Integrating Anticipative Replenishment Allocation with Reactive Fulfillment for Online Retailing Using Robust Optimization

Problem definition: In each period of a planning horizon, an online retailer decides how much to replenish each product and how to allocate its inventory to fulfillment centers (FCs) before demand is known. After the demand in the period is realized, the retailer decides on which FCs to fulfill it. It is crucial to optimize the replenishment, allocation, and fulfillment decisions jointly such that the expected total operating cost is minimized. The problem is challenging because the replenishment allocation is done in an anticipative manner under a push strategy, but the fulfillment is executed in a reactive way under a pull strategy. We propose a multiperiod stochastic optimization model to delicately integrate the anticipative replenishment allocation decisions with the reactive fulfillment decisions such that they are determined seamlessly as the demands are realized over time. Academic/practical relevance: The aggressive expansion in e-commerce sales significantly escalates online retailers’ operating costs. Our methodology helps boost their competency in this cutthroat industry. Methodology: We develop a two-phase approach based on robust optimization to solve the problem. The first phase decides whether the products should be replenished in each period (binary decisions). We fix these binary decisions in the second phase, in which we determine the replenishment, allocation, and fulfillment quantities. Results: Numerical experiments suggest that our approach outperforms existing methods from the literature in solution quality and computational time and performs within 7% of a benchmark with perfect information. A study using real data from a major fashion online retailer in Asia suggests that the two-phase approach can potentially reduce the retailer’s cumulative cost significantly. Managerial implications: By decoupling the binary decisions from the continuous decisions, our methodology can solve large problem instances (up to 1,200 products). The integration, robustness, and adaptability of the decisions under our approach create significant value.