Bandit's Paradise: Customer Acquisition Through Online Display Advertising

• Activity Number: 219
• Type: Invited
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Marketing
• Author(s): Eric M Schwartz*+ and Eric Bradlow and Peter Fader
• Companies: University of Michigan and Wharton, University of Pennsylvania and Wharton, University of Pennsylvania
• Keywords: Adaptive marketing experiments ; multi-armed bandit ; probability matching ; Thompson sampling

Abstract

As business experiments become more popular, firms move beyond "testing and learning" towards "earning while learning." Online advertisers, for instance, deliver dozens of different ad creatives across websites with the goal of acquiring customers. Over time, they adapt by allocating more impressions to better performing ads on each website. But how can they do this as profitably as possible? We frame this as solving the multi-armed bandit problem. This paper's contribution is to extend existing methods for bandit problems to handle various components of real-world adaptive marketing experiments (e.g., attributes of actions, unobserved differences in context, batched decisions). The key innovation is applying hierarchical models to randomized probability matching to solve the bandit problem with those components. The benefits of this new approach are demonstrated through a field experiment with a large retail bank using online advertising to acquire customers. The approach is compared to benchmark methods via simulation studies, providing insights into the types of the bandit problems in which certain types of methods to outperform others.