Supply‑Demand Coordination in E‑commerce: Challenges and Algorithmic Solutions

In the Internet era, leveraging data and algorithms to optimize the coordination between the supply side and the sales side of e‑commerce is a core problem for reducing costs and maintaining competitive advantage.

1. Background

The "Yanxuan" model offers a wider variety of products than ordinary brand e‑commerce, longer replenishment cycles (1‑4 months), and multiple sales channels (self‑built app, platform stores, live streaming, offline stores), making supply‑demand coordination especially challenging.

2. What is Supply‑Demand Coordination?

It is the practice of balancing inventory turnover (库转) and stock‑out (缺货) by aligning the replenishment chain ("采") with the sales chain ("销"). The ideal state is when the quantity replenished matches the actual sales, eliminating both inventory cost and lost sales.

3. Three Stages of Coordination

3.1 Pre‑stage (Prediction) – Create activity plans and demand plans based on forecasts. Activity plans are iteratively refined using profit‑margin calculations; demand plans are derived by forecasting SKU‑level sales.

3.2 In‑stage (Strategy) – Because forecasts are probabilistic, replenishment strategies must be optimized under constraints such as probability distributions, MOQ, lead time, and safety stock, targeting specific business goals (e.g., avoid stock‑out during promotions).

3.3 Post‑stage (Adjustment) – After sales, micro‑adjust traffic and, if necessary, perform clearance operations when inventory accumulates.

4. Difficulties in Each Stage

Prediction challenges : massive SKU count (~160k, 2.2k self‑operated), long production cycles (1‑6 months), and short sales history for many items.

Replenishment challenges : balancing inventory turnover and stock‑out rates, and the separation of activity and demand planning across different teams.

Sales‑adjustment challenges : SKU‑level activity planning limited to 1‑2 weeks ahead, multiple sales channels, and external factors such as pandemics or special events.

5. Algorithmic Solutions

5.1 Sales Forecasting – Features include potential users, reach, conversion rate, and external factors. Model evolution: time‑series (ARIMA), tree‑based models (XGBoost/GBDT), and deep learning (DeepTCN) that output probability distributions. Tree models are often preferred for their accuracy and interpretability.

5.2 Replenishment Model – Three versions: (1) safety‑stock based replenishment, (2) SKU‑level expected cost minimization (considering over‑stock and under‑stock costs), and (3) multi‑SKU shared MOQ constraint minimization.

5.3 Traffic Control – Integrates forecasting, automatic control, and optimization. Uses WaveNet for traffic prediction and PID control for demand flow. Model evolution includes: (1) WaveNet + PID, (2) PV‑value minimization with CTR constraints, (3) low‑efficiency item slot replacement for revenue maximization.

6. Summary

Combine algorithmic models with expert knowledge for robust forecasting.

Use a shared prediction engine for both activity and demand planning.

Avoid over‑optimizing forecast accuracy; focus on business‑level objectives like cost reduction.

Post‑adjustment via traffic control remains essential due to inevitable variability.