Network inventory optimisation: bu ne anlama geliyor
Set safety stock correctly at every location and the network still holds too much. Each site is right on its own terms, and the sum is wrong.
The reason is that locations are not independent. Stock at one covers demand at another, either formally through transfers or informally because customers shift. Optimising them separately prices that coverage at zero.
What single-location optimisation cannot see
Pooling. Two locations each holding for their own variability hold more together than one location covering both. The gain is real and it does not appear in either location computed alone.
Transfers. If stock can move between sites in two days, the second site is partial coverage for the first. Safety stock set as though it cannot move is safety stock paid for twice.
Correlated demand. Locations whose demand rises together need more cover than the arithmetic suggests; locations whose demand offsets need less. Correlation is a network property and is invisible from inside one node.
Why the echelon question comes first
Before optimising, decide where inventory is allowed to sit. A network holding at central, regional and store level has three places to be wrong, and the levels interact: central stock that cannot reach a store within its cover window is not cover, it is cost.
Most network projects begin by tuning quantities at existing levels. The larger number is usually in the structure — how many levels, and which SKUs are held at which of them.
Segmentation does most of the work
Not every SKU deserves a network answer. High-volume, stable items are close to optimal under simple rules. The gain sits in the middle: intermittent demand, high value, or long lead time, where pooling matters most and local rules are worst.
A network project that treats all SKUs alike spends its effort where there is nothing to gain and dilutes the case for itself.
The organisational obstacle
Network optimisation moves stock away from locations whose managers are measured on local service. It is correct at the network level and reads as a downgrade locally.
This is not a modelling problem and it does not have a modelling solution. If local availability metrics stay unchanged while the network reallocates, the reallocation will be quietly reversed — through expediting, through safety stock overrides, through orders placed outside the system.
What makes the result durable
Agreement, before implementation, on what each location is accountable for after the change. If a store no longer holds cover because a regional site holds it, the store’s service metric has to reflect that the cover exists elsewhere.
Otherwise the model is right, the network is better, and within two quarters the inventory is back where it started.