Theme: Learning Event Probabilities by Doing: The Case of Taxi Drivers

Lecturer：AssistantProfessor Wang Shuxiao (Department of Marketing, International Business School, University of International Business and Economics)

Host: Associate Professor Chen Xingyu (College of Management, Shenzhen University)

Time: 10:00 am -12:00 pm, December 8, 2023 (Friday)

Location: A203, Mingli Building, Lihu Campus, Shenzhen University

Lecturer profile: Wang Shuxiao is currently an assistant professor in the Department of Marketing at the International Business School of the University of International Business and Economics. He graduated with a PhD degree from the School of Business at the National University of Singapore. His research interests include quantitative marketing, platform economy, and social media.

Lecture Summary: The literature on learning has focused on learning the payoffs of events. We propose a dynamic discrete choice model that incorporates a Dirichlet learning framework to study the learning of event probabilities. Empirically, we use a unique data set of detailed trip records and trajectories to study taxi drivers’ location choices via learning by doing. Taxi drivers are assumed to choose where to search for passengers by solving a dynamic programming problem, and update their belief about demand conditions by accumulating experience in a Dirichlet learning fashion. We find that taxi drivers improve their productivity mainly by learning from experience to make more informed decisions on where to search for passengers. Loss aversion plays an important role in driving the learning rate: Drivers learn faster in low-expected-income areas than in high-expected-income areas due to the perceived earnings loss. Moreover, the learning rate evolves dynamically, starting at a limited level, gradually reaching its peak after 50 to 80 shifts, and then declining to a higher level. Counterfactual analyses reveal that new taxi drivers experienced a significant daily earnings loss – equivalent to 14% of their net daily earnings – due to their inaccurate prior belief about the demand conditions. We propose recommendation systems that guide new drivers to locations with the best earning opportunities and find it can accelerate their learning and mitigate their earnings loss.