A data-driven approach for strategic inventory placement in multi-echelon supply networks

This paper provides a data-driven solution for optimizing inventory buffers in large-scale supply networks. We study the placement and sizing of strategic inventories in multi-echelon supply chains where the decision maker faces uncertain demand with an unknown distribution influenced by explanatory variables. State-of-the-art multi-echelon inventory optimization models, such as the well-known guaranteed-service model (GSM), are non-linear and typically informed by distributional and parametric assumptions. They often rely on dynamic programming and are difficult to solve for large networks. We adapt the GSM to introduce a nonparametric, feature-driven approach to supply chain safety stock optimization that is based on mixed-integer linear programming (MILP). The MILP formulation sets cost-optimal base stocks, which are learned as linear functions of feature data under consideration of service level requirements. This integrated estimation and optimization approach is solved with commercial mathematical programming solvers and is enhanced by a Benders decomposition method for large networks. We extend the literature on data-driven inventory control by a multi-period and multi-echelon approach for safety stock planning in general, acyclic networks. On the real-world networks from Willems (2008), we find that incorporating feature information when setting safety stocks in large supply chains, on average, reduces operational costs out-of-sample. This value of feature information that the proposed model offers to decision-makers increases in demand volatility and is dependent on certain network characteristics.