Modeling Customer Choice in Retail Settings Using Novel Data Sources

The customer choice model plays a pivotal role in descriptive, predictive, and prescriptive analysis in theretail sector. It provides insights into customer preferences and behaviors, enabling retailers to tailor theirproducts, services, and operational strategies accordingly. As technology advances in both brick-and-mortarand online retail contexts, new sources of data become available. Internet of Things (IoT) technologies areparticularly transformative; for instance, heatmap sensors capture invaluable data on customer foot trafficpatterns and dwell times. In online contexts, clickstream data offers a granular view of customer behavior. Inmy dissertation, I delve into how we can utilize some of these novel data sources to enhance customer choicemodels in retail environments.In the context of brick-and-mortar stores, I introduce the utilization of heatmap sensors for modelingcustomer choice decisions. In the second chapter of my dissertation, I demonstrate that the addition ofheatmap data to sales data makes the identification and application of “consider-then-choose” modelspossible. Also, in collaboration with an apparel retailer, I demonstrate that using heatmap data can resultin distinct coefficient estimates, enhanced predictive accuracy, and improved revenue in product placementdecisions. In the third chapter, I develop a new consider-then-choose model for customer multi-item, multiquantitydecisions, grounded in the concept of direct utility. I utilize this model to study a retailer’s clearancedecisions, specifically focusing on location-tying strategies. Shifting to the online realm in my fourth chapter,I focus on the temporal information inherent in clickstream data. I harness this information to redefineconsider-then-choose models in online contexts, challenging the common assumption that clicking on aproduct equates to considering that product. I propose four distinct models to offer different perspectives onthe relationship between clicks and considerations. I assess the performance of these four models using acomprehensive dataset from JD.com.