Supply Chain Algorithms and Fresh Food Automatic Replenishment System at Hema

Introduction – The article shares practical experiences of supply‑chain algorithms deployed on the Hema platform, organized into three parts: Hema supply‑chain overview, algorithm positioning, and the fresh‑food automatic replenishment system.

1. Hema Supply‑Chain Overview

Hema operates a technology‑driven, consumer‑centric O2O model that integrates online and offline channels, delivering groceries within 30 minutes. Its core business includes Hema Fresh, X‑member stores, Hema Super Cloud, and Hema Neighborhood, with fresh food accounting for 60‑70% of sales.

The logistics‑inventory cost balance is crucial: bulk (whole‑item) logistics is cheaper than parcel logistics, while dispersed inventory raises stock‑out risk and waste. Hema’s supply‑chain network aims to share trunk routes and inventory across business lines to improve resource utilization.

2. Algorithm Positioning

Supply‑chain algorithms enhance traditional methods by leveraging data and AI. In e‑commerce, algorithms are divided into foundational (vision, speech, text) and business‑specific categories: front‑end (search, ads, recommendation), middle‑end (product, price, inventory prediction), and back‑end (fulfilment, warehousing, delivery optimization). The bull‑whip effect, where demand variability amplifies upstream, is mitigated by data‑driven forecasting.

Hema’s supply‑chain logic follows a closed loop: sales planning → supply capacity → inventory → fulfilment → sales adjustment, forming a complete commercial cycle.

3. Fresh‑Food Automatic Replenishment System

Background: Hema offers ultra‑fresh products with very short shelf‑life, including daily‑fresh items, demanding precise supply‑chain control.

System Modules: (1) Demand forecasting – high‑dimensional feature processing; (2) Inventory model – balancing user demand and holding cost; (3) Dynamic control – automatically generating promotions and traffic adjustments for items deviating from forecasts.

Achievements: Integration of spatio‑temporal heterogeneous graph neural networks improved forecast accuracy; the algorithm won top positions in Alibaba’s time‑series competition. Business impact includes >96% recommendation acceptance, 70% increase in order‑taker efficiency, 30% reduction in waste, and 25% reduction in stock‑outs. The solution was a finalist for the 2022 Franz Edelman Award.

Prediction Model Evolution: simple models → machine‑learning models → deep time‑series models → spatio‑temporal graph networks, each improving coverage, stability, and ability to capture item‑item and cross‑sample interactions.

Inventory Model: Handles single‑SKU stock and multi‑warehouse aggregation, addressing the “shared‑stock” issue where online orders may be pre‑empted by offline sales.

Dynamic Inventory Control: Monitors real‑time sales, updates forecasts, and triggers promotional or pricing actions to avoid over‑stock and waste, requiring joint traffic‑price optimization.

Q&A

Q1: Evaluation metric – weighted MAPE (accuracy = 1 – weighted error/total actual sales). Q2: Reference paper – STGAT (Huang et al., 2019) for spatio‑temporal interactions. Q3: Simple models used as a fallback (e.g., moving average, same‑week‑last‑year).

End of the sharing session.