JD.com Enhances Fulfillment Efficiency with Data-Driven Assortment and Inventory Management

Research Background

JD.com, one of the largest E-commerce companies in the world, has made significant strides in enhancing its fulfillment efficiency through data-driven assortment planning and inventory allocation. This research addresses the critical challenge of ensuring rapid and reliable delivery, which is essential for customer satisfaction and business growth.

The problem of optimizing fulfillment efficiency in a two-level distribution network, consisting of Regional Distribution Centers (RDCs) and Front Distribution Centers (FDCs), is multifaceted. FDCs, which are closer to customers, have limited capacity and rely on daily shipments from RDCs. If an FDC is out of stock, the order is fulfilled by the RDC, leading to increased costs and potential lost sales due to longer delivery times. The goal is to maximize the number of orders fulfilled by FDCs, minimize sales losses, and reduce inventory transfer costs.

In the context of the supply chain and AI decision-making, this research is particularly relevant. Traditional approaches often struggle with the complexity and scale of the problem, especially given the vast number of SKUs and the variability in demand. Prior methods, such as those based on discrete choice models, are not well-suited for warehouse settings where demand is highly dynamic and influenced by various factors like promotions and marketing events. Additionally, the computational intractability of large-scale mixed-integer programs limits the effectiveness of these methods in real-world applications.

JD.com’s two-level distribution network includes eight RDCs across China, each serving broad regions and stocking all SKUs sold online. FDCs, which are closer to customers, provide rapid delivery but have smaller capacities and rely on daily shipments from RDCs. When an order is placed, it is fulfilled from the corresponding FDC if sufficient inventory is available. If the FDC is out of stock, the RDC fulfills the order, incurring extra costs due to longer distances and potential lost sales from extended delivery times. Given the importance of FDCs for fast delivery and their capacity constraints, the selection of products to store at FDCs and the inventory management at FDCs and RDCs are critical for maintaining fulfillment efficiency.

The problem is challenging for several reasons:
1. **Coordination Complexity**: A single RDC serves multiple FDCs and has its own demand requirements, necessitating intricate coordination of its own demand fulfillment and inventory transfers to various FDCs. This also involves various business constraints, such as capacity constraints at FDCs, logistics throughput, and sales plans.
2. **Scalability**: JD.com manages a vast number of SKUs across different product categories, requiring the algorithm to be scalable.
3. **Demand Variability**: The demand at FDCs is relatively sparse and variable, complicating demand prediction.
4. **Future Uncertainties**: Assortment and inventory decisions must account for future uncertainties to align with long-term sales plans and mitigate risks from unforeseen events.
5. **Interpretability**: The interpretability of the prediction algorithm is crucial for coordinating different sections of the supply chain.

To address these challenges, the authors propose an integrated optimization model aimed at maximizing the number of orders fulfilled by FDCs while considering various business constraints. Due to the complexity of the problem, even a static version with deterministic demand is NP-complete. Additionally, the problem size is substantial since each RDC manages over 10 million SKUs. Consequently, the problem is decomposed into two subproblems: the assortment planning problem and the inventory allocation problem.

Key Findings

The paper presents several key findings that demonstrate the effectiveness of the proposed data-driven approaches in improving fulfillment efficiency. These findings are supported by rigorous experimental setups and concrete numerical evidence, highlighting the superiority of the new methods over existing ones.

Assortment Planning: ML-Top-K Algorithm

The ML-Top-K algorithm uses machine learning techniques to predict the number of orders containing each product and ranks them in descending order. The top K products are then selected for stocking at FDCs. In numerical experiments, this method achieved a local order fulfillment rate of 87.5%, compared to 86.9% for the baseline method. This improvement, though seemingly small, translates to a significant number of orders when scaled across JD.com’s extensive network.

The ML-Top-K algorithm leverages historical order data and machine learning models to predict the demand for each SKU. The algorithm first trains a predictive model using historical order data, which includes features such as product category, seasonality, and promotional activities. The model then predicts the number of orders containing each product for a specific period, typically one week. The products are ranked in descending order based on these predicted order counts, and the top K products are selected for stocking at FDCs. This approach ensures that the most frequently ordered products are available at FDCs, thereby maximizing the number of orders that can be fulfilled locally.

