In a recent report on data analytics, Hedgeweek explored how sophisticated technology tools are being deployed by fund administrators like SEI to enhance the investor experience as well as within the front-office of fund management groups to improve portfolio and risk management.
In that report, Tobias True, Partner and member of the Portfolio Construction Committee at Adams Street Partners, a leading private equity group, explained that as the world continues to evolve around the use of data and the ability to map out different relationships, it has led to an evolution of the process in terms of how Adams Street uses data to gather insights and come up with hypotheses.
As True remarked in that report: “It has helped us enhance our ability to manage risk in our portfolios and to construct different portfolios. Data analytics has been a critical supplement to our bottom-up research by helping us construct top-down portfolios.
“If we have a portfolio with too much exposure to an industry or source of risk, we can’t just go out and rebalance it tomorrow like we could in the public markets. We therefore use data to think one or two steps ahead to help us (optimise) the way we allocate capital.”
Such is the volume of data at hand, it is allowing investment teams – hedge or PE – to use data analytic tools and machine learning applications to better understand the sources of risk and return, and derive fresh insights into what they think will drive performance.
In this respect, machine learning technology, which has really risen to prominence in the last five years, could help risk teams operate in a more integrated fashion with portfolio managers, and in effect push risk management further into the front-office.
“In the old world, risk management was done as a separate exercise by an independent department stressing portfolios and so on,” says Andrew Downes, COO of New York-based Atreaus Capital LP, a discretionary global macro hedge fund.
“With greater and more sophisticated data collection and faster, more powerful computing power and more real-time analytics, that segregation is less pronounced. It means if you are looking at risk analytics you can apply them to influence the investment process, making risk management less of a backward-looking exercise and more part of the portfolio investment programme.”
To illustrate how much artificial intelligence and machine learning is being used within the hedge fund community, a recent poll by industry data firm BarclayHedge found that 56 per cent of respondents confirmed they were using these tools in their investment process. More specifically, 33 per cent of respondents said they were using machine learning for risk management activities.
BlackRock has developed a unique operating system for investment managers called the Aladdin Risk Platform. The platform uses machine learning algorithms to provide users with risk analytics to monitor risk in their portfolios and can, according to BlackRock, automatically screen over 2,000 risk factors per day.
Massive computing and analytic power, whatever the machine learning application, have the ability to offer significant benefits to risk managers because the more the machines learn from the data sets, over time, the better they get at pattern recognition. Risk is becoming even more active than before because people have the tools at hand to run more detailed scenario models, at the pre-trade level, to see what affect a particular trade or set of trades might have on the VaR.
Downes believes that if one wants to make risk management implicit in the investment process it is necessary to perform risk analysis at the time a trade is being placed.
“A simple version of this is to have real-time pre-trade checking versus risk parameters: the system can see if the trade about to be put on will keep the portfolio within the risk rules or exceed them, in which case it would therefore modify the trade going through or generate an alert.
“A more sophisticated version would be to enter trades or exit trades based on market conditions, probabilities of loss and size of loss. If you are holistically looking to improve performance by reducing losses, then your analysis tool would solve for that. However, the difficulty for machine learning tools – actually machines in general – is to identify the regime change or paradigm change in the markets, which is something that humans are particularly good at,” says Downes.
We are still very much skimming the edge of what could be possible in years to come, as machine intelligence grows exponentially and the machines themselves reach a singularity point: a tipping point, at which time machine intelligence exceeds the limits of human intelligence. Quite how we will think about monitoring risk in 10 years’ time is anyone’s guess. Will a form of augmented intelligence in the front-office control risk management on our behalf because it knows better?
For now, fund managers are well placed to use the analytic power of the machines to support their decision making.
A report by Chartis estimates the annual spend by hedge funds on risk, analytics and trading technology to be approximately USD9 billion, with a particular focus of that technology spend on modelling, portfolio management and analytics, and risk data aggregation and reporting.
George Kaye is the founder and CEO of Derivitec, a risk analytics software vendor. The firm’s central philosophy, according to Kaye, is to make it as easy as possible for people in the financial industry to provide validated risk management reporting: “That includes everyone from small hedge funds to global sell-side institutions,” he says.
In his view there have been “seminal changes” in technology over the last five years which have radically transformed the way that people do business. “From our point of view at Derivitec, the most important change has been the cloud. The cloud is now so prevalent that it almost ceases to be a topic of conversation. There is so much happening with cloud technology. For example, the speed at which we can compute, the volume at which we can compute, the security of how we compute.
“That means things like large-scale risk calculations now happen in timescales that would have been unthinkable a few years ago,” comments Kaye.
Risk analysis creates virtuous cycles
Better, faster, more powerful processing capabilities are the engine behind machine learning tools. Machines have become a good decision-making tool for portfolio managers to road test strategies, validate them and stress test them but it is still early days to expect the machines to provide predictive diagnostics on market risks. They are still in the second innings of self-supervised learning.
