TL;DR
Amazon researchers have developed a multi-stage optimization framework that automates the selection of fulfillment centers (FCs) for international shipping. By combining causal inference (Doubly Robust Estimation) with multi-objective linear programming, the model achieves up to a 51% reduction in delivery time and a 48% reduction in logistics costs compared to human-expert policies.
What Problem Does This Paper Address?
In cross-border (XB) e-commerce, shipping an item from a warehouse in the US to a customer in Germany or Australia is a complex logistical puzzle involving customs, volatile freight conditions, and long transit times. A critical decision is "regionalization": selecting which specific subset of existing warehouses should serve a particular international destination to minimize costs and maximize speed.
Currently, these subsets are often defined by human experts. However, manual selection is difficult to scale as networks grow and doesn't easily account for the complex trade-offs between delivery speed, shipping cost, and warehouse profit. Furthermore, historical data is "biased"—a warehouse might look efficient simply because it was historically assigned "easy" orders, making it hard for standard machine learning models to identify the truly optimal performers.
Methods
The authors introduce the XB FC Selection Model, a three-stage pipeline designed to select the best FCs for any given shipping corridor:
- Forecasting: The model uses short-horizon multivariate forecasting to predict next-period operational metrics (time, cost, and volume) for each origin-destination pair.
- Causal Candidate Generation (DRE): To strip away historical bias, the team employs Doubly Robust Estimation (DRE). This causal inference technique estimates the "Average Treatment Effect" (ATE) of selecting a specific FC. It answers the counterfactual question: How much profit would this FC generate if we selected it, regardless of how it was used in the past?
- Constrained Optimization (LP): Finally, a multi-objective Linear Programming (LP) model selects the final subset of FCs. It balances competing goals—minimizing time and cost while maximizing profit—under "Soft Regionalization" constraints.
As shown in Figure 1, the "Soft Regionalization" approach is key: it prefers selected in-region FCs but keeps out-of-region FCs as backups, ensuring that no customer demand goes unfulfilled even if the primary warehouses run out of stock.
Datasets
The model was evaluated using 12 months of production-scale shipment data from a large e-commerce company (Amazon). The study focused on three major international shipping corridors:
- US → Singapore (SG)
- US → Australia (AU)
- US → Germany (DE)
The researchers used "offline replay" and shipment-level counterfactual matching (Same-ASIN and Nearest-Neighbor) to simulate how the model would have performed compared to historical human decisions.
Key Insights and Ideas
- Profit as a Natural Regularizer: The authors found that optimizing for speed or cost alone often leads to "volume concentration," where a tiny number of warehouses get overwhelmed. By including profit in the objective function, the model naturally favors a broader, more sustainable distribution of orders across the network.
- The Power of Causal Inference: Standard "top-k" rankings based on observed speed are misleading because they don't account for selection bias. DRE allows the model to identify "hidden gems"—warehouses that are fundamentally efficient but were underutilized by human experts.
- Soft vs. Hard Constraints: Moving from "Hard Regionalization" (where only specific FCs can ship) to "Soft Regionalization" (where they are merely preferred) allows the system to capture efficiency gains without the risk of losing sales due to local stockouts.
Key Results
The XB FC Selection Model consistently outperformed human-expert policies and standard algorithmic baselines across all tested corridors:
- Delivery Time: Reduced by 8.5% to 51.1%.
- Logistics Costs: Reduced by 4.4% to 48.3%.
- Target Compliance: Improved the "5/5" target (delivery under 5 days and cost under $5 USD) by up to 6.5 percentage points.
- Pareto Dominance: Analysis showed the model consistently occupies the "Pareto frontier," providing the best possible trade-off between operational efficiency and volume retention.
Limitations and Future Work
The primary limitation noted is the lack of real-time inventory and capacity data at the FC level; the current model uses a 2% inventory proxy. Additionally, because only one shipping policy can be active per corridor at a time, traditional live A/B testing is impossible, necessitating the complex offline counterfactual evaluations used in the paper.
Future research aims to integrate explicit inventory data and explore Reinforcement Learning (RL) to jointly optimize both the selection of warehouses and the physical movement of inventory between them.
Summary updated