William Essex: Tell us a bit about your background. You're widely described as a 'former academic'. What took you from there to here?
Sonia Schulenburg: I joined an investment bank after I graduated with a BEng in Computer Engineering because I wanted to learn more about fund management. My interest was to model the behaviour of financial traders and develop a fully systematic investment management process, but as the problem is fairly complex, I quickly realised I needed to learn more about artificial intelligence. I felt the answer was there, but did not know exactly how to get to it.
So I left the Bank, completed a PhD in AI, became a scientist, and then went back to fund management.
I chose Edinburgh University for my PhD because it is one of the world's leading artificial intelligence research centres, and Edinburgh is also very strong in fund management, so I thought it was the perfect place to bridge the gap between these two quite distinct areas.
William Essex: Could you talk about performance to date? Also, what upgrades, enhancements, plain old changes (if any) have you made to the overall process since the fund began live trading?
Sonia Schulenburg: Since the initial production phase in 2007, the MAYA technology has undergone several significant stages of development. We have gone through these stages in parallel with stages of beta testing, paper trading and live trading.
It was like building a 300 bedroom hotel but having to operate it commercially while it was being built. We had to 'rent the rooms' that were available (2 or 3 at a time) while the rest of the infrastructure - like pools, restaurants and more rooms - were being developed. And we also had to raise finance to build the hotel at the same time! This, as you can appreciate, makes the challenge so much harder. I cannot begin to describe how hard conditions have been for us.
However, managing change is key and I am proud of how we have performed considering the difficulties we had to overcome. Now, let me leave the analogy and go back to the fund.
Live trading with a managed account for Sir Peter Burt, our long-term backer and investor, began in June 2008. This account had a limited amount of funds and therefore we could only trade one or two of our strategies, but we kept trading the fund. Then in August 2009 we started a second live trading phase with a larger pension fund account from another of our long-term backers, my husband Mark Calvert.
Then with seed capital from Baillie Gifford, the MAYA Fund officially went live in mid-April 2010. Despite significant financial limitations during this period (June 2008 to April 2011), which did not allow us to launch with the fully developed infrastructure that we have today, the fund has produced highly competitive real time results that are in line with back tested performance, as well as exhibiting low volatility and low correlation with the market.
The full technical infrastructure has been in place and the fund has been fully invested since January 2011. Since then we have been proud to be included in the Top 10 Long/Short Equity funds for our category in global hedge fund data bases.
This year, which is a difficult one not only for quant and CTA funds but also for long-only funds, our returns continue to be positive, competitive and uncorrelated with other hedge fund styles. Comparative performance of the MAYA Fund, the market and another 11 major hedge fund styles and indices, including quantitative directional, long/short equity, multi-strategy and managed futures amongst others, shows the MAYA Fund is ranked 2nd in these 11 categories from January 2011 to end of June 2011, which I believe is very encouraging for an emerging fund manager.
William Essex: How are your returns audited and verified?
Sonia Schulenberg: By Deloitte, with Bank of America Merrill Lynch as custodian and prime broker; we have been careful to ensure the fund is administered to the highest standards; we are, after all, building a long-term, sustainable business. In addition, the fund offers daily pricing and an independent administrator calculates and sends the daily NAV directly to the investors.
William Essex: Please explain the methodology. You have a decision-making process based on "cutting-edge advances in Artificial Intelligence". What are those advances, and how do they give you an edge? Can learning classifier systems represent competent traders?
Sonia Schulenburg: The system uses thousands of predictive features (or building blocks) to model stock price behaviour which is found to be persistent in different market regimes and time horizons. The most consistent, robust and historically informative features are automatically selected and used as inputs to MAYA's financial market models. The financial features are chosen to be fully independent and robust across multiple time scales and fluctuations in the data.
Features classified as being the most informative of future changes in the market are selected automatically during the portfolio construction phase to be used as a predictive input of the models. The market models are adaptive in real-time. This allows for changes in the market environment and gives them the flexibility to be adapted into other markets.
