Edgar van Tuyll van Serooskerken, head of Quantitative Strategies at Pictet Wealth Management, explains how Pictet uses machine learning for two purposes: first, to hunt for profitable investment opportunities; and second, to search for patterns in big data that are similar to those of past periods.
For the first use, machine learning can sift through the mass of data we now have on the past prices of assets in various classes, on company performance and on macro-economic developments, in order to look for hidden treasure. It will hunt for relationships between data points that could allow us to make returns and we study them with our experts to see if they make sense – we don’t just trust the machine and trade off them.
Nonetheless, since 2013 we have come up with a whole series of automated strategies, which have beaten the hedge fund index by a wide margin. They tend to reflect human behaviour which doesn’t change very much and makes those strategies stable for long periods. For example, the herding effect which leads investors to want to be in with the crowd – buying stocks that have been rising for a long period, and selling them if they have been falling.
Another example is that people tend to overestimate risks in the short term – for example, of China’s slowdown, the Brexit vote or the Italian referendum in 2016. They overprice volatility in the next few days and underprice it over many months because nobody thinks about such risks in the longer term.
For the second use, we give the huge amount of data we now have to a machine, ask it to look at the past few months and see how close or how distant it is to previous periods. What the machine is telling us now is that the current period looks like the tail-end of a bull market, similar to what happened in the tech bull market which ended in 2000. It was preceded in 1998 by a crash in emerging markets, which we also saw last year. And the share prices of tech companies such as Nvidia which make chips for machine-learning units have seen exponential growth similar to those of fibre optic cable companies in 1999.
The nature of data varies from industry to industry, and this affects how we can use it. Google can use deep neural networks to teach self-driving cars by repetition on the same road many times. But we can’t do that because we have just one history of the S&P500, and if we try to fit rules to a price history they may apply only once or twice. That’s why when machine learning finds relationships, we then search for a debatable explanation of why they work. Experience shows that relationships can break down as soon as you trade on them.
This access to machine learning in fund management has led to huge demand for ‘quants’– quantitative analysts who specialise in the application of mathematical and statistical methods to financial and risk management problems. With the availability of big data and computers that can process it, we hire only mathematicians and physicists with PhDs in university specialisms like number theory, chaos theory and the study of critical phenomena in nature such as the formation of ice crystals.
Other users of machine learning such as Google and Amazon carry out simple operations on huge amounts of data, which requires enormous processing power. But we do a lot of calculations on relatively little data, which requires much less processing power – it is the intensity of the calculations on much less data that is key to our use of machine learning.
One feature of the current enthusiasm for AI is the belief that we are creating super-intelligence (there were similar expectations in previous periods of scientific innovation). But all we are doing is creating more data and using more automation to process it, and machine learning works well only where there is a lot of data and there are rules to process it.
Anything that involves creativity, cooperation and strategy is hard to model because there are no clear rules like in chess or Go. People can choose to break any rules there are – to cheat, fight or take over companies. And human story-telling is still needed to explain these factors to investors, something that chatbots cannot do.
They also want to socialise more by going out with friends to eat and drink – a trend reflected in sales data for premium wines and spirits. A sub-trend here is sporting goods, where there has been extraordinary growth in sales as health-conscious consumers do more running, exercise and activities such as yoga.