Using robo-advisers as commitment devices

A TFI research project by Alexia Gaudeul and Caterina Giannetti
Posted on December 03, 2020

“Our setting implies that a rational investor should buy when stocks go up and sell when they go down, which is the opposite of most people’s behaviour.”

Alexia Gaudeul and Caterina Giannetti report results from a stock trading experiment in which they tested the use of commitment devices to help individuals remedy with the disposition effect – the tendency to sell winning stocks and keep those that lose value. A commitment device is a self-imposed arrangement that helps people to stick to a decision they have made.

Download the report or read the summary bellow.


Most households don’t care much for the stock market, even though stock investments generate higher returns than other asset classes. Only about 1 in 5 households in the EU own any risky financial assets like mutual funds, bonds or shares (Arrondel et al, 2018). How can more households participate in the stock market, or at least gain awareness of its opportunities? We decided to find out to what extent robo-advisers might help. After all, robo-advisers - with their automated portfolio management algorithms - can provide personalized advice cheaply and transparently.

The use of robo-advisers is relatively high in the US (8% of households, 1/8 of the total value of ETFs), but it’s much lower in the EU (2% of households, 1/10 of the total value of ETFs). One of the main reasons might be what is known as ‘algorithm aversion’ (Dietvorst et al, 2015, 2018), occurring when individuals are reluctant to let algorithms decide for them - even when those algorithms are proven to perform much better. How could we tackle this phenomenon? We set up a field experiment in which participants could trade on an artificial stock market over a period of three weeks, playing out various scenarios with our participants whilst offering them advice.

The experimental setting

Our experimental setup draws on previous experimental research (e.g. Frydman et al 2014). Stock prices were made to evolve randomly, trending either upwards or downwards. The participants only saw prices, deciding when to buy or sell depending on how they thought prices would evolve in the future. In our very simple setting, it made sense to buy when the stock went up, because it was a sign it would go up and up, and to sell when the stock went down.

Our experiment reflects real trading conditions more closely than traditional experiments, while remaining very simple to analyse. Rather than running this experiment in an experimental laboratory, in which people sit in front of a computer continuously for about one hour, we decided to run the experiment online over the course of three weeks: participants could trade online from their own computer or smartphone. This allowed us to increase the interval between trading periods from seconds to hours (and days), so participants could develop their understanding of the market and devise their own strategies at their own pace.

In total, 450 people participated in our experiment, from June to August 2020. They were recruited from a pool of students at the University of Pisa (Italy) and were paid based on their trading performance: the higher the value of their portfolio at the end of the experiment, the more they earned. Surprisingly for such a long-running experiment, there was very little drop-out (6%), and participants traded regularly (about twice per day, out of three possibilities per day).

The algorithms

The advice we gave our participants varied in its type and its strength. We considered two types of algorithms: one imposing a passive “buy and hold” strategy by blocking trading every two periods, and one imposing an active “long-short” momentum strategy by buying recent winners and selling recent losers, also every two periods. The advice also varied in terms of whether our participants had the power to override the decisions taken by the algorithm.

The first algorithm (preventing trading decisions) was meant to help people make more considered decisions, while the second one (imposing trading decisions) was meant to help them learn how to trade optimally. Making the advice optional - not forcing individuals to follow it - was meant to overcome algorithm aversion (making participants feel more in control).

Our experiment lasted 21 days, with 3 trading rounds each day, each round lasting 8 hours.

  • In the first 7 days (week 1), all participants traded without the help of any algorithms, i.e. no robo-advisers;
  • In the following 7 days (week 2), participants were assigned different types of algorithms: one type allowing participants to trade only every two rounds, the other one trading according to the probability of the stock going up. With both types of algorithm, we implemented treatments in which participants could either override (soft) or not (hard) the algorithm’s actions.
  • In the last 7 days (week 3), participants were asked to choose whether they preferred to play as in the second week or as in the first week. See figure 1.

Figure 1: Timeline, algorithm (i.e. robo-adviser) imposed in the second week, optional in the third week

We also ran opposite scenarios: in those cases, the algorithm was in place in the first week and participants could play freely in the second week. In the third week, as in the standard treatments, the participants needed to decide whether to play with or without the help of a robo-adviser.


