Yin Luo is the vice chairman QES (Quantitative Research, Economics, Strategy) at Wolfe Research, and has over 14 years of experience with machine learning in investment management, primarily in stock selection and global macro.
He talks to MarketBrains about the momentum of the last few years, the hurdles for machine learning adoption in investment management, and what the future may bring.
MarketBrains: There’s been a lot of talk about machine learning in investment management, how do you see that unfolding?
Yin Luo: It has changed very dramatically in the past couple of years.
Even five years ago, the investment world’s perception was not positive to machine learning and AI.
Most asset owners and asset managers would think that’s probably a complicated black box, being overfitted and doesn’t work very well in real performance, as opposed to backtested performance.
And even if it does work, I don’t understand this, so I will probably not want to invest in it.
That of course has changed very quickly in the past couple of years. It’s top of the mind among most large asset owners (pensions, endowments, and FoFs).
Fundamental managers want to use big data and machine learning to help them to make better decisions. For quantitative and systematic investors, machine learning is the next frontier to add alpha.
Many managers want some sort of machine learning and AI in the investment process, but it’s still very early stages.
“For quantitative and systematic investors, machine learning is the next frontier to add alpha”
MB: How do you see AI and machine learning being used in investment management?
YL: Most of it is still marketing rather than reality. Many managers may use some of these techniques to do some of the R&D work, but there are few managers really using machine learning to actually manage money.
First of all, it depends on how you define machine learning and AI. It’s very hard to define what is really AI, because the definition is mixed with big data, with machine learning, they all somewhat overlap, and they are all somewhat different.
The usage of alternative data is definitely far better developed. People do use, or try to use, many alternative data in their investment process. Some of these unstructured data are naturally better fitted to use machine learning techniques to analyze.
Analyzing textual data is very natural because you have to use computer algorithms to quantify and give you some sort of intelligence. So, that fits very well with natural language processing or NLP, which you can argue is one form of AI and machine learning.
I would say most managers are still trying to use the patterns that have been identified by these algorithms as a way to help them make investment decisions, rather than investing solely based on machine learning algorithms.
“Few of the firms using ‘true AI’ are well-established firms, they are more likely to be newcomers”
MB: Are there any particular firms you’d point to that are going past that, into using the algorithms for making investment decisions?
YL: There are investment managers, a selected few, that are more advanced, they use machine learning not only in the pattern recognition phase but also build that into their investment decision process.
Every manager claims they use machine learning to some extent, but it’s hard to say who is actually using them, but again, it’s more like anecdotal evidence, well-known firms such as Two Sigma, Millennium, Winton, Man Group.
This is a general market perception, but they never really tell you exactly what they do.
Few of the firms using “true AI” are well-established firms, they are more likely to be newcomers. They typically don’t have great brand names yet, they don’t have a lot of recognition, they have a fairly small AuM. Generally speaking, true AI firms are few in the actual investment business.
To some extent, I see many firms are using it in a data exploration stage, in the data discovery stage, and maybe some form in actual management, so it’s very hard to define what is pure and what is not.
“…the biggest hurdle to applying machine learning techniques is investment philosophy”
MB: What do you think is one of the dominant trends you’ve seen in 2017?
YL: The most interesting development is how quantitative firms think about utilizing alternative data and machine learning.
You would think they are cutting edge adopters, but maybe not. Especially for long only quant firms, they tend to face significant pressures of fees and have a very tight budget.
For the typical quant, you need to have an investment hypothesis, you collect your data, you do your backtesting to check if the data is supporting your hypothesis; this is the very much the conventional research in economics and finance.
The problem is that, in theory, if the data does not support your hypothesis then you should move on.
In reality, what almost everybody does is try the backtesting on different data or using different models, until they find the patterns that do support their hypothesis.
Almost nobody tells you how many iterations they have gone through in the process, and if you try many times almost for sure you will find some data, in some market, over some time periods, and it works exactly as you think.
With AI and machine learning, your hypothesis is not important, it is based on the pattern identified in the data. If you believe your algorithm, identify the patterns, the patterns are likely to survive, then you would invest in it.
“The most ironical thing is that the biggest black box is the human brain”
MB: And that’s foreign for quantitative investment management?
YL: It’s very hard for them to embrace it. Hiring and finding the right talent is very difficult, and yes, the algorithms need to be tweaked to be applied in finance, but the biggest hurdle to applying machine learning techniques is investment philosophy.
Their way of thinking is: well, that’s not how we do things. It’s fine if the fund performs well but if you lose money, how I am going to explain it to the asset owners?
When we (Wolfe Research team) meet with asset owners — the fund of funds, the pension funds, endowments, and so on — they don’t seem to care as much about model transparency.
The most ironical thing is that the biggest black box is the human brain.
In machine learning, you can look at exactly why the pattern was identified. Maybe it’s not intuitive to you, but you know why the machine made the decisions.
“…the best adopters of machine learning may actually come from not the quantitative community, but from the fundamental community”
MB: And what do you think the dominant trend next year will be?
YL: Many asset owners are convinced this is the future, they may not know what to do or how to build such models, but we’re hearing the question: how is the manager who actually manages money truly using machine learning?
We struggle to find such managers, but we definitely see the demand every day. And the best adopters of machine learning may actually come from not the quantitative community, but from the fundamental community.
This interview is edited and condensed.