February 4, 2021

What Artificial Intelligence Can—and Can’t—Do for Investors

While machine learning still has major shortfalls, it is gaining ground at divining useful patterns, like analyzing how executives talk in earnings calls.

Reported by Sarah Min

Art by Pete Ryan

Science fiction has had a way of anticipating the future. For a financial version of that, check out HBO’s third season of “Westworld”: Business mogul Engerraund Serac and his brother create an artificial intelligence (AI) program that predicts the stock market, allowing them to see what will happen 15 minutes into the future. Their hedge fund scores breathtaking gains.

Such miracles aren’t yet here. Indeed, artificial intelligence (AI), aka machine learning, today has yet to play a large role in investing. Certainly, investors are willing to work toward greater adoption, but machine learning is costly, hard to scale, and still very much in its nascency. It has probably played a greater role strengthening back office operations than it has augmenting investment decisions.

Still, artificial intelligence continues to increase its impact. As investors solve data problems, they’re rendering the trading process more efficient, reducing market frictions, and driving insights. Even if, for the moment, the process requires guidance by humans.

“Artificial intelligence techniques are already being incorporated in investment management processes, and these are bound to accelerate as more data and talent becomes accessible,” said Harshal Chaudhari, chief investment officer for global pensions at General Electric.

The Fast-Moving Hedge Funds

Hedge funds in particular have benefited from machine learning. More than half of hedge funds that responded to a BarclayHedge 2018 survey said they use artificial intelligence to inform investment decisions, up from just 20% the year prior. Around two-thirds said they were generating trade ideas, and more than a quarter said they were automating trade execution.

Those have helped some hedge funds outperform. In Europe, the total return of AI-oriented hedge funds from 2016 to 2019 was 33.9%, beating the 12.1% return overall for all hedge funds, according to an August report from Cerulli Associates. Some hedge fund shops in the game include Aidiyia Holdings, Cerebellum Capital, and Numerai.

A major weakness of machine learning today is that its ability to react to unforeseen or unprecedented events is yet untested. That point was highlighted in November, when Bloomberg reported that two AI-using investment firms, $75 billion Renaissance Technologies and $56 billion Two Sigma Advisers, saw losses, thanks to unforeseen volatility from the pandemic.

“Any data that you feed into an AI, into a machine learning algorithm, is by definition biased in its history,” said Adam Broun, chief executive officer at Kensho Technologies, a machine learning company based in Cambridge, Massachusetts.

AI shops face other challenges: Quantitative analysts, who construct mathematical models for investment decisions, tend to spread their probabilistic bets over many hundreds of companies, while traditional fundamental investors are more likely to invest in just a few dozen.

“[Quants] are taking far less business risk,” said Bill Coaker, chief investment officer at the San Francisco Employees Retirement System. “And I think where you make large sums of money and post high returns is by investing in great businesses, in leaders in innovation in how business will be done in the future.”

Where to Find Enough Data?

Ultimately, massive data sets are required to train machines to think. It’s the big reason why institutional investors find it so hard to incorporate AI capabilities in-house. Accessing and analyzing data sets is extremely expensive, costing a hefty chunk of operating budgets, and they are hard to scale. Skilled data and AI workers remain in short supply. Which is why good old-fashioned human thinking has continued to be the driver behind most investments.

Needing large sets of data is one reason why it’s still early days for investors to use financial reporting data or pricing data to make investment decisions, experts say. Using data from quarterly financial reports going back 20 years would yield a computer just 80 data points. That’s hardly enough to train a machine—not to mention, some companies are not that old.

But there are other ways to harvest data points. An alchemy of satellite imagery, courtesy of firms such as Orbital Insight, as well as social media, geotagging, credit card use, and other alternative data sources have helped those investors who will spend the money to find insights that deliver alpha.

Those alternative data sources can be especially helpful right now. For real estate investors, geotagging (mining digital information from phones, social media, etc., to see where people go) can show where people relocated during the pandemic. That has serious implications for both the residential and urban markets, both in a muddle.

For retail investors, credit card receipts, as well as satellite imagery of shipping terminals or parking lots outside shops, can generate a better picture of retail revenues. Satellite scanning can also determine the level of oil in tankers. Even vaccine treatments, including the success or non-success of thousands of clinical trials gleaned from company websites or social media alerts, can notify investors about the stock performance of various pharmaceutical companies, whether they be Moderna or Teladoc.

