Experimental Design
Participants take part in an experimental stock market that lasts 10 time periods t. In each period t, the price level of all stocks in this market either increases or decreases; a price increase is always +6% and a price decrease is always -5%. Each stock has a pre-determined fundamental quality, which is a stock’s probability of a price increase each period (randomly drawn with replacement from the set {0.1, 0.15, 0.2, 0.25, …, 0.9}), and which remains constant throughout the experiment. Participants are not informed about the true fundamental quality of the stocks.
Prior to the main experiment, we collect recommendation data. Both human and algorithm advisors receive a set of stocks and additional price information for prior periods to give recommendations on. A buy recommendation should be given for any asset with an estimated fundamental quality at or above 0.5. Thereby, we harmonize the instructions presented to both human advisors and the algorithm. The algorithm advice is based on a Bayesian posterior. We collect human advice data from real financial advisors, who are employed at a German bank. Since we expect variation in the quality of recommendations between the human and algorithm advisors, we match algorithm recommendations with equivalent human recommendations prior to the main experiment to mitigate noise.
In the main study, participants choose a stock based on advisor recommendations and remain invested in this stock for all 10 periods. They are randomly allocated to one of two main treatments; The HUMAN ADVISOR treatment and the ROBO-ADVISOR TREATMENT. In the beginning, participants select one of the recommended stocks. They solely decide based on their advisor’s recommendations, knowing their advisor type; here, they are not informed about any previous price levels. To incentivize participants of the main survey, they are informed about receiving a payment according to a fraction of the value of their assets at the end of the experiment.
The main experiment is divided into two blocks: (i) belief elicitation and (ii) changes in behavior. In block i, participants iterate through a sequence of ten belief elicitation rounds. In each round, we present them with additional price information and ask about their current belief of the fundamental quality of the asset.
Additionally, we vary how many previous price levels human and algorithmic advisors saw when selecting a stock on behalf of the client. This number of previous price levels is unknown to participants. We ask clients how many previous price periods they believe their human/algorithmic advisor saw when selecting the stock three times in the experiment (at t=0, t=5, and t=10). This gives us an approximation of the perceived decision quality of the advisor.
In block ii we collect data on subsequent investor behavior conditional on the advisor type. Participants across both treatments are allocated to one of two tasks: (1) risk taking and (2) advisor switching. Participants in the risk-taking task are requested to distribute a portion of a fictional portfolio between the current (risky) stock and another risk-free asset. Both assets exhibit different expected outcomes. Participants in the advisor switching task choose between sticking to their initial stock (as recommended in block i) and investing in a new stock out of a recommendation list including both human and algorithm advice.
In a post-survey questionnaire, we collect data on participants’ self-perceived risk preferences (Dohmen et al., 2011), financial literacy (van Rooij, Lusardi & Alessie, 2011), and socio-demographics.