Institutional investors turning to AI, data science to improve processes, yield alpha
BY Kelsey Rolfe | October 24, 2019
What if institutional investors could turn their portfolio companies’ quarterly earnings calls into actionable data by analyzing their sentiment over time?
The Alberta Investment Management Corp. is attempting to do just that, in partnership with Edmonton-based machine learning company AltaML Inc. Using a form of artificial intelligence called natural language processing, AltaML is analyzing the text and tone of companies’ quarterly and annual results calls, and using that information to predict the movement of stock prices.
Cory Janssen, AltaML’s co-founder and chief executive officer, says the project could provide the AIMCo’s investment managers with a significant edge. “Sentiment analysis is one area being done by a few groups out there, but right now it’s still in its infancy and we think there are probably factors we can develop for different signals.
“Think of, what’s the sentiment overall? But [also], what is the [chief financial officer] thinking? What about analysts — is that positive or negative? And . . . how does that change over time? As you link [that] together with other economic factors, or even other companies in the supply chain, we think there’s actionable signals we can generate now because we have the ability to pull data on a much greater scale than analysts could do on their own.”
The partnership has already begun to bear fruit beyond sentiment analysis. In its first pilot project, AltaML automated the AIMCo’s capital call and distribution process — cutting it down from between 15 and 20 minutes to less than five seconds — and eliminated the need for employee involvement.
Previously, when one of the pension organization’s real estate asset managers put out a capital call or returned funds to the investment manager, it would go to a central email box checked by an operational group employee. The employee would then have to manually enter it into an internal system to submit or receive the funds. Using machine learning and natural language processing, AltaML automated the process to allow an algorithm to read the emails and send the funds.
“It was really freeing up that time of work of little value,” says Michael Baker, senior vice-president of investment operations at the AIMCo. “Most of the people doing this would have an accounting designation, so we have other things that they can do besides doing this type of thing.”
Moving forward, the AIMCo is expanding the technology to other liquid asset classes with similar distributions and capital calls, including infrastructure, private equity and timber. The partnership is also looking to pilot other small projects, split evenly between operational work that will automate and simplify back-end processes and those that have alpha-generating potential in the front office.
“We held a bunch of workshops with different groups across AIMCo and came up with potentially over 100 different use cases,” says Baker, adding the partnership has six or seven pilot projects on the go.
The AIMCo is just one of many institutional investors increasingly recognizing the value of using artificial intelligence applications and data science and analytics in areas like asset allocation, trading, research and risk modelling. In many cases, these technologies are being used to automate cumbersome parts of the investment process and free up analysts and investment managers for more complex work.
Big bets on big data
Vast amounts of data exist behind the fancy algorithms deployed by institutional investors, says Bradley Hough, a senior consultant at PBI Actuarial Consultants Ltd. “We’re seeing a lot more interest in the data science area, in institutional investors saying, ‘Well, can we get better data and can that give us better insights into things?’” he says. “A lot of institutional investors [are realizing] that data is extremely valuable.”
He sees investors benefiting from increasingly sophisticated software that can help them analyze and display data in a clearer way for better decision-making. “What’s happened in the last few years is much more effective software that allows [PBI] to take data and present it live . . . so we can tell a story with that data.”
Jordan Vinarub, head of T. Rowe Price Group Inc.’s technology development centre in New York City, says the firm is using machine learning techniques to segment its clients, including plan sponsors, into groups that behave similarly.
From there, it can “understand what they were likely to buy based on what they held, and what the next best product [might be], like an Amazon recommender model,” adds Jane Conway, the firm’s head of enablement and global distribution.
“[With plan sponsors], they have a mandate for the plan, they have schemas and they have asset allocations that they have committed to their investment committee and to the board they will abide by. In that case, you need to understand the role of each of those people at the plan sponsor and how to be relevant with your messaging to them.”
The firm’s technology development centre has also expanded the use of data science and machine learning to other areas of its business, says Vinarub, including assisting its equities research team to use segmentation and classification in its work, collecting more data and assessing the predictability of its data sets.
Managing the portfolio
One area where some investors have started to see the benefits of machine learning is adjusting asset allocation to respond to macroeconomic factors and changes in the market, says James Davis, chief investment officer of the OPSEU Pension Trust. However, for pension plans, this is a bit more complicated.