“We’re all trying to get a picture of where we are in the global economy or where the market is,” said Chris Natividad, chief investment officer at EquBot, the AI platform startup that in 2017 released an exchange-traded fund (ETF) with ETFMG, powered by IBM’s Watson. So far this year, the AIEQ has outperformed the S&P 500 by 8%, versus the benchmark, which is up less than 1%. “So, with more data, it’s more pixels and more precision.”

All of this means that, if investors can get it right, they can get better information about a company even before the company’s chief financial officer.

How AI Uses Language Analysis Now

One area where artificial intelligence is driving insights is in what’s called natural language processing. Namely, analyzing corporate and other communications for insights. Documents can be examined for clues about company plans. Company earnings calls can also be mined—not just for earnings data, but for the tone the speakers use, which can tell listeners how confident they are in their results.

Natural language processing is so crucial that in 2018, the S&P 500 Global acquired machine learning company Kensho Technologies. The startup built technology to anticipate stock movements prior to the company’s acquisition. But afterward, Kensho pivoted its business to organize the massive amounts of historical data from the index provider.

The many hours of earnings calls and transcripts helped the firm create what its CEO says is the world’s best transcription service for financial speech. The firm claims its offerings outclass other such services.

“Google does a great job and Apple does a fine job with kind of general-purpose audio for your shopping list or whatever, but they tend to be terrible when it comes to financial speech,” said Broun, the chief executive officer at Kensho.

“If you say something like ‘GAAP accounting,’ Google will say ‘gap accounting’ because it doesn’t understand ‘GAAP,’ right? Whereas we understand what GAAP accounting means, what, you know, 3 cents a share means or things like that. And so, that engine is actually currently being used by our parent company,” Broun added.

Meanwhile, asset managers in the Lausanne, Switzerland-based BERG Capital Management, which focuses on investment governance, have trained their natural language processors to separate well-run businesses from the poorly governed ones using 10-K filings.

Businesses with a forward-looking view are more likely to use the words “century” or “long-term” or “perpetual” in their 10-K forms, according to BERG Partner and Chief Technology Officer Rens Götz. Meanwhile, poorer-run firms are more likely to use shorter-term time references: “month,” “quarterly,” or “recently.” Thus, Götz said, “There’s a temporal distinction in those forms that can be measured.”

AI can draw other distinctions between companies. When it comes to human capital, artificial intelligence can measure the merit of words used to describe a company’s workforce like “expertise,” “grooming,” or “building” in a 10-K filing. A business with a poor track record of dealing with its employees? Prevalent words are: “dismiss,” “replacing,” and “non-qualified.”

At BERG, those findings are a starting point to evaluate and rank companies across categories and sectors—and they can drive results. Last year, the top quintile of businesses in the S&P 500 that scored best in governance outperformed the benchmark by 15 percentage points, according to the firm’s governance methodology. By and large, those top firms were in the tech sector, with future-oriented business plans, as well as with large budgets to allocate toward research and development.

Other applications for machine learning include back office operations, which can organize the hundreds of notices investors can get from limited partners or fund of funds for upcoming investment actions. AI can also be used for compliance checks, which is especially useful when cross checking investments against different geographies and legal structures around the world.

Goal: Artificial Humanity

Perhaps in the future, artificial intelligence will play a more altruistic role in investors’ portfolios than the dystopian vision that the “Westworld” creators employed. According to Andrew Lo, finance professor at the Massachusetts Institute of Technology (MIT), what is truly needed in the future is not artificial intelligence, but “artificial humanity.” In other words, machines that may one day make all our investing decisions based on people’s characteristics. In other words, AI would develop a bedside manner.

“What we really need is algorithms that understand us, that understand our psychology and are able to deal with that psychology, while maintaining our trust,” Lo said.

For retail investors, algorithms in effect could act as doctors that both individualize portfolios, as well as make the tough investing calls more timid investors may be unwilling to make. For institutions, artificial intelligence may be able to pick out and identify investing preferences among trustees.

Of course, much more work is needed to lay the legal, ethical, and political groundwork to create that future. But such a leap is only a matter of time, and trust.