Data-driven Dynamic Pricing and Ordering with Perishable Inventory in a Changing Environment

We consider a grocery retailer selling a perishable product in a dynamic environment where consumers’ price sensitivity changes at unknown times (due to pandemics, weather events, etc.), and the product perishes at an unknown rate. We design online price experiments for learning about these unknown features over time. We then prescribe how to use the newly gained knowledge and the most up-to-date data to make informed joint pricing and inventory ordering decisions. Depending on whether the demand shock distribution is parametric or nonparametric, we design two versions of the data-driven pricing and ordering (DDPO) algorithm with the best achievable performance guarantee. Implementing our algorithm on a real-life data set from a supermarket chain, we show that our data-driven, learning-and-earning approach significantly outperforms the historical decisions of the supermarket chain by reducing the profit loss due to uncertainty by over 80%. In particular, avoiding active learning for price-sensitivity changes leads to an annual profit loss of over 62 million U.S. dollars; avoiding active learning for perishability results in a yearly profit loss of over 11 million U.S. dollars. (Joint work with Bora Keskin and Yuexing Li of Duke University.)