The Gateway to Algorithmic and Automated Trading

Me and My Machine: The Automated Trader Interview

Published in Automated Trader Magazine Issue 12 Q1 2009

Launched in November 2007, the Ion Fund is “a 100% systematic fund designed to operate in highly liquid and listed instruments” [major stocks, stock indices, government bonds and currencies]. It came through 2008, the year of vanishing liquidity and corporate bail-outs, with a net 2008 return of 24% (fund objective: an uncorrelated 20%pa). How does a fund like that provide a return like that in a year like that? The Ion Fund’s manager Dennis Lohfert spoke to Automated Trader’s Editor, William Essex, about achieving outperformance in interesting times.

Dennis Lohfert

To start with performance, what did you like most about trading conditions in 2008?

I didn't like the trading conditions in 2008. Were they good for us? Yes. Did our systems perform well? Yes. Did I enjoy it? Not really.

Did you see the trading conditions of 2008 coming?

I'm not in the business of making predictions. I'm a trader and those two are not the same thing. There's far too much noise in making trading decisions to be able to honestly call it a "prediction" without offending people that actually make proper predictions, like, say, meteorologists. It's a long-run statistical expectation coupled with strong risk management to deal with the surprises that crop up.
Did I foresee that it was going to be a good year for us? No. But I trusted the numbers enough to have a fair amount of confidence.

If we had spoken immediately before the fund's launch, I would have asked you, "Why now?" So now I'm going to ask you, "Why then?"

Why did we launch when we did? I'm of the conviction that in trading, whether systematic or discretionary, one eventually needs to pull the trigger. I would have loved to spend another year refining everything. But at some stage, you have to do it, or you'll never get anywhere. So, long story short: I think November 2007 was as good a time to launch as any other month.

“I try to build systems that work well over very long periods of time.”

"I try to build systems that work well over very long periods of time."

The fund uses trend following, mean reversion and opportunistic trading strategies. What worked in 2008? Did last year provide an external evolutionary pressure on the fund to adapt, evolve, improve?

Performance was pretty consistent across all groups of strategies. And what's more, not a single system that we have in production actually needed removing from the production set. That was very encouraging and indeed, 2008 was certainly the acid test for our systems. They held up even better than I expected, which is encouraging. It shows that the emphasis we've put on finding long-term system robustness is promising. But there's still a lot of work ahead of us.

The fund is 100% systematic. It's a rule-based and rational quantitative approach to investing. Does it like surprises?

Who doesn't like surprises? Yes, it likes surprises. In economic terms a surprise is better known by the less euphemistic term: "shock". So, in slightly more technical jargon: Yes, it likes shocks.

Have you ever disagreed with your model over a trade? If not, have you ever come close? And were you proved right?

Almost every day, there are some trades that I disagree with. And I'm happy to admit that I'm wrong on average. And it's very important that this is so. If I could consistently beat my own creation, it would mean that I clearly haven't put enough work and thought into improving it. After all, only once the student can eclipse the master, does the master know that he has taught him everything he knows. And that's why the system trades, and not I.

Do you use, or are you considering the use of, algorithms to finesse your order execution? If so, why? If not, would you consider using them? If you do, do you use third-party providers?

We do, but I want to emphasize that that this is not where we focus our research. That area has gotten very efficient over the last five years. Even though people always try to portray this field as something they are doing something really elegant with, in my opinion there just isn't that much you can do anymore. You're talking about very marginal improvements so your core research efforts are best focused elsewhere. Remember: algorithmic trading is not a direct source of Alpha, it's merely a cost-reduction technique.

Dennis Lohfert

You produced some volatility of your own from April through to August 2008. What happened there?

Actually, that is mostly due to sampling. What you're referring to is seen as blips in the monthly returns, whereas if you look at the daily return series you'd probably agree that it was all in line with historical expectations.
There wasn't anything that unusual about those periods, from a market-environment or trading point of view. If you recall, May and June weren't actually that remarkable, market-wise.

