Solving a Newsvendor Problem Driven by Environmental and Individual-level Data

Abstract

The extant data-driven newsvendor literature has not addressed the issue of incorporating individual-level data. We take up this issue and begin with learning a potential customer's buy/no buy pattern during each store visit, which depends on the interaction between individual features (such as gender, customer loyalty, and price sensitivity) and environmental features (such as seasonality, market price, and economic indicators). Suppose the purchasing patterns of all potential customers can be characterized by a target concept, i.e., a mapping to the Boolean domain from the Cartesian product of the environmental and individual feature domains. We then develop a model that integrates learning customer arrival and purchasing patterns, forecasting the aggregated demand based on the learned knowledge, and solving the optimal ordering quantity. Our solution adapts to the evolving environment and learns how to act effectively within it. Given accuracy and confidence levels, we also show the sample complexity required to learn an approximately correct concept under the framework of probably approximately correct (PAC) learning. Finally, we evaluate the information value of individual-level data by comparing our model with a quantile regression model. We find that the proposed model offers excellent performance, results in positive information values, and also provides useful managerial diagnostics concerning the purchasing patterns underlying aggregated demand data. In particular, we observe that individual-level data is especially valuable when the environment evolves to a new state.