Date : Jeudi | 2025-05-22 à 12h30
Lieu : Salle des thèses

Lien TEAMS : Cliquer ici pour rejoindre le séminaire doctorant du LÉO

Mehdi LOUAFI (LEO, Université d’Orléans)

This paper examines how algorithmic explainability shapes user decision-making and engagement on a leading French robo-advisor platform. In a large-scale field experiment, we randomly assigned 4,646 prospective clients to one of two conditions: a treatment group that viewed a graphical breakdown of the factors driving their personalized risk-profile recommendation, or a control group that received no additional information. Our primary analyses show that, on average, providing an explanation neither changes the probability of accepting the recommended risk score nor affects deviations from it. However, we uncover significant heterogeneity in treatment effects. First, among mobile users, explanations reduce the number of profile-adjustment attempts before contract signing. Second, for users who initially deviate conservatively from the algorithm’s suggestion, explanations prompt even more conservative portfolio choices. Finally, leveraging machine-learning methods, we estimate conditional average treatment effects across socio-demographic and economic subgroups, identifying additional segments for which explainability meaningfully alters behavior.