Why Betterment’s robo-advisor doesn’t use AI

AI-based techniques were less effective than simpler approaches, the company found.
article cover

Dianna “Mick” McDougall, Photo: Max Lirnyk/Getty Images

· 5 min read

Betterment is one of the world’s largest robo-advisors, but its consumer-facing investment offerings make virtually no use of machine learning.

And according to Betterment, they just work better that way.

“We haven’t used machine learning [in our consumer-facing robo-advisor product] for a very specific purpose, because it hasn’t really addressed a core problem that we felt like machine learning could directly address in effective ways,” Mychal Campos, Betterment’s senior director of investing, told us in an interview. “Us not using it doesn’t mean we didn’t put the research into it. We did, and we found that, actually, we have much simpler approaches, which actually are more effective.”

Despite the “robo” part of robo-advisor, Campos recalled that when the team explored using machine learning for portfolio construction, they went down “kind of a rabbit hole,” and the AI-based portfolio optimization resulted in lower expected returns. The team only wanted to use more advanced technology if it filled a clear need and led to clear benefits, he said, so when faced with these results, they ditched the flashier tech for more traditional statistical modeling.

With $33 billion in assets under management through March 2022, Betterment eclipses independent competitors like Wealthfront but reportedly falls behind traditional investment companies like Vanguard and Schwab as far as robo-advisor asset value. The New York-based company also touts 730,000+ open accounts—for comparison, Schwab has 495,000 accounts and Wealthfront has 490,000. The company declined to disclose long-term performance statistics to Emerging Tech Brew.

Stickin’ to statistics

One machine-learning technique the team tried out was using neural networks for portfolio optimization.

“Neural networks can be great in some areas, but they can also lead to overfitting of models,” Campos said. “What that means is if you train a model on a specific set of data, it might be very good at explaining that data—but then when you actually put it onto the real world and its predictions, it actually doesn’t do very well.”

He added that typically “what you have with neural network models is you try to throw everything in the kitchen sink to try to predict things, and that oftentimes leads to less-than-optimal outcomes—where you’re oftentimes left with a model that doesn’t really predict the data it hasn’t seen.”

So if it doesn’t use machine learning, then how does Betterment’s robo-advisor product work? The answer involves a lot of math and statistical modeling, according to Campos, with the crux depending on two different models averaged together.

Keep up with the innovative tech transforming business

Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.

One set of techniques is based on Monte Carlo simulations, “where we are basically simulating thousands of market environments and essentially creating optimal portfolios in each of those thousands of scenarios” to come up with the best possible option, Campos said. The other is based on the Black-Litterman model, a portfolio-allocation framework.

“We can actually use those models as well, to tailor them to someone based on what someone’s risk tolerance is, what their time horizon is, for their investing goal or for their savings goal,” Campos said. “We can then go to the next layer of actually providing advice on what are their projected outcomes…likely to be, based on the portfolios that we’ve constructed for them.”

When a user signs up for Betterment’s consumer robo-advisor, they’re met with introductory questions like “What would you like to do?” and “What are you saving for?” and many different answer options—but according to John Mileham, CTO of Betterment, it didn’t start out that way.

“At the beginning, it was a very simple model that allowed you to choose your risk tolerance on a scale of zero to 100, and we had a bunch of pre-baked portfolios,” Mileham told us. He added, “But it didn’t have the ability to think about different kinds of advice, different kinds of goals that people might have…[or] the ability to take inputs about things like people’s actual retirement plans, their timelines, the different components of that retirement and their different tax preferences.”

Even though Betterment doesn’t use machine learning widely, the technology does factor into some of its offerings. One example: To give its advisor clients more choices in constructing portfolios for their clients, Betterment onboarded a lot of exchange-traded funds (ETFs) at once, so the team used machine learning to calculate the expected returns and volatility of each fund, according to Campos. And for Betterment’s upcoming cryptocurrency investment offering, he added, the company uses machine-learning clustering methods to construct portfolios, since crypto is a nascent market without as much data.

“[When] applying machine learning and AI, we always want to make sure that we know what problem we’re trying to solve, and we want to have a strong hypothesis that a machine learning method is the way to do that,” Campos said. “What we want to steer away from is having a whole suite of machine learning tools to try and find problems to solve.”

Keep up with the innovative tech transforming business

Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.