“Timeframe matters a whole lot. We’re a pension fund, we’re investing for the long term, so we’re thinking about earning the returns we need for our members over the long term at the lowest risk possible. Dynamically adjusting the portfolio to take into account changes in the market’s perception of what’s happening in the macroeconomy can help us in that regard, but many of the folks who have had success in this have had success by making those adjustments in very short timeframes.”
The shorter the timeframe, such as daily or intra-daily, the more data an investor has to help them make decisions and to train machine learning models. Investors looking to make changes on a weekly, monthly or yearly basis don’t have that level of information, says Brandon Da Silva, the OPTrust’s associate portfolio manager. In order to address this, the pension fund is factoring uncertainty into its machine learning model, which it has been working on for about two years.
“What we found is that it’s not as simple as just saying historical volatility is your uncertainty measure,” he says. “Our uncertainty predictions are actually uncorrelated with historical volatility.”
Instead, the OPTrust’s model is teaching itself to read market proxies for movements in the economy. “There could be a lot of volatility in the market, but if your model actually knows how to interpret that volatility on a forward-looking basis, then there might not be a lot of uncertainty going forward,” says Da Silva. “If a model is confident in a trade, it will have less uncertainty. The confidence is not necessarily correlated with volatility.”
The model helps the OPTrust make decisions, along with other inputs such as external managers and macroeconomic analyses, but humans still make asset allocation tweaks. “The models Brandon’s been building are providing me with more confidence to act with more certainty, and [I’m] able to act in a more timely way,” says Davis.
Unigestion’s quantitative equities team is looking at using machine learning to forecast single-stock returns and risk measures, such as market beta, says Salman Baig, the firm’s senior vice-president and portfolio manager. He notes machine learning has been used more commonly in equities because the space has one of the “longest and largest data sets” available to train algorithms.
“If you look at the entire equity universe, there are tens of thousands of equities. With, let’s say, fixed income, you have a lot of bonds traded, but not nearly as much as you do equities, and also a lot of the data in the bond space is . . . not nearly as harmonized.”
However, he adds, it’s entirely possible to create algorithms that work for fixed income, commodities, currencies or other asset classes. Indeed, the firm is looking at using machine learning algorithms for private equity fund selection.
Hough expects institutional investors will continue to invest heavily in artificial intelligence and data science applications, especially in the face of a low return environment. “Interest rates seem not to be rising as quickly as everyone expected,” he says. “If those rates are lower for longer, you have to look harder to find returns. I think [institutional investors] will have to invest more in modelling tools and techniques . . . to see how they can get additional returns from their portfolios or managers, and how their risks pan out over time.”
That data will become more sought-after and more proprietary, predicts Baig. “Ownership of the data becomes pretty important because [if] the data is out there, other investors start looking at it and you start to see your returns deteriorate. I know there are data providers in exclusive deals with some of the largest investment firms.”
And investors will be required to do more due diligence on where their data is coming from, he adds, as areas like Canada, the European Union and the United States increasingly scrutinize technology companies for how they’re using the data they collect from users.
“A lot of data . . . are sourced from Facebook, from Google, from LinkedIn, and I think I could see, in five years, that [users] don’t automatically opt into Facebook distributing data with their partners, [they] have to voluntarily choose to do that,” says Baig. “A lot of policy-makers and regulators have come to realize data is not a commodity that can just be shared with everybody, that you should have control of your own data. And so I think that’s a potential challenge for some of these data providers and people who rely on it.”
As for future applications, institutional investors are continuing to look into their options. The OPTrust has worked with external providers that use artificial intelligence to better understand and predict the impact of climate change from a physical perspective on the plan’s assets, says Davis. “Getting a better handle on the physical risks of climate change is a big thing for us.”
Citing the example of an Australian wind farm owned by the OPTrust, Davis says artificial intelligence could help to estimate how climate change will affect wind patterns in that area, which could be a longer-term physical risk to the asset.
But even as the use of these technologies continues to expand in the investing space, they won’t replace humans, says Baker. “Frankly, if we are not in this space and starting to learn how to use these tools, we’ll be left behind. We feel strongly that the use of artificial intelligence and machine learning can help investment professionals, risk professionals and also help operations staff to make better decisions or to improve efficiencies.
“Robots are not going to be replacing anyone, but machine learning really is going to help people make better, more informed decisions.”
This article originally appeared on CIR’s companion site, Benefitscanada.com. Read the full story here.