Can Humans Understand How Robots Invest?
For a guy who built a robot he hopes will banish human emotion from the investing process, Chida Khatua spends a lot of time trying to figure out how it thinks.
Khatua is chief executive officer of EquBot, a San Francisco company that’s built an artificial intelligence system for investing. In October, a day before the launch of an exchange-traded fund that uses EquBot recommendations, his team was going over stocks the computer wanted to buy. One name popped out: Brookdale Senior Living Inc., which operates retirement communities and nursing homes. This was back when wildfires were burning parts of California, where some of Brookdale’s facilities sit.
The trade looked off to Khatua, a former Intel Corp. engineer. Buy a company caught up in a natural disaster? But on a second look, it wasn’t hard to put together what the computer might have been thinking. News reports and press releases—all fed into the system—showed how Brookdale was responding to the threat. “We found, hey, that senior living facility—they have a very good, organized setup” and could provide backup housing, Khatua says. The ETF bought the stock and made a small profit on the trade. Scanning news wires may not sound like a human equity analyst’s idea of deep research, but for a computer it’s all data that can be combined with other information to make statistical predictions.
That’s the tricky thing about artificial intelligence and investing: If AI has an edge, it’s because it’s putting together a jumble of information in ways that humans wouldn’t. But when people trust their money to a fund, they’d like to be able to understand how the manager—or computer—makes decisions. A program reads about wildfires and buys a stock after deciding management will end up looking good in a crisis? Talk like that gets some AI purists’ hackles up. “It’s very important to separate the reality of what’s going on from the marketing being put around it,” says Andrew Dyson, CEO of QMA, an investment firm that uses quantitative techniques and big data. “People love a story, right? And there’s a real danger that these things are stories, and people have to get beyond the story and actually understand what’s going on.”
The fund EquBot’s model makes recommendations for, the AI Powered Equity ETF, launched in October and has quickly amassed $136 million in assets, making it one of the most successful ETF debuts of 2017. Drawing computational muscle from International Business Machines Corp.’s Watson platform, EquBot’s system assesses more than 6,000 U.S. publicly traded companies each day. It scrapes millions of regulatory filings, news stories, management profiles, sentiment gauges, financial models, valuations, and bits of market data. Then it chooses about 30 to 70 stocks for the fund, which is run by ETF Managers Group LLC. It’s not the first ETF to use AI in some way—one employs it to spot changes in market sentiment—but its backers say it’s a pioneer in using the technology to look at multiple components of an investment to build a portfolio. “It’s like employing an army of equity analysts,” says Khatua.
Since it started, it’s beaten the S&P 500 in 12 weeks and trailed in 13 of them. A Bloomberg analysis shows that after being hurt early on by its bets on smaller and more volatile companies, the ETF recovered by buying bank stocks. Its 1.9 percent total return falls just shy of the S&P 500’s 2.5 percent. So it’s been about average—but that record is much too short to be meaningful. Luck and the behavior of the overall market are the main influences on a diversified portfolio’s performance for much longer than most people realize, says James White, CEO of Elm Partners Management, an investment adviser. It could be a decade or more before anyone can say whether EquBot is selecting stocks with more skill than a dart thrower, White says.
EquBot’s chief operating officer and co-founder, Art Amador, says he trusts the AI to make a decision and run its course. “From a principle standpoint, we don’t want to intervene, we don’t want to create any bias under any circumstance,” says Amador, who met Khatua in class when they were students at the University of California at Berkeley Haas School of Business. “We didn’t tell it, ‘Oh, no, you’re not going to do this because it doesn’t make sense logically.’ ”
Any AI system is likely to make investment decisions that look puzzling, says Zachary Lipton, an assistant professor in the machine learning department at Carnegie Mellon University. In the strictest sense, a model “is not operating according to logical rules. The model is just spewing out statistical correlations,” he says. “It’s not giving you logic—there isn’t actually a chain of coherent logical reasoning that tells you how to invest in the stock market. If there was one, you wouldn’t need the model in the first place.”
Even so, Amador and Khatua say they run additional checks on the bot’s output. One is to make sure that the data under review are solid—for instance, that the computer isn’t scraping a website that recently changed its format, which could cause the computer to misread it. The other is what they call a “sanity test” to see if the choices make sense, based on how the program was trained. “Our core philosophy is that we don’t want to create a black box for AI,” says Khatua. Their goal is to have a system that operates “the way a good rational investor would think about and go through the process to decide whether a good investment opportunity is there or not,” he says.
The EquBot system is also designed to learn as it goes, according to the team. In the early days of the ETF, the program started buying companies without knowing how much money was about to land in its coffers. A bunch of small stocks that made sense as slivers of a $5 million portfolio proved too hard to trade when the portfolio quickly swelled to $40 million. The program eventually learned how to account for a stock’s trading volume when deciding what time to buy or sell it, and stopped picking up microcap stocks.
Tammer Kamel, CEO of Quandl Inc., an alternative data platform, understands the part about EquBot’s system seeing opportunity in California’s fire-ravaged countryside. “That’s classic AI,” he says. But caring too much about a program’s reasoning starts to get problematic. “As long as you persist in sanity-checking the output of your AI, then it will never be smarter than humans,” Kamel says. “I get it. In the early stages, you want to see if this thing is incorrectly programmed or has bugs—yeah, you have to watch out for that. But sooner or later, you have to take the reins off and trust in the technology.”