An array of machine learning methods are used to quantify the performance of various stocks, producing a set of return predictions for multiple model-stock pairs, and recommended investment actions. The stochastic nature of the market is addressed formally using a range of machine learning methods which coexist in a grid. The grid of models outputs actionable signals associated with a confidence level of their prediction. The models learn features using historical stock price data at the tick level.
And what is significantly different from other machine learning methodologies is that the data is flushed through the models only once. Models are continuously updated as more data becomes available (in a truly on-line fashion), maintaining model relevance to the current market climate. Furthermore, changes in the market regime can be detected by analysing the investment trends and market distributions, allowing more relevant models to be activated or current models to be updated appropriately.
An investment portfolio is constructed using the signals produced by the grid of predictive models. At this stage a second level of learning takes place, where model predictions are combined to meet specific strategy constraints such as risk and execution style, and to ensure diversification of exposures across the portfolio. This stage analyses the effect that the time of day has on the distribution of all the individual stock prices which can be exploited to optimise the investment strategy.
We believe that Level E's inter-disciplinary approach encapsulates some of the most recent and promising developments in the fields of machine learning, distributed computing, data analysis, computational economics and behavioural finance, which combined are particularly suited to the highly dynamic nature and intrinsic uncertainty of financial markets. The end result is an absolute return equity portfolio aiming to deliver consistent returns, lower volatility and low correlation with the underlying markets and other alternative trading strategies.
William Essex: What is a 'potentially predictive market feature'? Please give me a recent example.
Sonia Schulenburg: Market features are complex mathematical expressions that describe price behaviour which is found to be persistent in different market regimes and time horizons.
A market feature is a signature, building-block or motif found in raw price tick data, or in transformations of the data.
William Essex: What is 'continuous learning' and - if there is a meaningful answer to this question - how do you monitor and quantify lessons learned? This may be a subjective human question, but can useful lessons be drawn from all market conditions, including the big (apparently) one-off events?
Sonia Schulenburg: Continuous learning basically means finding systematic ways for an algorithm to test its process while being engaged in that process, and then to alter its behaviour based on the results of its own performance.
In order to adapt to different market regimes, in my PhD work I introduced the term 'continuous learning' in a multi-agent setting of artificial financial traders. In this view, continuous learning rejects the traditional view that learning takes place in two independent stages; the first one using a training set of a fixed sliding window, and then validating and testing the learned behaviour in a subsequent window of fixed size, where learning no longer takes place. The general assumption is that the training set is large enough to contain a representative sample of possible events, which can, after being learned, be applied successfully in the future. Typically this is how the learning process of systems such as neural networks is structured. Level E does not use neural networks in any of its predictive technologies.
The Level E approach overcomes these limitations by focusing on modelling aggregate behaviours - that is, learning to infer near optimal decisions in an on-line fashion. In this framework, lessons can be learned from one-off events as well as from more frequent ones. The human body is excellent at dealing with this type of problem, and this has motivated research in Artificial Immune Systems which is a type of learning classifier system.
William Essex: Is there any role for human intelligence? Please describe your team, and members' roles within it. How do you approach recruitment?
Sonia Schulenburg: Of course there is a role for human intelligence! People often think that because we use machines we don't have to do any work but it is not the case; work is very intense.
At Level E we have a strong team of top PhDs with backgrounds in areas such as computer science, engineering, distributed systems, physics and AI, and together we design, develop and maintain the trading engine and its many components. The technical infrastructure is constantly developing. At first we started trading companies from the FTSE 100 Index, but now the system is being applied to other markets, such as the US and Brazil's equity markets, and in the pipeline we have Canada and other asset classes such as FX.
From its conception, Level E has taken on researchers of an exceptionally high standard. I do the hiring by finding the appropriate candidates first and then approaching them. The requirement is that the PhDs do not have any investment knowledge whatsoever, they must have in-built engineering skills (in their genes), be fast learners, and have degrees in computer science or related and AI at the PhD and Post-doc level. Then they experience a very steep learning curve.