Other than that, does the fund's performance reflect the consistent application of a trading strategy, or constant change in response to constantly changing market conditions?

If I may be so bold and put it like this: consistent application of robust adaptation. I know that sounds grand, but that's what I try to focus on. I try to build systems that work well over very long periods of time. I'm not a fan of switching systems in and out and deploying systems that have not been rigorously (and I mean rigorously) examined and tested. We won't even look at anything that hasn't been tested over at least a decade of time and sales and order book data, and we really want to go back more than that still. Preferably longer than 15 years or until the inception of a market. Some markets are much younger than this, so there we tend not to be active. There are some techniques to come to grips with this, most notably data proxying, but we prefer mature markets.

If anything, what have you learned from your first 'live' year that all the testing didn't, or couldn't, teach you?

Yes, it seems that the NYSE and LSE can take on very emerging market-like characteristics, particularly on the regulatory front.

Looking ahead now, as well as back, how does the fund react to unprecedented market shocks? Can there be such a thing as too much volatility, or even perhaps the wrong kind of volatility?

If I may answer the second part first. To go back to time series analysis: volatility is really just one characteristic of a time series and it's not enough to be able say "oh it's very volatile". There's more to it than that. Simply put, yes, there's volatility that tends to help a system and there is a wrong kind of volatility.
If I may get a little philosophical here: Traders live off the exploitation of price differences. There isn't more to it than that. If prices don't move, you can't make any money. As such, a little volatility is always required.

Our systems have in the past reacted reasonably well to shocks and the accompanying volatility, but we are certainly not claiming to be immune. Shocks are by definition something rare, or else it wouldn't be too shocking, would it? So, we have to be very careful to make generalizations about shocks, but on average, our systems seem to like them, as I've said earlier.

How does your risk management process work?

I divide risk into four components, of decreasing impact. The most important risk is adverse movement. This may not seem that profound, but the fact is, every time you have a position open, there's risk that you get hit. And there's a very small probability that you get hit hard. And this is actually where most of your risk comes from, on the higher frequency side of things. This drives the main risk overlay. It's pretty standard stuff and everybody, especially on the high-frequency side knows what numbers to use. There really isn't much modelling you can do here because you just have to be so conservative and pretty much all assumptions have to go out of the window and one is left with a higher frequency version of Expected Shortfall. More interesting, especially from a longer term perspective is the risk posed by drawdown (that is, time spent underwater). This comes in second after adverse movement. Now this is an area where you can really busy yourself with modelling and simulation and we've done a fair amount of work to get a handle on drawdown behaviour and how to get a clearer picture of what the drawdown risk really looks like. The next component is Performance Degradation. While this is very important in the long run, from a business perspective, it's not that significant a source of risk to investors because what it means is that performance starts flattening out over time, and not making any money is a lot better than actually losing money. We've got systems in place to watch for this, but it happens on a much longer time-frame than Drawdown does and is about an order of magnitude smaller in terms of numbers. The last risk component is liquidity.

Dennis Lohfert

What about liquidity, which has been fragmenting across Europe for most of the lifetime of the fund? Given that ongoing change in the trading environment in one of your two major geographical markets, how has the fund evolved to cope with it, and how is it evolving?

As we are primarily market takers we're obviously interested in the trading community concentrating as much liquidity as possible in one central book for each market. I find it incredible that everyone and their mother in Europe wants to have their own little pseudo-exchange. This happened in the US in the late '90s and I'm sure it'll consolidate itself to a large degree. It's good to see competition on execution, exchange and clearing fees, but market fragmentation is inefficient for price discovery and I believe that efficiency, over the long run, will win out.

Tell me more about the Ion Interconnected Computational Cluster (IonICC). What is it, and how does it work?

It's basically a cluster (or grid) computer with our own framework for doing distributed processing. It takes our Trading System implementations and completes a three-stage process, which is at the core of our research pipeline, in parallel. This three-stage process is there to combat curve-fitting, which is the number one problem in Trading System Engineering.
It's very computationally intensive, but fortunately, it can be computed in a distributed manner. IonICC allows us to complete that part in a reasonable time period. Our research process is very long and it would be much longer still if we had this amount of computational power available.