Risk analytics can offer tremendous value and insight for real-time risk management, but the human still has to exercise caution that false flags are not being created; after all, every algorithm is susceptible to bugs and glitches in code.
Kaye says that the greater processing power is helping firms like Derivitec to deliver even more granular, detailed risk reporting: “Absolutely. It’s all about granularity. We can run VaR calculations back 10 years on portfolios of thousands of trades with varying degrees of complexity on the trades. It doesn’t matter if the client is a small fund trading 20 instruments or a large fund trading 10,000 instruments, we have that elastic capability built in to our system. If we want to slice and dice the risk and look at it in all different ways, that means running lots of different risk reports with a high degree of position-level detail. And that is absolutely possible. The technology allows us to do this.”
Speed is another element that is helping fund managers think about risk and fine-tune their portfolios. As risk systems improve, fund managers can understand their risk exposures and respond more adroitly; in the past, this may have been limited, in turn limiting how much capital they could safely deploy. Now, risk technology advances are presenting multifarious risk metrics, improving how portfolio risk is managed. A virtuous cycle ensues, as the more they analyse risk – with or without the help of machine learning tools – the more confidently they can deploy capital, improve the Sharpe ratio, and attract new investors in to the fund.
Ivan Popovic is Managing Partner at Tolomeo Capital, a Swiss-based quantitative fund manager that relies heavily on technology-driven investment strategies. Rather than buy best-of-breed risk systems, Popovic explains that the team built its own proprietary state-of-the-art risk measurement framework.
“The approach is fully generic and can incorporate any kind of asset class: from plain-vanilla exchange traded products to hedge funds, private equity and complex OTC derivatives. At the heart of the system lies the pricing kernel, which statistically analyses the return drivers of an asset.
“At the first stage, each position in the portfolio is modelled by itself. As a next step, the dependence structure of those return drivers are modelled, which results in Monte Carlo simulations of the P&Ls of each position whose co-movements behave as in the real-world. The advantage of this approach is that it allows for straightforward analysis of certain types of shocks which affect multiple assets classes,” explains Popovic.
Signalling risk
Additionally, risk factors which are currently driving the market can easily be identified in the system. These models are constructed bottom up and are not exposed to data-mining or over-fitting risks.
“It is our last line of defence to make sure we are aware of what kind of risks we are currently facing,” adds Popovic, confirming that AI and machine learning have been an integral part of Tolomeo’s trading system since day one.
“As the timing of an investment is not completely random, the risk in a trade is therefore dependent on the underlying signal. As the sizing depends on the signal it is therefore derived from machine learning and AI techniques. This is the first line of defence. We do monitor risk and performance in real-time but at that level, no AI/machine learning techniques are used, they are found at the heart of the trading systems.”
Over at Atreaus Capital, being a discretionary macro manager there is a clear and obvious human element to managing the portfolio process. They are not looking to employ sophisticated machine learning tools just to improve risk analytics, per se, but more to analyse cost of trading and performance. Downes is in no way dismissive of machines. He sees value in using them alongside the human, but in a very controlled way.
“Human plus machine beats machine only. That’s what I think is an interesting space to explore as a discretionary portfolio manager,” he says.
TCA and trade optimisation
To that end, one area that Atreaus Capital is using sophisticated analytics is in relation to trade execution; specifically the cost of entering and exiting trades.
“If you look at FX liquidity it is spread over multiple venues, multi-bank platforms and ECNs, etc, so what we’ve done is build an execution system through which everyone trades. That gives us all of the trading data, with which we can perform TCA analysis to see how each trader compares in terms of the cost of each trade executed, to try and optimise the process.
“We have the ability to measure real-time price movements and how difficult it is to get certain trades done in different circumstances with different banks. Using algorithms to optimise TCA addresses, in some part, concerns we might have about liquidity risk through the course of day or during different markets.”
Another possible application of data analytics is to think about how to optimise the timing of entering and exiting trades. Collating such data, regardless of whether it is a machine or a human doing the trade execution, gives investment managers the ability to convert that data into information, from which they might derive insights and arrive at new knowledge.
When someone is making money, how can they make more money? Equally, when someone is losing money, how can one use analytical insights to get them to reduce the amount being lost?
“You want to analyse what they do and make it better,” says Downes. “Reducing costs plays a part in both increasing returns and reducing losses; both of which are exercises in risk management.”
IBM Project Debater
Technologists are making huge strides in the application of AI and machine learning to help businesses of all shapes and sizes improve their operating models, and streamline workflow processes. IBM has leveraged Watson – the incredible supercomputer that uses deep learning algorithms to analyse data sets to generate insights – in all areas of industry. Its AI capabilities are being applied in healthcare diagnostics, trade finance, logistics, education and advertising.
