Smart inventory allocation for delightful customer experience

Mohak Kumar Sukhwani21 Feb 2022

Problem

An online fashion retailer buys stock continuously and must decide, for every purchase order, how many units of each SKU go to which warehouse. If stock sits far from the customers who buy it, orders ship from a distant warehouse, delivery is slower and logistics cost rises. The share of demand served by the nearest warehouse is called Regional Utilisation (RU). Warehouses have limited capacity and already hold old stock, so the demand-proportional split often cannot be executed as is.

Example from the article: a purchase order is a list of (SKU, quantity) pairs for M SKUs, and the system must return a quantity per SKU per warehouse, plus a "not assigned" quantity when the order exceeds total free capacity.

Solution in brief

A classifier predicts, for each SKU, the probability that a purchase comes from the region of each warehouse, which gives an ideal split. An integer program then moves as little stock as possible away from that ideal, paying a penalty per move, until warehouse capacities and existing inventory are respected.

Questions this case answers
  • How to distribute a fixed batch of items across capacity-limited locations so that each item sits close to where it will be demanded, when the ideal demand-proportional split is infeasible
  • How to split a purchase order across warehouses in proportion to regional demand?
  • How to keep an allocation feasible when warehouse capacity and existing stock limit the ideal split?
  • How to raise the share of orders shipped from the nearest warehouse without a manual allocation policy?

Tools and methods

  • Split prediction classifier — predicts warehouse probabilities per SKU from style attributes
  • Ideal split matrix — order quantity times predicted probabilities, assuming unlimited capacity
  • Redistribution penalty matrix — cost of serving demand from a warehouse other than the nearest
  • Integer programming — finds the feasible allocation with minimum total penalty
Continue reading — it's free

Google or a one-time email code. No password.