Alibaba's Smart Supply‑Chain Forecasting: Scenarios, Algorithm R&D, and Application Cases

Guest speaker Wang Tong, a senior algorithm expert at Alibaba, shares Alibaba's research on intelligent supply‑chain forecasting, emphasizing that uncertainty is the core challenge and that prediction serves as the first line of defense.

The discussion is organized into three themes: (1) forecasting scenarios and their characteristics, (2) the R&D roadmap and results of forecasting algorithms, and (3) concrete application cases.

1. Forecasting scenarios are divided by prediction horizon: long‑term GMV forecasts (up to one year), medium‑term sales forecasts (weeks to months), and short‑term warehouse order forecasts (days). Each horizon requires different algorithms and data granularity, and both offline (day‑level) and real‑time (hour‑level) solutions exist.

2. Algorithm evolution spans three generations. The first generation uses traditional statistical time‑series methods, offering high stability but low accuracy and adjustability. The second generation adopts machine‑learning models such as XGBoost/LGB, improving accuracy but demanding extensive feature engineering and offering limited adjustability. The third generation applies deep‑learning architectures (CNN, LSTM) which increase adjustability but suffer from stability issues and high data‑volume requirements.

To combine the strengths of all three generations, Alibaba built a proprietary model called Falcon . Falcon decomposes a time series into low‑frequency trend, medium‑frequency seasonality, and high‑frequency pulse components, using specialized blocks for each. It keeps the parameter count low (a few thousand) to avoid over‑fitting and enables fast training on millions of records.

Falcon was evaluated in internal and external competitions (Kaggle M5, Tianchi AI Earth, power‑load forecasting) and consistently ranked first, demonstrating its general‑purpose capability across e‑commerce and non‑e‑commerce domains.

3. Application cases include the 2018 Double 11 shopping festival and Taobao live‑streaming. For Double 11, forecasting enabled pre‑packing of hot‑selling items and strategic distribution of inventory to reduce peak‑time labor costs. In live‑streaming, a customized forecasting pipeline accounted for host‑product matching and promotional effects.

The article also discusses specific challenges in e‑commerce supply‑chain forecasting, such as handling product substitution relationships, seasonal and regional demand variations, and predicting demand for out‑of‑stock items.

Overall, the talk illustrates how Alibaba abstracts supply‑chain forecasting as a time‑series problem, iterates through statistical, machine‑learning, and deep‑learning solutions, and finally delivers a lightweight, high‑accuracy model (Falcon) that addresses the unique constraints of large‑scale e‑commerce operations.