How AI is Revolutionizing Supply Chains: JD.com’s Billion‑Parameter Time‑Series Model

During the "AI+Smart Logistics and Supply Chain" session at the 2025 AI Innovation & Entrepreneurship Conference in Hangzhou, Professor Shen Zuojun, a Hong Kong Academy of Engineering academician and JD Retail Supply Chain Chief Scientist, together with Dr. Qi Yongzhi, Technical Director of JD’s supply‑chain algorithm team, co‑chaired the event and delivered the keynote "AI Technology Empowering Supply Chain Industry Upgrade".

The speakers emphasized that the global supply chain is at a pivotal structural transformation point, requiring technological innovation to break traditional constraints. JD’s intelligent decision hub integrates dynamic demand forecasting, AI‑driven risk perception, and cross‑chain collaboration to form a perception‑decision‑execution loop, redefining efficiency boundaries and shifting the ecosystem from linear cooperation to networked intelligent coordination.

JD Retail manages over 10 million self‑operated SKUs across more than 1 500 smart warehouses, achieving over 90% same‑day or next‑day delivery for self‑operated orders. This performance is underpinned by a four‑in‑one intelligent inventory management platform—planning coordination, demand forecasting, intelligent decision, and risk perception—centered on precise time‑series forecasting.

The company’s proprietary time‑series large model reaches industry‑leading accuracy, supporting intelligent product selection, dynamic allocation, and fulfillment optimization through a "predict‑decision‑execute‑feedback" enhanced learning loop.

Traditional methods such as ARIMA, Prophet, LSTM, and TCN struggle with complex patterns and zero‑shot generalization. While large language models adapted for time series (e.g., GPT‑4TS, TimesFM) show promise, they face data scarcity and RLHF adaptation challenges.

JD’s algorithm team addressed these issues by building the industry’s first billion‑parameter pure time‑series model, incorporating multi‑scale feature fusion, adaptive temporal attention, and weak‑supervised pre‑training tailored to supply‑chain scenarios. The model surpasses state‑of‑the‑art benchmarks, especially in zero‑shot cross‑domain prediction.

To support the model, the team assembled a 1.5‑billion‑sample high‑quality dataset, employing time‑series splitting, data ratio balancing, and synthetic data generation. They introduced the PCTLM architecture, which processes data in patches with enhanced projection and grouped time‑position encoding attention. Additionally, a reinforcement‑learning‑from‑human‑feedback (RLHF) framework named TPO was developed for pure time‑series models, markedly improving prediction accuracy and generalization.

Beyond demand forecasting, JD implemented data‑driven inventory selection and allocation strategies. Algorithms such as ML‑Top‑K, Reverse‑Exclude, and Hybrid Selection enable precise high‑potential product identification and dynamic inventory structuring, raising local order fulfillment rates by 2.19% and increasing high‑value order share by 1.44%.

The end‑to‑end allocation algorithm fuses demand forecasts, multi‑objective optimization, and simulation to produce real‑time, billion‑variable solutions that lower inventory holding costs by tens of millions of yuan annually, cut allocation expenses by over a hundred million yuan, and improve on‑hand inventory rates by 0.85%.

Looking ahead, JD anticipates a shift from traditional chain structures to ecosystem networks, driven by large models and multi‑agent collaboration. In partnership with Tsinghua University, JD will continue to innovate in intelligent prediction, dynamic scheduling, and multi‑level inventory coordination, aiming to open its optimization outcomes to the industry and reduce logistics cost ratios, thereby fueling high‑quality development of the real economy.