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

The "Smart Canal, Smart Computing Future" 2025 AI Innovation and Entrepreneurship Conference in Hangzhou featured the AI+Smart Logistics and Supply Chain sharing session, where Professor Shen Zuojun, academician of the Hong Kong Engineering Academy and JD Retail’s chief supply‑chain scientist, together with Dr. Qi Yongzhi, technical director of JD’s supply‑chain algorithm team, co‑chaired and delivered the keynote "AI Technology Empowering Supply Chain Industry Upgrade," presenting JD Retail’s frontier research and practice in intelligent supply chains and analyzing core AI‑driven transformation pathways.

Global industrial chains are at a structural transformation node; breakthroughs require technological innovation. JD’s chief scientist stresses building an intelligent decision‑center through dynamic demand forecasting for precise resource matching, AI‑based risk perception for resilient response, and cross‑chain collaboration to break industry silos, forming a perception‑decision‑execution loop that redefines efficiency and enables networked intelligent collaboration.

Dr. Qi highlighted JD’s large‑scale commercial case: managing over 10 million self‑operated SKUs across more than 1,500 smart warehouses, achieving over 90% same‑day or next‑day delivery, powered by an intelligent supply‑chain system.

To tackle massive SKU management, JD built a four‑in‑one intelligent inventory platform (plan coordination, demand forecasting, smart decision, risk perception) whose core is accurate time‑series forecasting.

JD’s self‑developed time‑series forecasting large model reaches industry‑top performance on accuracy and multi‑dimensional feature fusion, supporting intelligent selection, dynamic allocation, and fulfillment optimization, creating a "predict‑decision‑execute‑feedback" enhanced learning loop that boosts turnover and response speed.

Traditional methods such as ARIMA, Prophet, LSTM, and TCN struggle with complex patterns; LLM‑based time‑series models (e.g., GPT‑4TS, TimesFM) remain nascent due to scarce high‑quality datasets and RLHF adaptation challenges.

JD’s algorithm team built the industry’s first billion‑scale pure time‑series model, introducing multi‑scale feature fusion, adaptive temporal attention, and a weak‑supervision pre‑training task tailored to supply‑chain scenarios, surpassing SOTA on public datasets, especially in zero‑shot cross‑domain prediction.

For training, they assembled a 1.5‑billion‑sample high‑quality dataset, devised time‑series split, data ratio, and synthetic data construction methods, proposed a generic PCTLM model with patch‑based processing and grouped time‑position encoding attention, and introduced a reinforcement‑learning‑from‑human‑feedback (RLHF) framework (TPO) for pure time‑series models, markedly improving prediction accuracy and generalization. The model is now deployed in JD’s supply‑chain system, delivering substantial accuracy gains.

Beyond demand forecasting, JD created data‑driven intelligent selection and allocation strategies: ML‑Top‑K for high‑potential product identification, Reverse‑Exclude for dynamic inventory structuring, and Hybrid Selection for multi‑objective balance, raising local order fulfillment rates by 2.19% and high‑value order share by 1.44%.

The end‑to‑end allocation algorithm integrates demand forecasts, multi‑objective optimization, and simulation, handling billions of variables in real time, reducing warehouse holding costs by tens of millions RMB annually, cutting allocation costs by over 100 million RMB, and improving inventory availability by 0.85%.

With ongoing breakthroughs in large models and multi‑agent collaboration, supply chains are shifting from linear chains to ecological networks, moving value focus from pure efficiency to global resource optimization. JD’s practice demonstrates AI’s commercial impact, reshaping business models and user experience.

JD continues to deepen collaboration with Tsinghua University, focusing on supply‑chain efficiency and cost optimization through intelligent prediction, dynamic scheduling, and multi‑level inventory coordination, aiming to open innovations to the industry and drive high‑quality economic development.

JD invites industry peers to co‑build intelligent supply‑chain infrastructure, contributing Chinese wisdom to global supply‑chain upgrades.