The State of AI in Investment Management
On this panel, industry experts (listed above) discussed the affects of AI on Investment Management. We’ve included a short transcription of the panel, beginning at 4:18 of the webinar.
Patrick Dugnolle, BNP Paribas: I’d like to start with a story about AI before computers even existed.
It’s a story which happened in 1904 when the New York Times published an article about Clever Hans. Clever Hans was an intelligent horse who had been going around Europe for many years. You know horses are known to be extremely intelligent, but Clever Hans is different. He displayed human intelligence. Clever Hans knew colors, numbers, advanced type method. He could remember and spell the names of people he met. However, it was all biased.
He had learned the quickest way to receive his favorite reward was to find the correct times squares he could read minimal changes in the eyes of the questioner. So what does it tell us?
What does this animal intelligence tell us about artificial intelligence? Like Hans, AI carefully reads all minimum variations in the input data. And when human intelligence is mimicked better and better, everyone gets mesmerized. However, there is no bronchi where the intelligence is coming from. The only sure thing is that it is highly dependent on input data and functions.
Just like clever Hans, there is nothing magic about AI. So with that said, I’ll ask the first topic of this webinar and I’m gonna ask you, Victor or Nico, what you think about the challenge that has occurred to integrate financial data into AI models. Is the financial industry effectively using big data and AI. Victor, do you want to start with that?
Victor Martinez, State Street Corp: To address that particular question, I think we can take it from two different angles: on the part that AI has allowed us to differentiate in data is between structured data and unstructured data.
In terms of structured data, I don’t think the finance industry has changed too much about how they actually use data. Especially in the financial data set, with market data. So I think we’re still using the same sort of systems and I think we are, in some ways, stuck to that kind of framework that we had in the past. Now, in terms of unstructured data, that’s where AI has given us a huge, different overlook on how data can be incorporated into the different operations of finance.
On that angle, AI has provided a big leap in terms of extracting information from different sources and being able to apply AI to financial models for different reasons. Now, the effect on that, for example, into trading models or other types of things like that, has been limited. However, in operations and in other aspects of finance, that unstructured processing of data has given us big advances in many different aspects of finance. Also, there’s a big use of AI and big data in issues like portfolio management. Not necessarily enough for generation, but in optimization of how to construct portfolios or how to do effective trading algorithms. So I think that’s how I will describe it.
Patrick Dugnolle, BNP Paribas: Nico, would you like to have something?
Nico Smuts, Ninety One: Yeah, I think on the topic of big data, I was listening briefly to the, to the banking panel, and I think banks are generally ahead of fund managers on this because they’ve been dealing with unstructured data for much longer. The banks have to understand their millions of customers and there isn’t a lot of structured data available. Whereas in fund management, we’ve been kind of spoiled in the sense that we only need to cover thousands or tens of thousands of companies and securities and there is so much structured data out there that that’s kind of what we tend to go for.
I think especially in the current market conditions where more and more of the insights come from unstructured data, that pushes fund managers more towards the AI route where you need to deal with larger data sets and need more sophisticated algorithms, high performance computing clusters, cloud based solutions. I think that we were playing catch up with other parts of the financial industry, but definitely the momentum is picking up speed.
Patrick Dugnolle, BNP Paribas: Do you think, Nico, that is why AI has been used more effectively in risk management perhaps or operations?
Nico Smuts, Ninety One: Yeah, that’s a really good question. I think as data scientists and people working on AI, we’re all looking for the Holy grail, which is that algorithm or that formula that will give you an edge and allow you to see around the corner to predict what’s coming next. But I think there’s a trade off, right? Because this approach of investing a lot of time in alpha generation and using more complex algorithms is difficult. The risk is that you can get it right, but you can also get it wrong. And even if you do get it right, those signals fade into decay over time.
So if you take into account the sustainability for each you create, often the numbers come out in favor of incremental gains in other areas. So, an example, trying to predict time series is extremely difficult, and that’s essentially what a lot of the predictive algorithms strive to do in financial markets.
It’s just a very difficult problem to solve. If you look at something like optimizing trading or finding liquidity or quantifying and automating some peripheral part of the investment process… often that operational game, you can lock it in. It’s not going to get decayed away as quickly and the gains go straight to the bottom line.
So as an example, and this is something we’re doing – we are using machine learning to identify liquidity when we’re trading in opaque markets. If you can find an algorithm that can just save you one basis point in implementation shortfall per trade, and each trade has two legs and they tell you to turn over your portfolio twice, that’s four basis points you’ve saved, which has a high probability of being repeatable. Whereas using it for alpha generation, it’s a bit less repeatable. It’s a bit more chunky in terms of, there are periods when it works but then there are also periods when it doesn’t work. At Ninety-One, we try and spread across those two domains, work on offer generation, but also lock in those incremental gains on the operational and implementation side.
Victor Martinez, State Street Corp: To that point about why AI has made major inroads in areas of investments, specifically in training. I think that there is also an underlying issue. Not just the complication of the problem at hand, though it’s a very challenging problem, I agree. But on the other hand is the nature of markets.
For example, people who probably implemented AI techniques to extract information from unstructured data, they have been doing it for a while, and then there were opportunities at that time where this actually broadly made a difference in strategies.
Nowadays the markets are adapted to that and all of these techniques and all of these procedures or methodologies are widely known and incorporated and the data is widely available to so many people. So I think the market has already taken some of that edge away. I think the nature of the market is being adaptive and is part of the reason why this is so challenging, which brings us to the other part of why alpha generation is difficult.
That is one of the things that AI hasn’t been able to completely crack – that aspect of it is that the nature of the market is being adaptive. This means that they are not stationary by nature and that the stationarity is not so far compatible with a lot of the techniques in AI.
This is, for example, when you try to do image recognition, you are always relying on the fact that your images are going to keep those parts in time and it’s true. A phase doesn’t change that quickly. If you take 10,000 pictures of your face over the last five years, most likely your features are going to be pretty much the same; they’re not going to change. With the markets, it doesn’t happen that way. Every time you take a picture, the pixels on your face are going to be changing completely so you have a different face every time you take a picture.
That’s why it’s so challenging to introduce that type of AI technique into these markets. There is some hope in that direction, and I think the new roles that they have been making in terms of reinforcement learning have some potential to address those problems in a realistic way.
Now, the value of AI techniques is in handling large amounts of data. Finance, to a certain degree, has been limited because we have done very well in terms of one asset, like pricing, for example. We do very well with one, but when we deal with multiple assets and the interaction with multiple assets, the problem becomes a lot more complicated. I think that from that perspective, AI has still a lot of potential because it’s still opening different areas which could be applied to finance. On the other hand, the success in other fields are not easily translated into finance right now because of those issues.