Multi-Echelon Retail Inventory

DC ↔ store network · Pooling

A retail network of \(N\) stores fed by one distribution centre can pool safety stock at the upstream node, exploiting the statistical cancellation of independent local demand shocks. Total inventory falls roughly as \(1 / \sqrt{N}\) (Eppen 1979) while service is preserved. Foundational papers: Clark & Scarf (1960) base-stock decomposition; Eppen (1979) square-root law; Federgruen & Zipkin (1984); Fisher & Raman (2010) on retail accurate response.

Why it matters

Square-root pooling · documented retail savings

√N
Eppen (1979) square-root law: pooling \(N\) iid stores reduces total safety stock by factor \(\sqrt{N}\) — the foundational pooling benefit.
Source: Eppen (1979), Management Science 25(5).
30–50%
Documented safety-stock reduction when retailers consolidate from store-only to DC + store inventory pooling (industry benchmarks).
Source: Axsäter (2015), Inventory Control, Springer.
−50%
Fisher & Raman (2010): accurate response cuts pre-season buy by half via late-season replenishment from a pooled DC.
Fisher & Raman (2010), The New Science of Retailing, HBR Press.
N × 1
Clark & Scarf (1960) decomposed an \(N\)-echelon DP into \(N\) single-echelon newsvendors via echelon-stock variables — tractable optimum.
Clark & Scarf (1960), Management Science 6(4).

Where the decision sits

Brick-and-mortar chains · omnichannel networks · replenishable assortments

Multi-echelon allocation is the dominant lever whenever a retailer holds inventory at two or more levels — a DC and \(N\) downstream stores, or a national hub and regional warehouses. Holding all safety stock at stores is wasteful (each store buffers its own variability); holding it all at the DC is service-blind (no on-hand at the customer touchpoint). The interesting policy lives between: stores carry just enough to cover replenishment lead time from the DC, while the DC holds the pooled buffer that catches network-wide demand shocks. Cross-link upstream to single-store replenishment and the newsvendor.

Forecastpooled demand
Set echelon stocksDC \(S_0\), stores \(S_i\)
Allocate echelon stockDC → stores
Replenishstore from DC

Problem & formulation

Two-echelon base stock with Clark-Scarf decomposition

OR family
Multi-Echelon Inventory
Decomposition
\(N\) echelon-newsvendors
Solver realism
★★★ Closed form
Reference
Eppen (1979), Clark & Scarf (1960)

Sets and indices

SymbolMeaningDomain
\(i \in \{1, \ldots, N\}\)Downstream storediscrete
\(0\)Upstream distribution centre (DC)singleton

Parameters

SymbolMeaningUnit
\(D_i \sim \mathcal{N}(\mu_i, \sigma_i^2)\)Per-period demand at store \(i\), possibly correlatedunits / period
\(\rho\)Pairwise demand correlation between stores (uniform)\(\in [-1, 1]\)
\(L_{\text{store}}\)Lead time DC \(\to\) storeperiods
\(L_{\text{DC}}\)Lead time supplier \(\to\) DCperiods
\(\alpha_i\)Permitted stock-out probability at store \(i\)\(\in (0, 1)\)
\(z = \Phi^{-1}(1 - \alpha)\)Service-level safety factor—

Decision variables

SymbolMeaningDomain
\(S_0\)Echelon base-stock at the DC\(\geq 0\)
\(S_i\)Local base-stock at store \(i\)\(\geq 0\)

Pooled demand at the DC

Aggregated across \(N\) stores, total per-period demand is normal with mean \(\sum \mu_i\) and variance accounting for correlation:

$$D_{\text{total}} \;=\; \sum_{i=1}^{N} D_i \;\sim\; \mathcal{N}\!\left(\sum_{i=1}^{N} \mu_i,\; \sum_{i=1}^{N} \sigma_i^2 + 2\sum_{i < j} \mathrm{Cov}(D_i, D_j)\right)$$

Eppen's square-root law

For iid stores \(D_i \sim \mathcal{N}(\mu, \sigma^2)\) with no correlation, the safety-stock comparison is closed-form:

$$\underbrace{N \cdot z\,\sigma}_{\text{stores only}} \;\;\longrightarrow\;\; \underbrace{z\,\sigma\,\sqrt{N}}_{\text{DC pool}} \qquad \Rightarrow \qquad \frac{\text{pooled}}{\text{stores only}} \;=\; \frac{1}{\sqrt{N}}$$

Doubling the network (\(N \to 2N\)) cuts pooled safety stock by a further \(1/\sqrt{2} \approx 29\%\). Eppen (1979).

Correlated case

With uniform pairwise correlation \(\rho\), the pooled standard deviation inflates:

$$\sigma_{\text{total}} \;=\; \sigma\,\sqrt{N + N(N-1)\rho} \;=\; \sigma\,\sqrt{N}\,\sqrt{1 + (N-1)\rho}$$

Perfect correlation (\(\rho = 1\)): pooling gives no benefit. Independence (\(\rho = 0\)): full \(\sqrt{N}\) reduction. Negative correlation: pooling does even better.

Clark-Scarf echelon decomposition

Define the echelon stock at node \(i\) as local on-hand plus all downstream on-hand and in-transit. Clark & Scarf (1960) showed the \(N\)-echelon problem decomposes into \(N\) single-echelon newsvendors solved sequentially:

$$S_i^{\ast} \;=\; F_i^{-1}\!\left(\frac{c_u^{(i)}}{c_u^{(i)} + c_o^{(i)}}\right) \qquad \text{for } i = N, N-1, \ldots, 0$$

Each level uses an effective overage / underage cost rolled up from the downstream solution. Reduces an exponential-state DP to \(N\) one-dimensional optimisations.

