How the ‘world's first’ AI-managed ETF stacks up, almost five years later

A director of engineering at Intel, a vice president at Fidelity Investments, and an investment portfolio manager for Apple walk into a business school class…and walk out with an idea for the first-ever ETF that’s managed entirely by AI.

It was 2016, and Chida Khatua, Art Amador, and Chris Natividad were hunched over desks at the University of California, Berkeley, each in search of a way to leverage their resumé milestones into a successful company. Day in and day out, they listened to guest lecturers speak about their experiences managing hedge funds. But though each hedge-fund manager had a wealth of knowledge in their own domain—like commodities or global markets—their expertise seemed to stop there.

For the future co-founders, it sparked an idea: What if they could use artificial intelligence to bring all that sector knowledge together, capturing data faster and more broadly than humans could—and turn it into investment insights?

Within a year, they were doing just that. In 2017, the three co-founders debuted AIEQ, billed as the “world’s first” equity ETF fully managed by AI. It started with a team of about 10—themselves, along with seven developers in India. That first year, the fund managed $60—$70 million in assets. Since then, the company has tripled the team, and the fund now uses 80,000 models to analyze approximately 6,000 US companies. It manages about $150 million in assets, a small slice of the ~$5.2 trillion managed across the 1,596 ETFs in the US.

In the years since AIEQ’s debut, it first underperformed against the US market, then matched its performance, and finally surpassed it in 2020, when it beat the S&P 500 by at least seven percentage points. But AIEQ has also hit some snags: Tech and health care bets temporarily derailed performance in Q2 2021, and the recent tech sell-off contributed to the fund dropping 12% year to date. It's up 57% since inception, though the S&P 500 Total Return index is up 84% over the same span.

“We have seen investors become more and more comfortable with quantitative, algorithmic funds over time, and AI-run funds seem like a likely next step in the evolution of asset management,” Daniel Weagley, assistant professor at the Georgia Institute of Technology, told us via email. “In the longer run, I don’t expect the average AI-run fund to outperform index funds, especially as more funds enter the market and compete away any alpha that the AI-models identify.”

How it works

AIEQ is run by EquBot, the co-founders’ AI investment platform and portfolios as a service (PaaS) company. The ETF uses EquBot’s tens of thousands of proprietary models, and each day, the platform focuses on pulling data points related to the 6,000 US companies it tracks. Since 80,000 AI models is far too many for humans to manage, the company leans on IBM Watson to monitor them.

AIEQ learns from structured data, the traditional kind that’s always been used in the finance world, like revenue, growth, R&D expenditure, and market movements. But it also gets insights from unstructured data—think: insights from news articles, blogs, corporate innovation announcements, and social media.

The models have been trained on historical data ranging from five to 30 years, weighing recent data more heavily. The models are trained on a cost function, Khatua told us, meaning that for each historical data point—like, say, a news article from 2001—the model predicts an expected return. There are different trust scores involved, too; for instance, it’s trained to weigh a New York Times article differently than a blog post. EquBot feeds all of that data into knowledge graphs, which can then be used by AIEQ as training tools.

While the system relies on 80,000 models, three of the types hold outsized influence.

First up is a model that charts out a company’s financial picture for varying time horizons, using mostly earnings and spending data. Then there’s a model that uses about 170 line items, like innovation ranking, to determine a company’s current quality. Finally, there’s what Khatua calls the “infamous model,” which uses IBM Watson’s natural language processing tools to extract metadata and analyze the sentiments of more than a million pieces of content per day.

EquBot uses internal tools and IBM Watson’s OpenScale tool to monitor 10 metrics on each model that can help flag bias or “model drift,” as well as track each model’s judgment calls via decision trees of sorts. If customers ask for the “why” behind a decision, Khatua said, EquBot can provide them with the data points behind it.

“Right from the beginning, we wanted to make sure that every insight, every decision, we can go back and look at the data point, and we can explain it,” Khatua told us.

Two people on the company’s nearly 30-person team are responsible for watching those potential bias-related metrics full-time, while the individual owners of each model are responsible for checking for any red flags once per day.

This year, the team plans to introduce a public tool showing the underlying data and explainability behind the fund, Khatua said. The new platform is chiefly designed for asset and wealth-management companies to experiment with, as well as to gauge customer interest in different aspects of the fund, but since it will be available for anyone to look at, it could also make external audits possible.

“That platform is in development; we’re trying to bring it out as soon as possible for asset management companies, for wealth management,” Khatua said. “From there, we can also learn what exactly our customers [are] interested in, so we can tune our product lineup along with it.”

Update: This piece has been updated to include both AIEQ and the S&P500's total returns, which includes dividends and capital gains in addition to price returns. The initial version of this piece included only price returns.