Please describe the gestation process of IonICC.

Let me start with this statement: Silicon Time is cheaper than Carbon Time. In other words, a computer's time is much cheaper than that of a human being. Trading System testing is what is termed an embarrassingly parallel problem. As such, the natural choice for it is cluster or grid computing. More nodes mean simply this: more results, in a shorter period of time. And that in turn means better productivity for our biological systems.

Ultimately it boils down to this: how can you effectively accelerate the main workload in the Trading System Development pipeline? Which, when you look at it from the perspective of time spent in each stage, number crunching is still the largest slice of the pie. So I thought: OK, take that piece, parallelize the workload and reduce the total time spent in that stage. I had done a reasonable amount of HPC (High-Performance Computing) before so rolling out IonICC wasn't that much of a problem. IonICC is just a tool in the end. Think of it as the analogue of what a chainsaw is to a lumberjack. It allows us to cover a lot of area quickly and broadly.

Your system development process emphasises realism as well as consistency and robustness. What do you mean by realism?

Realism is the key to developing viable trading systems. Without it, unless you're developing very long term systems (where you can conveniently "assume away" a number of issues), any results generated are inherently suspect.

There are so many factors to take into account when simulating a system historically: transaction costs, liquidity and market impact, latency variability and data errors, funding availability and pricing, repo and securities lending, availability and pricing, order book queuing and execution probabilities, all these things have to be accounted for before you can really say that the results you are looking at are realistic. One has to look at a great many things in great detail.

Your system development and evaluation process uses historic trading data. Will the economic crisis alter that process by (for example) distorting the definition of "normal" market behaviour?

What's a "normal" market? I don't think I've seen anything normal in the last 18 months. I think most people would agree that this is not a normal market and most people I talk to wouldn't redefine a "normal" market because of the last 18 months. What it does do, however, is give us another challenging regime for systems that are developed in the future that they must navigate successfully in testing. It just makes the problem a little harder still. A successful trading system survives a large number of different regimes.

Dennis Lohfert

How does 2009 look so far? Do you see any new themes emerging, or is this year, so far, looking like a seamless continuation of last year?

Again, I'm not in the business of making predictions. I do think there are some economic reasons for why things can't continue to be quite as volatile as they have been in the last six months, but then again, economics is not nicknamed the "dismal science" for nothing.

The one thing that has become obvious to my mind is that the "hidden risks" (risks that are very difficult if not impossible to capture using quantitative analysis) have increased so dramatically that a very conservative approach to risk management will continue to be at the forefront throughout 2009.

You stress the fund's non-correlation to any given market's performance. Does that become a kind of discipline in itself? Is it monitored on a regular basis?

I was never a believer in benchmarks, so I do not want to benchmark our performance. I think the majority of hedge fund managers seem to have forgotten that we really are supposed to be total return orientated. I come from the proprietary trading side, as do the majority of people at Ion. As such, benchmarking is just not in our blood. It's relatively easy for us to de-correlate from any given market, given the variety of instruments we trade and how often positions change their sign. We do monitor correlation, but it's not the primary objective. The primary objective is risk-adjusted return.

Do you see Ion's performance during such extreme market conditions as a vindication of systematic trading as a concept? After all, many discretionary managers were losing their shirts.

Systematic Trading, and the type of Systematic Trading we do has advantages. But it also has disadvantages. Systematic Trading works, but that has been well known for over two decades. I'm not sure that it needed any vindication. Certainly not in the circles that I move in.
Describe your development process. Is it a collaborative effort or do your developers work on a concept from beginning to end in isolation?
It's mostly a collaborative effort. Four eyes see more than two and six eyes see more than four. There are so many pitfalls and details to examine, it's very important that the Trading System Engineers do as much cross-checking as possible.

When you get a trade idea, what happens then?

