The rule of thumb says folding four warehouses into one halves your safety stock [3]. The band that reaches the finance desk says 15% to 30% less inventory [4]. A peer-reviewed test on a real factory network beat the company’s own stock rule by 6.6% [1]. The gap between those numbers is not effort. It is the shape of the network, and the rule only holds when yours fits it.

The big picture:

Multi-echelon planning sets stock levels for a whole network at once. Plant, main hub and local sites get planned as one system, not four [4]. Each site’s repeated buffer folds into one shared pool, and one pool is smaller than the parts it replaces [3].

The size of that gain comes from the square-root law. Fold four equal sites into one, and total safety stock is said to drop by half [3].

The law is neat and widely quoted. It also needs four things that real networks rarely have [3].

  • Demand at each site must be independent.
  • The sites must be the same size.
  • Demand must sit on a normal curve.
  • The main lead time must not grow when sites merge.

A peer-reviewed study ran a planning model on three network shapes. The gain moved with the shape. It beat its baseline by 16.4% on a straight chain, 11.3% on a split network, and 6.6% on a real factory network [1].

The factory case is the one I keep reading. The baseline it beat was the company’s own rule.

The firm promised customers a 98% fill rate, meaning the share of order lines it could ship from stock. It then put the same 98% on every internal link between its paper mills and its box plants [1].

What strikes me is that the 98% is not a customer promise there. It is the customer promise copied onto internal links. That copy rebuilds the double buffer the network was meant to remove.

Service targets add to the bill. A 95% fill rate needs a stock multiplier of 1.645. A 99% target needs 2.326 [2].

That is 41.4% more buffer for the last four points. The 41.4% is my arithmetic on the two table values, not a measured saving.

By the numbers

  • 50% to 0% — Overlap penalty: Four equal sites cut safety stock 50% at zero correlation, 31% at 0.3, 16% at 0.6 and nothing at 1.0 [3]. The table is my arithmetic on the pooled-variance rule, not a source result.
  • 6.6% — Real-network gain: A planning model beat a real factory’s own 98%-on-every-link rule by 6.6%. The same model beat a textbook chain by 16.4% [1].
  • 1.645 to 2.326 — Service multiplier: Moving a fill-rate target from 95% to 99% raises the stock multiplier by 41.4% [2].
  • 15% to 30% — Quoted vendor band: The cut vendors and guides cite, with service held above 98% [4][5]. It is a claim, not a measurement.

What I’d watch:

The people closest to this choice are not arguing about the band. They are arguing about which number goes into the model.

  • Shared spread, not site count: Combined demand depends on how closely your sites move together. At four sites, a correlation (a shared swing) of 0.3 turns a 50% cut into 31%. A correlation of 0.6 turns it into 16%. Measure that overlap on two years of real shipments before the case is signed.
  • Where the 98% lives: A fill rate promised to the customer belongs at the final node. Once it is copied onto every internal handoff, each link carries its own buffer [1]. I would want to read the policy before the savings estimate.
  • The lead-time leg: Pooling moves stock further from the buyer. If the main-to-region lane adds transit days, the shared buffer has to cover that gap [3]. The lane time belongs in the model beside the demand spread.
  • Price of perfection: A 99% target at every node needs 41.4% more stock than 95% [2]. On a network where pooling returned 16%, that one setting can erase the gain. I would want both numbers on the same page.

The catch

My read on the 6.6% is that it is not a measurement of the pooling law. It is one planning model against three baselines on three networks of different size and shape [1]. Treat it as a sign that the gain thins as a network gets real, not as the saving your team will book.

The four limits behind the law come from undated warehouse notes, not a study [3]. The math checks out. Nobody measured how often real networks break those rules.

Overlap is also hard to estimate with care. Two years of shipment data may miss a shared sale or a single national buyer that makes four regions move together. A low reading can be a quirk of the sample, and that is the first flaw a finance reviewer will spot.

Pooling also shifts risk rather than removing it. One shared buffer in one building is cheaper to hold.

It is also a single point of failure that every region now depends on. The vendor pages say little about that trade [4][5].

Our own field note points the same way, and it is one case rather than proof. A medical device seller pooled six local buffers into one.

It freed $14.2 million and held 99.4% service. That is one company’s result, not a benchmark.

At a glance

  • The Big Shift: Multi-echelon planning promises 15% to 30% less inventory. The math behind it halves safety stock when four equal, independent warehouses become one [3][4]. Real networks rarely meet those terms.
  • Why It Matters: The saving is the working capital an operations leader is asked to release. A peer-reviewed test on a real factory network beat its own 98%-on-every-link rule by 6.6%, not 30% [1].
  • What I’d watch: I am watching which numbers feed the model before the target is signed.
  • Shared spread: The spread of combined demand, which depends on how closely the sites move together, not on how many sites there are [3].
  • Fill-rate rule: Where the 98% promise is written, because copying it onto internal links rebuilds the repeat buffers [1].
  • Service multiplier: The factor a target applies to the buffer, 1.645 at 95% and 2.326 at 99% [2].
  • The Catch: The 6.6% comes from one study across three network shapes, and the overlap estimate is easy to get wrong [1][3]. Pooling also leaves one building holding the risk for every region.

Related reading

Sources

[1] Geevers, van Hezewijk and Mes, “Multi-echelon inventory optimization using deep reinforcement learning”, Central European Journal of Operations Research 32:653-683, published 19 July 2023 — 16.4% / 11.3% / 6.6% improvement by network shape, and the CardBoard Company benchmark that enforced a 98% fill rate on every internal connection. https://link.springer.com/article/10.1007/s10100-023-00872-2 [2] NIST/SEMATECH e-Handbook of Statistical Methods, “Cumulative Distribution Function of the Standard Normal Distribution” — critical values 1.645 at p=0.950 and 2.326 at p=0.990. https://www.itl.nist.gov/div898/handbook/eda/section3/eda3671.htm [3] Inventory Path, “The Square-Root Law of Inventory Pooling Is Often Wrong” — the four-warehouse halving, its four assumptions, and the note that correlation, unequal sites and longer central lead times erode the benefit. https://www.inventorypath.com/the-square-root-law-of-inventory-pooling-is-wrong-more-often-than-its-right [4] ToolsGroup, “Multi-Echelon Inventory Optimization: Benefits & Best Practices” — the 15% to 30% typical reduction and service targets above 98%. https://www.toolsgroup.com/blog/multi-echelon-inventory-optimization-toolsgroup-guide [5] Re:innovation, “Multi-Echelon Inventory Optimization: The Complete Guide to Multi-Tier Networks, Safety Stock & Service Level Targets” — the 15-30% range as reported by multiple industry sources, and the z-score curve behind near-perfect targets. https://www.reinnovation.eu/post/multi-echelon-inventory-optimization-the-complete-guide-to-multi-tier-networks-safety-stock-serv