The financial sector is making a massive shift towards machine intelligence in capital markets. Panelists share their experience in using data science and domain expertise in fully understanding data context. They address how machine intelligence in capital markets can be useful in creating new alpha signals, as well as in the data generation/preparation process, in portfolio construction or risk management.
Moderated by: Roland Fejfar, Executive Director, Morgan Stanley.
Panelists: Mark Salmon, Professor, Cambridge University; John "Morgan" Slade, CEO, CloudQuant; and Andrej Rusakov, Co-founding Partner, Data Capital Management, held at the London Big Data and Machine Learning Revolution event in April 2018.
You can also access the full video below.
Machine intelligence in the capital markets over the last two to three years has received a lot of attention. Consequently, a lot of promises have been made. Some are ambitions, some are speculative and some are plainly wrong.
We are still in the very early stages of applying this technology in a capital market context and other sectors have been more advanced. Just applying other techniques from other sectors with a capital market context doesn't just work.
So with this panel, we explore the current state of machine intelligence in capital markets in recent developments, key challenges and direction of travel in applying this technology, in this context, in order to generate differentiated alpha in capital markets.
The Issue of Trust: Machine Intelligence in Capital Markets
Starting from an academic perspective. Where do we start? What do you think are the key challenges in applying machine intelligence to current market problems?
I think machine learning is evolving quite rapidly and I don't think we'll be using the same sort of techniques in the same way in three or five years. If you look at how machine learning has been applied in biomedical sciences I think the applications in finance first of all of this different nature of the animal in finance. But I think we're probably about five years behind what the equivalent applications are doing in biomedical research. The second part of your question I think is more difficult in that I think has been always a problem for quantitative methods in the finance industry. And as techniques evolve.
Personally, I found it very difficult to convince relatively conservative fund managers to take on machine learning techniques and it's always a permanent surprise to me that they don't listen to what I say because I think the techniques are actually very powerful and they have to be used with great care. I think there's a very clear conservatism which I think is justified within the fund management industry and other aspects of finance because you need a track record. And I don't think machine learning has got there yet.
You need to have interpretability, particularly from a sales side so the managers are wary of adopting techniques which they can't themselves explain to the people who are trying to sell their strategies too. And I think we need to establish credibility and I think the last paper was actually very interesting in the sense that it is really showing a value-added over traditional methodologies for adding to a strategy that's been well known for many many years.
Bridging the Gap
When you explain to investors what you do how do you bridge that gap of trust?
So what we do is the phenomenal black box. It is very important to explain to investors what exactly we do and to Mark's point, into predictability is very important. We view machine learning as a generalized regression problem whereby yes it does value, but even the most sophisticated machine learning techniques, in my mind at least, could be explained by traditional factors such as value, growth, trend etc. Now machine learning does add value whereby one can use it to enhance portfolio allocation decisions and timing the entry and the exits of a trade is better. However, predictability is something that we show how investors. I personally have a philosophy of machine plus human or human plus machine.
I think that a human armed with machine intelligence will always outperform machines alone or humans alone. And if I were to draw a parallel, computers have been playing chess much better than humans for the last 20 years or so. However, a very good chess player, not a grandmaster, armed with a great chess playing machine will outperform the best player machine up until today and this is certainly the case in finance, for now, I think it's likely to stay. And I guess the last point on predictability and how we explain the gap between what investors understand and what is actually doable is everything. Almost every alpha that you finds is in essence bait to factor your investor doesn't ask you about. And once you're comfortable with that, inter-probability should become much easier and this is really what we're doing.
Morgan from your perspective what is required to bridge that gap?
You know the established players have a process they follow. I worked for Melbourne Richfield 20 years ago which is a large CTA that had been around for 30 years at that point in time. They were entirely systematic. They had a way of doing things, their investors expected them to continue to do that, if they had started to talk about using this back then it was a regression trees, classification trees, they probably would have scared a lot of people.
There is a concern about upsetting the apple cart if you have something working or an established player you have a large AUM. Why do you want to introduce some new variable to the equation? Well I'll tell you why, I think that you bridge the gap by explaining to people that instead of trial and error processes or trying to build more and more complex regression models as Andrej alludes to, that kind of concept, you know machine learning gives you a boost in efficiency on the research side that you didn't have before. We can do things 100 times faster than we could before.
But there is a problem. It's very easy to move to the dark side. It's very easy to curve fit. And so you have to have an immense amount of discipline and essentially recipes within your organization that are established to ensure that people stay on the track and don't go off the rails like curve fitting. So I think convincing people that you have those safeguards in place is a key part of adopting this.