What a Robo-Advisor teaches us about controlling algorithms

What a Robo-Advisor teaches us about controlling algorithms

Every investor has a friend who checks their portfolio twelve times a day. When the market dips, they sell. When it rises, they buy. When a news headline appears, they rearrange everything. Their portfolio is less an investment strategy and more a nervous habit. And, as decades of research in finance have shown, it usually performs worse than if they had simply left it alone.

Now imagine handing that friend a robo-advisor, an app that uses algorithms to manage investments automatically. Should the app let them fiddle? A study presented at the International Conference on Information Systems (ICIS) 2025 by Subin Kim, Daehwan Ahn, Bong-geun Choi, Jiyong Park, Jinseok Yang, Byungjoon Yoo, and Kyeongmin Yum explores exactly this question, using six years of real data from a South Korean robo-advisor. Their answer reveals an uncomfortable tension between what is good for investors and what is good for platforms.

Why give people control at all?

People are often reluctant to trust algorithms, even when algorithms outperform human experts. Researchers call this algorithm aversion. One well-known remedy is to give users some control. Studies have shown that people are far more willing to use an algorithm if they can tweak it, even slightly. Control makes people feel safer and more in charge.

But most of that research looked only at the first moment of adoption, the decision to start using an algorithm. What happens afterwards, over months and years, has received much less attention. And afterwards is exactly where platforms make their money, because a platform’s health depends on users staying engaged, logging in, and adding funds.

The authors identify two very different kinds of control. Decision control means the algorithm proposes an action, such as rebalancing your portfolio, and you simply approve or reject it. You remain the final authority but leave the details to the machine. Process control goes further. It lets you initiate a rebalancing yourself, on your own timing and according to your own judgement. You are not just approving the pilot’s decisions; you can grab the controls.

Six years of real investor behaviour

The study draws on 312 weeks of data from a robo-advisory platform, covering well over a million weekly observations. Over that period, the platform offered different levels of control to users, which allowed the researchers to compare behaviour and results under three regimes: no control, decision control only, and decision control plus process control.

They looked at two broad sets of outcomes. The first was engagement, measured through how often users logged in and how much additional money they deposited. The second was performance, measured as risk-adjusted returns, which reflect how well a portfolio performs relative to the risk it takes.

The paradox of control

The findings are striking and a little counterintuitive.

When users were given decision control, the ability to approve the algorithm’s suggestions, they paid less attention to the platform. Weekly logins fell by about a third. Additional deposits also dropped significantly. Yet their portfolios did better, with risk-adjusted returns about 4.4 percentage points higher than for users with no control at all.

Why would less attention lead to better results? The authors suggest that once people feel they have ultimate authority, they relax and trust the algorithm. They stop hovering. And because they stop hovering, they stop making the impulsive manual interventions that so often harm investment performance. Feeling in control, paradoxically, makes people more willing to let go.

When the platform added process control, allowing users to trigger rebalancing themselves, the picture flipped. Attention rose again by almost 30 percent and deposits recovered. Users were more engaged and more active. But the extra activity came at a price. Engaged users intervened more often, rebalancing by hand, and their returns fell by about 3.9 percentage points, wiping out almost all the gains from decision control.

Put simply, the more users could fiddle, the more they did, and the more they fiddled, the worse their portfolios performed. The platform got busier, the investors got poorer.

A dilemma for every AI platform

This creates a genuine dilemma for platform designers. From the investor’s point of view, decision control looks ideal: they feel in charge, trust the algorithm, and earn better returns. From the platform’s short-term business point of view, however, decision control looks worrying, since logins and deposits fall. Process control does the opposite. It drives the kind of engagement that looks great on a dashboard but quietly undermines the very performance that the algorithm was built to deliver.

The authors frame this as a choice between features that boost visible, short-term activity and features that preserve long-term value for clients. A platform obsessed with engagement metrics might be tempted to give users more buttons to press. But those buttons may encourage exactly the behaviour that hurts customers. In the long run, satisfied customers with better returns may be worth more than busy customers with worse ones.

The lesson extends well beyond finance. Consider a self-driving car that lets you override its decisions, or a medical AI that lets doctors adjust its recommendations. The right amount of control matters a great deal. A little oversight can build trust and catch genuine mistakes. Too much hands-on intervention can undo the benefits that the system was designed to provide.

The authors are candid that their field data cannot prove cause and effect beyond doubt, since users were not randomly assigned to different control regimes. They plan further analyses and a controlled laboratory experiment to confirm the causal links and measure the psychological mechanisms more directly.

For individual investors, though, the message is already familiar and worth repeating. Checking your portfolio constantly and rearranging it at every twitch of the market is rarely a winning strategy. If you have chosen an algorithm to manage your money, sometimes the smartest thing you can do is approve its decisions, close the app, and go for a walk.

As the old investing joke goes, the best-performing accounts often belong to people who forgot they had them.

Reference

Kim, S., Ahn, D., Choi, B., Park, J., Yang, J., Yoo, B., & Yum, K. (2025). Controllability in algorithmic decision-making: Field evidence from a robo-advisor. In ICIS 2025 Proceedings (Paper 4). Association for Information Systems.

Leave a Reply