How JD’s AI‑Driven Inventory Selection & Allocation Earned the Wagner Prize

Abstract

In October 2024, JD Retail’s supply‑chain technology team won the Daniel H. Wagner Prize for outstanding innovation in inventory selection and allocation. This article introduces the award‑winning technologies and their practical applications.

Overview

Effective supply‑chain management and fulfillment are critical for e‑commerce customer satisfaction. JD achieves over 90% of self‑operated orders fulfilled within 24 hours, thanks to a robust logistics infrastructure.

At the INFORMS conference, JD’s team received the Wagner Prize for their data‑driven inventory selection and allocation algorithms, following a previous INFORMS prize earlier in the year.

Two‑Tier Distribution Network

JD operates a two‑tier network of Regional Distribution Centers (RDC) and Front‑line Distribution Centers (FDC). RDCs store a comprehensive SKU range, while FDCs, located closer to customers, have limited capacity and rely on daily replenishment from RDCs.

When an order is placed, it is fulfilled from the FDC if stock is available; otherwise, the RDC supplies the order, incurring higher cost and potential delay.

Problem Definition

The goal is to maximize FDC fulfillment rates while minimizing sales loss and inventory transfer costs, subject to capacity and business constraints. The problem is decomposed into two sub‑problems: inventory selection (which SKUs to stock in each FDC) and inventory allocation (daily transfer quantities from RDC to FDC).

Selection Algorithms

Two heuristic algorithms are proposed:

ML‑Top‑K : Uses machine‑learning forecasts of SKU sales to select the top‑K SKUs for each FDC.

Reverse‑Exclude : Iteratively eliminates low‑impact SKUs until only K remain, ensuring remaining orders can be satisfied.

Both methods improve local order satisfaction rates compared to baseline approaches.

Allocation Algorithm

An end‑to‑end learning framework integrates demand forecasting, safety‑stock calculation, and simulation‑based allocation. Multi‑task neural networks predict sales, generate target and safety inventory levels, and simulate fulfillment to minimize total loss.

The framework employs recurrent neural networks to capture multi‑period dependencies and uses a simulation module to evaluate inventory decisions.

Experiments and Deployment

Experiments on real JD data show that the ML‑Top‑K and Reverse‑Exclude algorithms increase local order satisfaction by 0.54 pp and 0.27 pp respectively, with a combined gain of 2.2 pp.

The end‑to‑end allocation algorithm achieves a 58.59% FDC demand satisfaction rate, outperforming parameter‑search (57.55%) and linear‑programming (57.12%) baselines.

Deployed across all eight RDCs (covering 5–13 FDCs each), the system reduces inventory holding costs by tens of millions of RMB annually, cuts transfer costs by over a hundred million RMB, raises inventory on‑hand rate by 0.85%, and improves local order satisfaction by 2.19%.

Conclusion

The paper presents data‑driven selection and allocation models that significantly enhance JD’s supply‑chain efficiency and customer experience. The algorithms have been successfully integrated into JD’s operational system, demonstrating the practical impact of AI and operations‑research techniques on large‑scale e‑commerce logistics.