Data driven forecasting and revenue management with exposure dependent reference prices

We use data on the e-retail business of TMall to consider scalable forecasting and revenuemanagement (RM) for thousands of items. We use forecasts that depend on the novel concept of asymmetric formation of RPs. The related literature studies RPs empirically and theoretically but ignores possible asymmetry in formation of RP and is often focused on selective items. Our main contributions are in establishing the extensive benefit to forecasting and RM from using RPs and in bringing forward the importance of asymmetry in the formation of RPs. We develop forecasts that capture the asymmetry in the formation of RPs via customers' exposure to prices measured by sales or clicks. We use binary classification to compare the effectiveness of different forecasts. We further formulate and numerically study the RM problem when forecasts are RP-dependent.We show that forecasts that capture asymmetry in the formation of RPs lead to 2%--5% higher revenue-weighted error than forecasts with a standard RP, and these have a 20%--25% revenue-weighted error than forecasts that ignore RPs. We observe that RM using forecasts that capture exposure effect on RP\ results in 3%--4% higher profit than RM using forecasts with standard RP and these may be 22% or more higher than when using forecasts that ignore RPs. This paper provides foundations for a data-driven, scalable, and effective implementation of forecasting and RM. Furthermore, the improved accuracy, tractability, and robustness of RM using forecasts with exposure effects on RP supports their usage in practice