Demand Forecasting Practices in Alibaba Retail: From Mean Models to Deep Learning

In the new retail scenario, accurate demand forecasting is crucial for intelligent replenishment; the article shares Alibaba Retail's practical experiences, starting with simple mean models, progressing through time‑series techniques, and culminating in machine‑learning and deep‑learning solutions.

Mean models provide a baseline accuracy and a scale for demand, while time‑series models (ARMA, ARIMA, GARCH, Holt‑Winters) capture trends and noise but rely solely on historical sales, limiting performance.

Machine‑learning models treat demand prediction as a regression problem, emphasizing feature engineering such as sliding‑window averages, statistical descriptors, polynomial coefficients, and warehouse attributes, and using a stratified modeling approach to handle the wide range of SKU sales volumes.

Operational plans, including marketing promotions and supply‑chain actions, are transformed into features to improve forecast accuracy during events like Double‑11.

Deep‑learning attempts adopt a Wide & Deep architecture, encoding continuous time‑series features with CNN/RNN/attention mechanisms in the Deep part and sparse categorical features in the Wide part, achieving more stable results for certain SKUs.

To feed inventory models, the forecast's uncertainty is estimated either via analytical error propagation for linear models (ARIMA, regression) or by measuring historical prediction deviations, the latter being model‑agnostic and preferred in practice.

Overall, the article provides a concise roadmap of model evolution, practical insights, and references for further study.