| Summary: | Customer demand constitutes a crucial source of uncertainty in designing and operating complex and costly urban last-mile distribution operations. To mitigate associated risks, companies are diversifying their last-mile delivery options, exploring new vehicle types, and engaging in varied contracting schemes, encompassing vehicle rentals and spot market capacity utilization. We introduce a sequential learning and optimization problem integrating demand forecasting into a tactical last-mile fleet composition problem under uncertainty. Specifically, we propose a novel forecasting infrastructure and several machine learning models to predict customer demand in the medium-term future with high granularity. These forecasting results are then integrated into a two-stage stochastic program to derive cost-optimal fleet compositions. A real-world case study focusing on an e-commerce retailer in São Paulo, Brazil, reveals the economic viability of stochastic fleet composition planning informed by highly accurate demand forecasts. Our results show that accurate
demand forecasts enable e-commerce retailers to make cost-minimizing tactical decisions about the size, vehicle type, and governance structure of the rented vehicle fleet. Furthermore, our framework underlines the importance of implementing integrated decisions, where a fleet composition design is interlinked with forecasting methods to mitigate uncertainties.
|
|---|