Choosing where an order ships from

1The problem

Once a merchant holds stock in two places, a warehouse and a shop, or two warehouses in different regions, every order raises a question: where does it ship from? The usual answer is the location that can fill the order most cheaply right now. That rule is easy to explain and easy to build, and it is often wrong over a day of orders, because it uses up stock that a later order will need from that same place.

Large retailers solve this with optimization models. A merchant with two to five locations needs rules that capture most of the benefit without an optimizer. This paper sets out three.

2A worked example

Two locations hold stock. Location A holds one unit of product X and one of product Y. Location B holds two units of X. Two orders arrive in sequence: order 1 wants one X, and order 2 wants one X and one Y. Shipping one parcel costs 4 from A and 5 from B for order 1, and 6 from A and 7 from B for each parcel of order 2.

Rule	Order 1	Order 2	Total cost	Parcels
Myopic: lowest-cost location now	X from A, cost 4	Y from A, cost 6, plus X from B, cost 7	17	3
Look one ahead	X from B, cost 5	X and Y from A, cost 6	11	2

Source: our constructed example.

The myopic rule ships order 1 from A because 4 is less than 5. That uses A's only X, so order 2 can no longer ship whole from A and is split into two parcels. Paying one unit more on order 1 saves seven on order 2 and one parcel for the customer.

3Three rules

3.1Ship whole when possible

A split order costs more than one extra label: it means two deliveries for the customer, two tracking numbers and two chances of a problem. The first rule is to prefer any location that can ship the whole order over a cheaper combination that splits it, unless the saving is larger than a set split penalty. Research on multi-item ecommerce fulfillment treats the cost of splitting as central to the problem and shows that assignments derived from a linear program, rounded with care, keep the expected cost within a proven bound of the optimum¹.

3.2Look one order ahead

The example turns on one fact: X is plentiful at B but scarce at A, and Y exists only at A. Stock that is the last unit of its kind at a location, or the only stock that lets a common order ship whole, is worth more than its shipping cost suggests. The second rule charges a penalty for using such stock, so the choice for order 1 compares 4 plus the penalty at A with 5 at B.

This is a simple form of a well-studied idea. A heuristic for fulfillment decisions made on the fly uses dual values from a linear program to estimate what each unit of stock at each location is worth to future orders, and adds that to the immediate cost; in tests on data from a large online retailer it reduced outbound shipping cost compared with a myopic rule². A small merchant can approximate the dual values by hand: a penalty for the last unit of a product at a location, and a larger one for a product that exists at only one location.

3.3Reconsider once before picking

An order is assigned when it arrives, but it is not picked at once. Between the two, more orders arrive, and some of them would have changed the first decision. Research on real-time fulfillment found clear benefits from periodically reevaluating assignments that had not yet been released to the warehouse, mainly by reducing the number of shipments³. The third rule reruns the first two once, for every order not yet picked, just before each pick wave.

4Inventory has to be right

All three rules assume the system knows what each location holds. When it does not, the rules make confident mistakes. Accurate availability information also changes what shoppers do: when a retailer began showing reliable store inventory online, alongside an option to buy online and pick up in store, sales shifted between its online and store channels⁴. A store that shows availability per location should treat that number as part of the product.

Allocation also affects replenishment. If a location's reorder is driven by what it shipped, a location that was used to rescue other locations' orders will look busier than its own demand, and its reorders will grow. Lee, Padmanabhan and Whang showed how distortions of this kind in demand signals amplify as they move up a supply chain⁵. Replenishment should be driven by the demand each location serves, not by its shipments.

5Limitations

The example is constructed to show the mechanism and does not estimate the size of the effect in practice. The penalties in the second rule need tuning to a merchant's costs and order mix, which we do not give. We consider shipping cost and split orders only; delivery promises, carrier capacity and location labor limits can override them. The cited studies use data from large retailers, and the gains at two to five locations are not established.

6Conclusion

Choosing the lowest-cost location for each order in isolation can cost more over a day of orders. Shipping whole when possible, charging a penalty for scarce stock, and reconsidering assignments once before picking capture much of what optimization models offer, provided inventory is accurate and replenishment follows demand rather than shipments.