This is cutting edge technology. Watson can work with the smallest data sets to learn and help businesses to advance their thinking. Soren Mortensen is Director, Financial Markets, IBM. He confirms that IBM has a series of API’s as part of their Watson Services that are geared towards cognition and problem solving (Thinking, Learning, Deciding & Acting) whilst others are geared towards human interaction (Perceiving and Responding).
“Within financial markets these API’s are used for the discovery of data and extracting any relationships that might come out of that data. An example of this is how quarterly weather improves revenue estimates in certain industries.
“The cognitive capabilities can identify, by analysing large volumes of both structured and unstructured data, underlying drivers affecting performance to offer better insight into companies of interest, which will enable the buy-side to make better investment decisions,” says Mortensen.
IBM also have debater technology where the system can listen to arguments, and respond convincingly with its own, unscripted reasoning to persuade an analyst to consider his position on a controversial topic. Based on vast, empirical sets of data, it can also help relationship managers create hypothesis to explore with their clients along with arguments to support these.
The technology can also help the investment manager to visualise the effect of news events on an individual company through natural language processing, graph database and visualisation technology.
“We are also using this technology, to assess how a price move on crude oil will likely impact all other asset classes and risk factors – and if your portfolio is exposed to those risk factors, as a consequence of the move, determine what the impact might be,” explains Mortensen.
The upshot to this is that the buy-side can leverage market data and cognitive capabilities to really understand underlying risk factors that impact the performance of a given portfolio.
“Investment managers want better insights into the companies that make up their portfolios to make better investment decisions – this is possible with cognitive technologies, which can analyse large volumes of structured and unstructured data to find patterns. It is augmented intelligence as opposed to artificial intelligence; augmenting information for a portfolio manager to make better decisions,” adds Mortensen.
Model complexity risks
The risks or challenges of very advanced nonlinear machine learning techniques (e.g. neural networks, deep learning, etc) are that they can be seen as black boxes. This means that it is sometimes not clear how the input data leads to certain output data, which hampers the interpretation of the results and hence the decision making in risk management.
“This additional complexity adds another dimension of risk at the core of the risk management process, which can be also seen as a disadvantage,” opines Popovic.
“The addition of model risk does not necessarily improve the overall result or capabilities of the system. Sometimes, increasing complexity actually hampers the overall results of any (risk) estimation and adds a lot of variance to the estimation.
“A slight change in the input data might completely change the output. Hence using very advanced, relatively new and complex AI tools in risk management might add other sources of risk rather than offer an improvement to the whole risk management process.
“They certainly need to be used carefully as increasing complexity does not automatically translate into better results.”
Python scripting for risk reports
Regardless of the extent to which hedge fund managers might use machine learning tools to improve risk management and measurement, what is irrefutable is that risk has become much more of an immediate requirement.
In the past it was more of a regulatory box ticking exercise. Now, managers want to know in real time what their risk exposures are to respond to their clients. Are they keeping everything within the right limits? If the markets are shocked in different ways and given the complexity of the instruments being traded, what does risk mean to me as a portfolio manager?
“Whereas managers might have been producing risk reports on a weekly basis they are now producing them several times a day to get close to real-time monitoring and management of risk,” observes Kaye.
Derivitec have also pushed their reporting capabilities to support Python scripting so that clients can interface with their analytics using Python APIs to generate their own reports.
“It’s very important to support this because if you can’t present risk in an easy to understand, visual format, people will just ignore it. It means that clients are now better able to manage risk, with the whole organisation better informed of where the predominant risks are, whilst at the same time able to drill down to the last detail within a clear, graphical interface,” comments Kaye.
Certainly with such massive volumes of data at their disposal, it is incumbent upon portfolio managers and risk officers to present portfolio risk analytics in a clean, uncluttered way. Otherwise, another risk is potentially introduced: paralysis by analysis.
Machine learning tools can help cut through the clutter and present salient points, either numerically or graphically, in order for investment teams to make split second decisions. This is the real power of the machine – crunching through zettabytes of data to seek out patterns in the numbers.
Back at Atreaus Capital, Andrew Downes cautions against getting too carried away by offering these final thoughts in conclusion:
“Some managers have spent a lot of money hiring Silicon Valley engineers and I’m not sure they’ve been as successful as they set out to be in using AI in their investment processes. Having an unsupervised machine simply be fed data and be expected to learn a way to trade and make money is still a long way off in my opinion. Even if one managed to achieve that Holy Grail, how would you know that it could be repeatable in all markets? It’s hard to know what the deep learning algorithms would be doing and why.
“At the end of the day, to be successful you need to start with ML and other tools by keeping things simple, intuitive and try to improve step by step what one is doing; such as optimising TCA. Start with a simple objective and other things can follow.”