Experimental results show that the ML-Top-K algorithm outperforms the baseline method, which relies on a simple ranking of products based on historical sales. The improvement in local order fulfillment rate, from 86.9% to 87.5%, may seem modest, but it translates to a significant number of orders when scaled across JD.com’s extensive network. For example, a 0.6% increase in local order fulfillment rate, when applied to millions of weekly orders, can result in thousands of additional orders being fulfilled by FDCs, reducing the reliance on RDCs and improving overall efficiency.

Assortment Planning: Reverse-Exclude Algorithm

The Reverse-Exclude algorithm iteratively eliminates products from the consideration set based on their historical order frequency. This heuristic achieved a local order fulfillment rate of 87.2%, slightly lower than the ML-Top-K algorithm but still higher than the baseline. The algorithm’s simplicity and efficiency make it a viable alternative, especially in scenarios where computational resources are limited.

The Reverse-Exclude algorithm starts with the full set of SKUs and iteratively removes the least frequently ordered products until the remaining set of products fits within the capacity constraints of the FDCs. This approach is computationally efficient and does not require complex machine learning models, making it suitable for scenarios with limited computational resources. The algorithm’s simplicity also makes it easier to implement and maintain, which is an important consideration for practical applications.

In the experiments, the Reverse-Exclude algorithm achieved a local order fulfillment rate of 87.2%, which is slightly lower than the 87.5% achieved by the ML-Top-K algorithm. However, the difference is minimal, and the Reverse-Exclude algorithm’s simplicity and efficiency make it a valuable alternative. For instance, in scenarios where the computational overhead of training and deploying machine learning models is a concern, the Reverse-Exclude algorithm can provide a cost-effective solution without a significant loss in performance.

Assortment Planning: Hybrid-Selection Algorithm

The Hybrid-Selection algorithm combines the outputs of the ML-Top-K and Reverse-Exclude algorithms to leverage the strengths of both. This approach achieved the highest local order fulfillment rate of 87.8%. The hybrid method’s ability to balance prediction accuracy and computational efficiency makes it particularly effective in large-scale operations.

The Hybrid-Selection algorithm integrates the predictions from the ML-Top-K algorithm and the rankings from the Reverse-Exclude algorithm. It first applies the ML-Top-K algorithm to generate a preliminary list of top-ranked products. Then, it uses the Reverse-Exclude algorithm to iteratively remove the least frequently ordered products from this list until the remaining set of products fits within the capacity constraints of the FDCs. This hybrid approach combines the high accuracy of the ML-Top-K algorithm with the computational efficiency of the Reverse-Exclude algorithm, resulting in a more robust and effective solution.

In the experiments, the Hybrid-Selection algorithm achieved the highest local order fulfillment rate of 87.8%. This improvement, compared to the 86.9% baseline, demonstrates the effectiveness of combining different heuristics to achieve better performance. The hybrid approach is particularly useful in large-scale operations, where the need for both accurate predictions and efficient computation is critical.

Inventory Allocation: End-to-End Learning Framework

For the inventory allocation problem, the paper introduces a novel end-to-end learning framework that integrates forecasting, optimization, and simulation. This framework minimizes total sales losses and inventory transfer costs. Experimental results show that the FDC demand satisfaction rate increased by 1.05% compared to the previous method. The multitask and self-learning capabilities of the framework, along with its interpretability, provide significant advantages in managing complex and dynamic inventory systems.

The end-to-end learning framework for inventory allocation is designed to address the challenges of demand uncertainty, the interplay between FDCs and RDCs, and the interdependence of inventory decisions across multiple time periods. The framework consists of three main components: forecasting, optimization, and simulation. The forecasting component uses machine learning models to predict future demand, the optimization component determines the optimal inventory levels, and the simulation component evaluates the performance of the inventory decisions.

The framework’s multitask learning capability allows it to simultaneously learn inventory allocation decisions across FDCs and RDCs, and demand prediction to minimize the total cost and demand prediction error. This facilitates information sharing and knowledge transfer between tasks, improving the efficiency and accuracy of inventory management. The self-learning mechanism eliminates the need for complex operations research algorithms previously used for data labeling, significantly enhancing the efficiency and accuracy of sample processing in inventory management.