We have a very long research pipeline. It's a long journey from idea to production. The idea is not to get something out of the door as quickly as you can. The idea is to get something checked out as thoroughly as you can. An extended time-to-market is not going to matter in the long run, but it may help you catch a lot of stuff that isn't going to work as well as you thought it was going to work. Development takes as long as it takes.

So how long does it take from that "light bulb" moment of inspiration to get a model into production? What could you do to speed that up?

It takes a long time. Months and months. Some people pride themselves that they can turn out new strategies in 30 days and deploy them in production. From my own experience, that is far too short a time to conduct all the necessary checks and double checks. If someone tells me they've finished a system in less than 30 days I can give them a list of two dozen things to do with it that will occupy them for another three months. I pride myself on the fact that at Ion it takes up to 180 days to get a system properly checked out and deployed, and that's despite all the cluster computing and automation tools we have. Good things do take time.

When was your last light bulb moment?

There are times when you look at the results and think: there's something there. Sometimes, when you set out, you know that's going to be there and it's just a question of doing the work and getting to it. Other times, it's just a very faint notion and you have no idea whether it's going to work. You definitely have those light bulb moments, but they're rare. My last big one was nine months ago.

So you spend 180 days working on an idea, developing a model. Then you plug it in. What happens then?

Then we're in harm's way, right? This is where the actual risk starts. Before, there was the research
risk and the business risk, but now the investment risk starts. Once it goes into production, it starts trading, and now the main issue is making sure that it doesn't fall apart, and pulling the plug on it as quickly as you can if you think something's going wrong. It's a continuous monitoring process, looking at the statistics, the general behaviour, the current market conditions, doing comparisons between current and previous market conditions, basically doing continuous homework looking for any signs of degradation or unusual behaviour.

A lot has been written recently about the ever decreasing half life of trading models. Has that been your experience too?

Of course it has. Market dynamics do change over time. Sometimes slowly, sometimes very quickly. But not every system is susceptible to an erosion of performance. I think a lot of people don't understand this point. Simply because an idea is systematic, that doesn't mean that it'll disappear with increasing efficiency. There are many trading systems that rely on very strong economic foundations to derive their edge. I'll name liquidity provision and risk transfer as two structural themes here. There are very solid reasons why these types of strategies are profitable in the long term, so there is no decay in that sense. There's competition, but that's true for anything in a free market.
Non-structural systems, essentially what you'd call perhaps technical (or fundamental) trading systems, are of course, subject to the curse of ever-increasing market efficiency. This is why our research process is so geared towards robustness, realism and consistency.

Dennis Lohfert

How many models do you currently have in production?

225, split across 32 groups.

Have you set any targets for the business in terms of the number of models you'll have in production in twelve months time?

No. I think that would be a foolish thing to do. Trading System Engineering, as I refer to it, is a long and difficult process. And as every engineer will tell you, you can't rush a good, solid product. Good things take time. It takes however long it takes and you don't want to take any shortcuts.

How much of the architecture that's required to run Ion models is built in house?

All of it. To this day, I fail to see any commercial product suite for Trading System Engineering out there in the market place that would get me to raise an eyebrow, let alone two. A big part of the edge in designing systems is how you set up the whole process front to back. And to do that effectively, you usually end up with a custom solution.

With the infrastructure you created yourselves, why did you choose to build rather than buy when vendors have invested millions in R&D to produce cutting edge trading tools for firms such as yours?

I suppose with that line of reasoning you would also come to the conclusion that General Motors must be building really good cars. The whole point about systematic trading is doing things (slightly) different from the rest of the herd.

Finally, how do you handle recruitment? Where do you find people?

Finding good people is one of the most difficult aspects of this business. We have a team now where I'm happy with everyone, and I think the key is to recruit on opportunity, not necessarily out of need. Take your time. I find it difficult to foresee a situation where we really need to recruit someone quickly. We're small, and I just don't see that happening. But now, we're doing well as a company, we're growing slowly, and in the current economic climate, that puts us in a good position to keep our eyes open for anybody out there who might be interesting.