TP_AI Team’s Winning Solution for JD Global Operations Optimization Challenge – Intelligent Inventory Management

The article shares the champion TP_AI team's analysis of the JD Global Operations Optimization Challenge – Intelligent Inventory Management, presented by Min Xu from Tsinghua University and Ma Siyuan from Peking University AI Innovation Center.

With rising e‑commerce demand and cost pressures, accurate regional demand prediction and efficient inventory allocation are critical; JD operates hundreds of warehouses with strict delivery standards (e.g., orders before 11 am must be delivered the same day).

Task 1 – Regional Sales Forecasting: Participants must predict the next month’s sales for each product in each region using two years of historical sales, product attributes, and promotion data. Evaluation uses a quantile‑loss function to assess the entire sales distribution, not just the mean.

TP_AI’s approach focuses on quantile prediction: compute historical quantiles as baseline forecasts, give higher weight to recent data, and discard distant observations. They found limited benefit from product/category/discount features due to data sparsity and privacy‑preserving anonymization.

Task 2 – Warehouse Network Inventory Allocation: JD’s two‑level inventory system (Regional Distribution Centers and Front Distribution Centers) imposes constraints on SKU count and total volume for daily allocations, simulated over 30 days.

The team first calculates a “target allocation” by keeping safety stock at RDCs and distributing the remainder to FDCs to meet expected demand multiples. Under SKU and volume limits, they greedily prioritize allocations that yield the highest expected gain, selecting the top 200 FDCs and then the top 3000 SKUs.

Final Round: The solution combines both tasks by predicting the 7‑day sales sum quantile, subtracting inbound stock and current inventory to obtain the optimal allocation. A non‑parametric sampling method estimates the quantile, and the authors note that incorporating EOQ concepts could improve performance when ordering costs are non‑negligible.

Overall, the TP_AI team emphasizes simple, explainable algorithms, careful handling of long‑tail demand, and the importance of aligning model design with real‑world operational constraints.