Pooling means bringing together in one place the stock that used to protect several separate points. Its foundation is statistical: when independent demands are added, their deviations partly offset each other, so the variability of the whole grows more slowly than the sum of the variabilities. A single shared buffer therefore protects better than the sum of local buffers, for markedly less capital.
Why it matters
Almost every lever in this domain is a trade-off: you accept less service for less stock, or you pay capital for availability. Pooling is the exception. At identical service and identical demand, grouping stock reduces the capital required, with no direct operational counterpart. That is what makes it the most powerful lever in the network cluster.
The reasoning is one of offsetting. Ten sites each holding their own buffer each protect against their own peak, yet those peaks do not occur simultaneously. When one site consumes more than expected, another consumes less. Grouping stock lets these deviations partly cancel, and it then suffices to protect against the variability of the whole, far smaller than the sum of the variabilities.
The signature quantifies this effect, and the result often surprises. The buffer needed does not grow in proportion to the number of sites but as its square root. Going from four warehouses to one halves the buffer; going from sixteen to one divides it by four. The gain is therefore considerable at first, then flattens, which indicates that full centralisation is not necessary to capture most of the benefit.
This square-root shape has an important practical consequence. An organisation with two or three sites has little to gain from merging, while a highly fragmented network holds a major opportunity. Before launching a logistics reorganisation, it is worth calculating where you sit on that curve.
Two major caveats frame this result, and ignoring them leads to promises that cannot be kept. The first is the independence of demands. The square root assumes sites do not react to the same causes; yet a national promotion, a season or an economic shock move them together. The more correlated the demands, the more the gain melts, and it vanishes entirely if they move perfectly in unison.
The second is delivery lead time. Centralising moves stock away from the customer, which lengthens the final lead time or forces express transport whose cost can wipe out the capital saving. The calculation must therefore always weigh stock saved against added distribution cost, and that balance depends heavily on unit value.
The expert lesson is that pooling does not necessarily require a move. Shared visibility between sites, with the ability to transfer a part to where it is missing, captures a large share of the statistical benefit without touching physical locations. This virtual pooling is often faster to implement, less risky, and it preserves customer proximity.
The mechanism
The starting point is an elementary property of independent random variables: their means add, but it is their variances, not their standard deviations, that add. Since the buffer depends on the standard deviation, it follows the square root of the sum of variances, not the sum of standard deviations.
For sites of comparable size this gives directly the law visible in the signature: the pooled buffer equals a single site’s buffer times the square root of the number of sites, whereas the decentralised total equals that same buffer times the number of sites. The ratio between the two is therefore the inverse of the square root of the number of sites, whatever service level is chosen.
When demands are correlated, the formula corrects and the gain shrinks. It remains substantial as long as correlation stays moderate, which is the usual case across regions or across customers in one market, but a strong shared seasonality reduces it sharply. Measuring the real correlation between sites is therefore a prerequisite, not a refinement.
Finally, pooling is not only about places. The same mechanism operates when close references are merged into one, when a component used by several products is standardised, or when the differentiation of a generic item is postponed. These are three forms of the same idea, and the neighbouring page on postponement explores the richest variant.
The traps
The gain promised by the square root assumes sites that do not move together. A national promotion, a shared season or a market shock correlate demands and sharply reduce the benefit. Without measuring that correlation first, the calculation promises a gain that will not materialise.
Centralising moves stock away from the customer. The capital saved can be entirely absorbed by longer lead times, express transport or lost proximity sales. The assessment must be made on full cost, and it tips differently depending on unit value.
Many projects stall at the cost of physical centralisation when most of the statistical gain comes from shared visibility and inter-site transfers. Abandoning pooling because merging warehouses is impossible confuses the means with the effect.
The rollout
Count the stocking points holding the same references and estimate the theoretical gain. A two-site network has little to gain, a twelve-site network a great deal.
Compare demand histories across sites before announcing a figure. Strong correlation reduces the theoretical gain to a fraction of itself.
Weigh the capital saved against added transport cost and longer customer lead times. The result depends heavily on unit value and on the commercial promise.
Install inter-site visibility and allow transfers before contemplating any physical reorganisation. The gain arrives faster, at far lower risk.
Look for standardisable components and near-identical references: merging two close items produces exactly the same statistical effect as merging two sites.
Neighboring concepts
From knowledge to action
A fragmented network funds the same protection several times over. Our Inventory & distribution file measures the real correlation between your sites and quantifies the gain available without relocating.