at a clip, I can pretty much do that on the touch, at the bid or offer on the screen. But when sizes get significantly larger, that is when our traders will reach out to their counterparties and ask for a ‘risk’ price, where the counterparty offers us a firm price for a given size and assumes the execution risk from that point on. It turns out, in the case of palladium, that the machines are cheaper for little tickets, but beyond a certain size, our data says we incur less slippage by calling a dealer and getting for a risk price. It is crucial to understand that whilst the machines have a role, there are certain times when their ability to understand what is going on inevitably runs into challenges, and in times like that, we believe that old school, manual trade execution is still the best route.

The Final Result

So what does this all mean in the end, in terms of our ultimate goal of looking to obtain higher returns for our investors? If you look at the last 20 years of AHL’s real-life trading, our estimated execution costs have come down by a somewhat astounding 80%. And if you look at just the latter portion of the chart, when we started imposing machine-human competition and increased use of algos, we think we’ve cut our slippage costs by about something of the order of a third, which is a fairly remarkable number and a positive result for our clients.

Scott Kerson, Head of Commodities for AHL/Man Systematic Strategies

Prior to joining Man, Scott ran a boutique consultancy focused on quantitative analysis and systematic trading strategies for the global commodity markets. Before launching his own business, he spent 15 years in commodities sales and trading, including primary responsibility for Barclays Global Investors’ active commodities model; proprietary trader and senior quantitative analyst at commodity-specialist hedge funds Ospraie and Amaranth, and Managing Director and Head of Commodities Sales & Structuring for Deutsche Bank and Merrill Lynch.

He graduated with Highest Honours in Economics from the University of California (Santa Cruz) and received his Masters in Financial Economics, concentrating on econometrics and macroeconomics, from Duke University, USA.