Our research team is based in Edinburgh.
William Essex: Please talk about capital-raising. You have attracted investment from a number of financial institutions, and prominent individuals, and if I understand the situation correctly - do I? - you also have research finance for your work in AI. How do you expect your investor base to develop?
Sonia Schulenburg: I launched Level E Limited as a research company to develop artificial intelligence-based models for trading, and through the process I have received a number of competitive research grants from Scottish Executive, Scottish Enterprise and The Royal Society of Edinburgh, which are publicly sponsored bodies that support highly innovative research and new business start-ups.
Level E Ltd does not have research funding, however, in the sense of a research institution. Level E Ltd is a commercial company. Level E Capital emerged a few years after Level E Ltd, as a way to create an investment product that would encapsulate all that Level E has to offer.
Our investor base is currently 100% institutional, and we are in talks with family offices and high net worth individuals who have expressed interest to invest in the fund. We expect the investor base to be well mixed, but perhaps with more concentration from large institutional investors.
William Essex: Please describe your development and testing process. What technology do you use in programming and execution?
Sonia Schulenburg: Ideas begin with basic and theoretical research, then we move into an applied research stage of experimental development and prototypes (alpha, beta and pre-production) prior to full productisation. This iterative process involves heavy software development and testing.
We fully support open source architectures and Linux-based systems. Windows is totally banned, except for a couple of cases where it is required for data feeds.
We favour Java but are not limited to a particular programming language.
We have developed all our trading infrastructure in-house and execution is via DMA using our FIX engine.
William Essex: How do you manage risk?
Sonia Schulenberg: Risk management is taken extremely seriously in all steps of the investment process. It is embedded in the decision-making process, in portfolio construction and in execution. Our proprietary risk systems ensure that every order has a stop order attached to it. Risk parameters are kept very tight both at the position level as well as the portfolio level.
William Essex: The AI model emphasises steady returns with low volatility. Your best and worst months seem to have been achieved by, as it were, maintaining a steady course and thereby missing out on an upswing and a downswing respectively. Is that fair, and if so, how do you feel about volatility today? Is your fund at home in these market conditions?
Sonia Schulenburg: The media refers to extremely low volatility but this was exacerbated by the fact that the fund was under-invested due to technical infrastructure developments during its first 9 months (from Apr-Dec 2010). This caused a volatility below 5, but under normal circumstances (being fully invested) the volatility of the fund is around 10 when the market might be around 14.
The fund has been fully invested since Jan 2011 and during this period, as well as historically and backtesting, the fund exhibits about 2/3 of the volatility of the market and this reflects in higher returns too during rising markets.
William Essex: Why 'Level E'? Is there significance in the name, and in 'MAYA'?
Sonia Schulenburg: I wanted to honour the Mayans, a Mesoamerican civilisation recognised for their mathematical and astronomical systems, their fully developed written language, art and ingenious architecture.
Level E comes from the floor where I was based during my PhD at the Artificial Intelligence Department of Edinburgh University. At Level E it was us, the Evolutionary Computation Group, and at Level F, the Maths Reasoning Group and often we were very competitive. Of course we always solved the hardest problems! But they were good at theorem proving.
William Essex: When you think about the future, do you think more about developments in AI, or about developments in markets?
Sonia Schulenburg: I need to think about both, and how they are interlinked.
William Essex: Do you live to work, or work to live?
Sonia Schulenburg: My husband says that I live to work, he said 'definitely'.
William Essex: Could MAYA beat you at chess?
Sonia Schulenburg: Chess has been a good environment for testing Artificial Intelligence techniques.
It is different to financial markets in the sense that in chess the information specific to the game is known to the players and the solution space is very large, but known. Early AI had successes when computing power increased and systems were capable of beating the best human players, as was the case with Deep Blue.
However, in a way this has been an AI disappointment because they tend to have a fairly simple static analysis routine. The power comes from number-crunching through thousands of positions, but they do not plan, or learn on the fly from experience like MAYA does.
William Essex: Sonia Schulenburg, thank you very much.