A Nonparametric Approach to Estimating Heterogeneous Demand from Censored Sales Panel Data

Analyzing historical sales data to draw conclusions on the underlying demand structure is a central foundation for sales planning, e.g. in assortment and revenue optimization. This contribution focuses on estimating the choice behavior of demand segments as well as their distribution from panel data featuring multiple consecutive sales observations. Existing methods in this area mostly utilize parametric models and estimation procedures that rely on some given information, i.e., expert knowledge. To overcome this requirement, we employ finite mixtures to model sales events over multiple time frames and obtain nonparametric demand estimators. The proposed approach requires no given assumptions over underlying distributions. Furthermore, we also introduce a hindsight approach to assign individual sales observations to demand segments. This contribution decreases the need for manual adjustments in demand estimation and allows practitioners to gain detailed insight in purchase behaviors. In an extensive simulation study, we benchmark the approach on different data sets and compare its results to those from published approaches. The study highlights that the approach shows superior performance for markets with heterogeneous demand.