Our setting implies that a rational investor should buy when stocks go up and sell when they go down, which is the opposite of most people’s behaviour. Indeed, most individuals tend to keep stocks that go down and sell those that go up (also known as the ‘disposition effect’).

We have two main measures of success for our study. The first one is whether exposing investors to robo-advising helped them overcome the disposition effect. The second measure is whether our participants chose to adopt the robo-advisers they were exposed to. After a week of experience with algorithmic trading, did they then choose to rely on robo-advisers voluntarily? We were particularly interested in whether people who were the worst at trading were sophisticated enough to realize that relying on an algorithm would benefit them.

In line with previous research, we find that our participants are subject to the disposition effect when they do not get help from an algorithm. Participants who were more often active and who were more reflective (according to a cognitive reflection test) were also less subject to the disposition effect. The use of a robo-adviser reduced this effect, no matter what type of adviser, and whether it could be overridden or not. The more sophisticated algorithm, which traded instead of the individual rather than preventing trading, reduced the disposition effect more.

In the last week of our experiment, participants could decide whether to proceed with the assistance of an algorithm or not. On average, the take-up rate was quite low: only 36% of our participants decided to rely on a trading algorithm in the third week. They preferred softer, more sophisticated algorithms (+7% adoption): they liked being able to override the decision of the algorithm, and they preferred active rather than passive robo-advising. Although participants seem to prefer soft interventions, which may be less efficient, this does not result in lower payoffs compared to hard algorithms. Finally, participants who exhibited a stronger disposition effect were less likely to opt for help from a robo-adviser in the third week.

Practical implications

Our findings imply that robo-advisers improve the performance of individual traders, but that most choose not to use them. This is the case even for those traders who perform the worst without the help of a robo-adviser. In terms of adoption, the way the robo-adviser is designed does not seem to matter much but giving people the option to override the adviser helps overcome algorithm aversion.

We believe that further improvements in robo-advising should focus on making individual investors more aware of the benefits of robo-advisers, especially those investors who perform badly on the stock market. In our experiment, we found that even bad investors with the opportunity to use a robo-adviser for a while did not seem to realize how much better they performed with it than without it. However, allowing participants to experience markets on their own before offering robo-advising helped increase adoption compared to offering them robo-advising right at the beginning.

Further development should also give traders the opportunity to override decisions by robo-advisers. This was shown to increase adoption, while not significantly reducing performance in our experiment. There’s a simple way to overcome algorithm aversion: by offering the option to override trades by the robo-adviser, thus enhancing the feeling of being in control.


  • Arrondel, Luc, Laura Bartiloro, Pirmin Fessler, Peter Lindner, Thomas Y. Mathä, Cristiana Rampazzi, Frédérique Savignac, Tobias Schmidt, Martin Schürz, and Philip Vermeulen. 2018. “How Do Households Allocate Their Assets? Stylized Facts from the Eurosystem Household Finance and Consumption Survey.” International Journal of Central Banking, 12.2 (2016): 129-220.
  • Black, Fischer, and Robert Litterman. 1992. “Global Portfolio Optimization.” Financial Analysts Journal 48 (5): 28–43.
  • Kruger, J., and D. Dunning. 1999. “Unskilled and Unaware of It: How Difficulties in Recognizing One’s Own Incompetence Lead to Inflated Self-Assessments.” Journal of Personality and Social Psychology 77 (6): 1121–34.
  • Fischbacher, Urs, Gerson Hoffmann, and Simeon Schudy. "The causal effect of stop-loss and take-gain orders on the disposition effect." The Review of Financial Studies 30.6 (2017): 2110-2129
  • Frydman, Cary, and Colin Camerer. "Neural evidence of regret and its implications for investor behavior." The Review of Financial Studies 29.11 (2016): 3108-3139.
  • Frydman, Cary, and Antonio Rangel. "Debiasing the disposition effect by reducing the saliency of information about a stock's purchase price." Journal of economic behavior & organization 107 (2014): 541-552.
  • Frydman, C., Barberis, N., Camerer, C., Bossaerts, P. and Rangel, A., 2014. Using neural data to test a theory of investor behavior: An application to realization utility. The Journal of finance, 69(2), pp.907-946.
  • Ploner, Matteo. "Hold on to it? An experimental analysis of the disposition effect." Judgment & Decision Making 12.2 (2017).