Shortening Decision Chains: End-to-End Inventory Management and Intelligent Replenishment in JD's Supply Chain

“Try to make the decision chain as short as possible. If a decision can be done in one step, don’t split it into two; this reduces personnel and improves precision, using an end‑to‑end approach for inventory management. This is a research direction we are pursuing in Silicon Valley.”

At the 2018 China‑Europe Global Supply Chain Forum, JD.com’s Chief Scientist of Intelligent Supply Chain and Chair of Tsinghua University’s Industrial Engineering Department, Dr. Shen Zuo‑jun (a tenured professor at UC Berkeley), presented to more than 500 experts from logistics technology, big data, intelligent supply chain, and supply‑chain finance, including members of the US National Academy of Engineering and leading scholars from CEIBS.

JD operates over 500 million SKUs, eight regional distribution centers (RDCs), and more than 500 warehouses, with a daily inventory value of 500 billion RMB. Despite this scale, JD achieves an average inventory turnover of 24 days, an in‑stock rate above 95%, and an automatic replenishment rate exceeding 85%.

How is this achieved? JD employs an intelligent multi‑level replenishment system that selects different replenishment models based on product demand patterns—for stable‑selling items it uses a basic safety‑stock model, for long‑tail items a continuous replenishment model, and so on. The system combines algorithmic optimization, machine‑learning techniques, and manual inputs from sales staff to enhance replenishment accuracy.

The supply‑chain community knows the “bullwhip effect,” where information distortion amplifies as it moves upstream, causing overstock or stock‑outs. Professor Shen proposes a novel concept: “shorten the decision chain.” By making decisions in a single step, personnel are reduced and precision improves; the approach skips separate forecasting and inventory steps, directly outputting replenishment quantities.

Using an end‑to‑end method for inventory management is a completely new theory in the supply‑chain field. According to Shen, this algorithm significantly improves all inventory metrics and reduces the need for parameter tuning.

Below is a concise version of Professor Shen’s speech:

1. Inventory Management

JD’s inventory landscape is complex: over 500 million SKUs, eight distribution regions, and more than 500 warehouses. Daily inventory value reaches 500 billion RMB, with an average turnover of 24 days, an in‑stock rate of 95%, and an automatic replenishment rate of 85%. Over 90% of orders are delivered the same or next day, far surpassing typical US retail delivery times.

To handle this complexity, JD uses an intelligent multi‑level replenishment system. From supplier to RDC to front‑line warehouses, the process is called the intelligent replenishment system, while transfers from RDC to smaller warehouses are termed intelligent allocation.

The core of the intelligent replenishment system is forecasting a distribution rather than a single point estimate. After forecasting, business‑person inputs are incorporated to fine‑tune parameters. Although many machine‑learning methods are employed, human expertise remains essential for such a complex decision‑making environment.

Based on current inventory levels and business goals (e.g., turnover rate, in‑stock rate), the system outputs a replenishment strategy, including target inventory, replenishment points, and quantities.

Different product types receive tailored replenishment models: stable‑sales products use a basic safety‑stock model, long‑tail products use a continuous replenishment model, and numerous optimization and machine‑learning techniques enable automatic parameter adjustment, achieving high‑precision automatic replenishment.

In intelligent allocation, algorithms achieve three goals: balanced warehouse inventory, stable daily allocation volumes, and a significant increase in the in‑stock rate of small warehouses (FDCs) to over 90%.

2. Sales Forecasting Platform

JD aims to build a world‑class sales forecasting platform because accurate forecasts positively impact distribution, allocation, promotion, pricing, and inventory. The platform is designed to be simple and usable even for those unfamiliar with machine learning, while providing explainable results to help business users understand the predictions.

Leveraging massive data and advanced machine‑learning techniques, JD’s research team develops models that capture diverse data characteristics, ensuring reliable forecasts and close collaboration with business units.

3. End‑to‑End Intelligent Replenishment

The classic bullwhip effect arises when inaccurate forecasts lead to inventory mismatches. Shen’s philosophy is to shorten the decision chain: make decisions in one step, reducing personnel and improving precision. By bypassing separate forecasting and inventory steps, the algorithm directly outputs replenishment quantities, dramatically enhancing inventory performance and reducing parameter tuning. This end‑to‑end inventory management approach is a new research direction pursued by JD’s Silicon Valley team.