90% to 95% service level in retail replenishment

My role: Lead data scientist; led a team of four data scientists (three junior and one senior).

Executive Context#

A US multi-brand fashion retailer planned inventory across 543 stores, four distribution centres, and three channels: brick-and-mortar, D2C, and e-commerce. The planning scope covered 28,000 base EANs and 84,000 SKUs.

Buying decisions had to account for 9–12-month lead times. By the time stock arrived, promotion plans and channel demand could have changed. The operating challenge was to maintain service while reducing inventory tied up in the wrong quantities or locations.

Diagnostic Approach#

• AI Necessity: Long-lead-time commitments required decisions under uncertainty. The problem included when to order, how much to buy, which promotion price to use, and how to split stock across DCs.
• Data Sufficiency: The planning unit mattered: 28,000 base EANs became 84,000 SKUs across three channels. Aggregating that into one demand number would hide the allocation decisions.
• Methodological Fit: Price, quantity, timing, and DC allocation had different decision structures. Each needed an appropriate optimization method.
• Infrastructure Readiness: The recommendations had to support buying 9–12 months ahead and allocation across four DCs serving stores and online channels.

Strategic Intervention#

We separated the replenishment problem into four decisions rather than asking one forecast to solve the entire planning workflow.

Decision	Method	Planning purpose
Promotion price	Bayesian optimization	Choose a promotional price
Order quantity	CMA-ES	Optimize how much to order
Long-order timing	Q-learning	Plan when to commit a long-lead-time order
Distribution-centre split	Dynamic programming	Allocate inventory across DCs

The forecast was part of the input. The value came from connecting it to purchasing, promotion, and allocation decisions at the required SKU and channel scope.

Outcome#

Measure	Before	After	Context
Service level	90%	95%	Improved availability alongside inventory planning
Net inventory saved	Baseline	Approximately $150K/month	Reported monthly inventory result
Planning scope	—	543 stores, 4 DCs, 3 channels, 84K SKUs	US multi-brand fashion retail
Buying horizon	—	9–12 months	Long-lead-time replenishment

Strategic Takeaway#

A demand forecast becomes useful when it informs the commitment. Here, the commitment included promotion price, order quantity, order timing, and DC allocation. Matching the method to each decision improved service while reducing inventory pressure.