Experimental results show that the end-to-end learning framework achieved a 1.05% increase in FDC demand satisfaction rates compared to the previous method. This improvement is significant, especially considering the high volume of orders processed by JD.com. The framework’s interpretability, provided by its modular structure, enhances the understanding of the inventory management process and facilitates more informed decision-making.

Limitations

While the proposed methods offer substantial improvements, they also come with certain limitations and areas for further research. These limitations include the challenges of demand prediction, the complexity of the optimization problems, and the need for continuous adaptation to changing market conditions.

• Demand Prediction Accuracy: The accuracy of demand predictions is crucial for the success of the assortment and inventory allocation algorithms. However, demand in E-commerce is highly variable and influenced by numerous factors, making it challenging to achieve high accuracy. This limitation can lead to suboptimal decisions, especially during periods of high demand volatility. Potential mitigations include incorporating more sophisticated machine learning models and continuously updating the models with new data. For example, using ensemble methods or deep learning models that can capture complex patterns in the data may improve prediction accuracy. Additionally, implementing a feedback loop where the models are regularly retrained with the latest data can help adapt to changing market conditions.
• Computational Complexity: Although the heuristics and end-to-end learning framework are designed to be computationally efficient, they still face challenges in scaling to extremely large problem sizes. For instance, the assortment planning problem is NP-hard, and even with heuristics, finding exact solutions may be infeasible. To address this, future work could explore more advanced approximation algorithms and parallel computing techniques to handle larger datasets. For example, using distributed computing frameworks like Apache Spark or leveraging GPU acceleration can help scale the algorithms to handle the vast number of SKUs and orders managed by JD.com. Additionally, developing more efficient heuristics that provide near-optimal solutions with lower computational overhead can further improve scalability.
• Interpretability vs. Performance Trade-off: While the end-to-end learning framework offers improved interpretability compared to traditional deep learning methods, there is a trade-off between interpretability and performance. More interpretable models may sacrifice some predictive accuracy. Balancing this trade-off requires careful design and validation, potentially involving human-in-the-loop approaches to ensure that the models are both accurate and understandable. For instance, using explainable AI techniques such as SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) can help provide insights into the decision-making process of the models. Additionally, involving domain experts in the model development and validation process can help ensure that the models are both accurate and interpretable, facilitating better decision-making in the supply chain.

Practical Implications

The findings of this research have several practical implications for supply-chain and AI practitioners, providing concrete implementation paths and decision-making frameworks. These implications can be applied in various scenarios to enhance operational efficiency and customer satisfaction.

• Enhanced Local Order Fulfillment: By implementing the proposed assortment planning algorithms, companies can significantly increase the number of orders fulfilled by FDCs. This not only reduces delivery times but also decreases the reliance on RDCs, leading to cost savings and improved customer experiences. For example, JD.com’s implementation of the Hybrid-Selection algorithm resulted in a 0.54% increase in local order fulfillment rates, which, when scaled across millions of orders, translates to substantial benefits. This improvement can be replicated in other E-commerce companies with similar distribution networks, leading to faster delivery times and higher customer satisfaction.
• Optimized Inventory Management: The end-to-end learning framework for inventory allocation provides a robust solution for minimizing lost sales and inventory transfer costs. By integrating forecasting, optimization, and simulation, the framework enables more accurate and efficient inventory management. This can be particularly beneficial for companies with complex and dynamic supply chains, such as those in the E-commerce industry. For instance, the framework’s ability to adapt to changing demand patterns ensures that FDCs are always well-stocked, reducing the risk of stockouts and overstocking. Companies can implement this framework to optimize their inventory levels, leading to reduced holding and transfer costs and improved stock availability.
• Data-Driven Decision-Making: The use of data-driven approaches in assortment planning and inventory allocation highlights the importance of leveraging advanced analytics and machine learning in supply chain management. Companies can adopt similar methodologies to optimize their operations, using historical data and predictive models to make informed decisions. This approach can be extended to other areas of the supply chain, such as demand forecasting, transportation planning, and supplier management, to create a more integrated and efficient system. For example, companies can use machine learning models to predict demand for different products, optimize transportation routes, and manage supplier relationships, leading to a more streamlined and responsive supply chain.

Source: https://arxiv.org/abs/2509.12183