Service-level constraint

$$\Pr\bigl(\text{stock-out at store } i\bigr) \;\leq\; \alpha_i \qquad \forall\, i$$

Interactive solver

Compare stores-only, DC-pooled, and hybrid policies under correlation

Pooling policy comparator
iid stores · uniform correlation \(\rho\) · closed-form safety stock
★★★ Closed form
Number of downstream stores
Units / period
Demand volatility
Pairwise ≬ stores
\(z = \Phi^{-1}(1-\alpha)\)
Periods
Periods
—
Stores-only safety stock
—
DC-pooled safety stock
—
Hybrid (DC + stores)
—
DC-pool savings vs stores
—
Hybrid savings vs stores
—
Safety factor \(z\)
Stores only (linear in \(N\)) DC pool (\(\sqrt{N}\) law) Hybrid (DC + stores) Savings % vs \(N\)

Under the hood

Three closed-form policies are evaluated under the iid-stores assumption with uniform pairwise correlation \(\rho\). Stores only: each store carries \(z\,\sigma\,\sqrt{L_{\text{store}} + L_{\text{DC}}}\), summing to \(N\,z\,\sigma\,\sqrt{L_{\text{store}} + L_{\text{DC}}}\). DC pool: all safety stock at the DC, scaled by pooled standard deviation \(\sigma\,\sqrt{N}\,\sqrt{1+(N-1)\rho}\) and total lead time \(\sqrt{L_{\text{store}} + L_{\text{DC}}}\). Hybrid: stores cover their downstream lead time \(L_{\text{store}}\) individually, DC pools the residual lead time \(L_{\text{DC}}\) across the \(N\) stores. The chart bars compare current-\(N\) totals; the dotted line shows DC-pool savings as \(N\) sweeps \(2 \to 12\).

Reading the solution

What a network planner actually does with the policy

Three patterns to watch for

  • Hybrid usually wins. Stores need some on-hand to serve customers within store-lead-time; pure DC-only is unrealistic when \(L_{\text{store}} > 0\). The hybrid policy — stores buffer their own \(L_{\text{store}}\), DC pools the residual lead time \(L_{\text{DC}}\) — captures most of the pooling benefit while remaining feasible.
  • Correlation kills pooling. Push \(\rho\) toward 1 (a regional weather event hits all stores together) and the \(\sqrt{N}\) law degenerates. Pooling pays off precisely when local demand shocks are statistically independent.
  • Service-level convexity. The safety factor \(z = \Phi^{-1}(1-\alpha)\) is steep near \(\alpha = 0\): going from 95% to 99% service nearly doubles the buffer. This is the cost of high service, regardless of pooling.

Sensitivity questions the model answers instantly

  • Open 4 more stores in the same region? — raise \(N\) and watch DC-pool savings widen by \(\sqrt{(N+4)/N}\) below stores-only.
  • Faster DC-to-store transport (cut \(L_{\text{store}}\) from 2 days to 1)? — hybrid total falls; DC pools more of the lead time.
  • What if a regional event correlates demand at \(\rho = 0.4\)? — pooling benefit shrinks visibly — sometimes by half.

Model extensions

From two-echelon baseline to industrial multi-echelon variants

Dynamic allocation

Eppen & Schrage (1981) fair-share allocation: when DC stock is short, divide proportionally to expected store demand rather than first-come-first-served.

Correlated demand pooling

Move beyond uniform \(\rho\): regional clusters, seasonality-driven correlation, promotion-event spikes. Pooling benefit becomes a function of the full covariance matrix.

Multi-SKU coordination

Joint replenishment across SKUs sharing transport / handling. Saves on order setup; complicates the per-SKU base-stock optimisation.

Perishable / fashion variants

Short-life products (groceries, weekly fashion drops): salvage replaces holding cost. See fashion buying.

Fashion buying →
Single-store (s,S) link

Drop the upstream node and the problem reduces to classic single-echelon (s,S) replenishment at each store.

Store replenishment →
Omnichannel pooling

One DC serving both physical stores and online fulfilment. Pooling extends across channels; ship-from-store is a fluid recourse.

Supply-disruption robust

DC supply itself is uncertain (port closures, supplier outages). Robust / chance-constrained reformulation hardens the upstream node.

AI-driven dynamic allocation

Reinforcement learning policies that re-allocate DC inventory across stores in response to live POS signals, beating static base-stock under non-stationary demand.

Key references

Foundational multi-echelon and pooling literature

Clark, A. J. & Scarf, H. (1960).
Optimal policies for a multi-echelon inventory problem.
Management Science 6(4): 475–490. doi:10.1287/mnsc.6.4.475
Eppen, G. D. (1979).
Effects of centralization on expected costs in a multi-location newsboy problem.
Management Science 25(5): 498–501. doi:10.1287/mnsc.25.5.498
Federgruen, A. & Zipkin, P. (1984).
Approximations of dynamic, multilocation production and inventory problems.
Management Science 30(1): 69–84. doi:10.1287/mnsc.30.1.69
Eppen, G. & Schrage, L. (1981).
Centralized ordering policies in a multi-warehouse system with leadtimes and random demand.
In: Multi-Level Production / Inventory Control Systems, North-Holland.
Fisher, M. & Raman, A. (2010).
The New Science of Retailing.
Harvard Business Review Press.
Axsäter, S. (2015).
Inventory Control (3rd ed.).
Springer International Series in Operations Research & Management Science. doi:10.1007/978-3-319-15729-0
Graves, S. C. & Willems, S. P. (2000).
Optimizing strategic safety stock placement in supply chains.
Manufacturing & Service Operations Management 2(1): 68–83. doi:10.1287/msom.2.1.68.23267
Cachon, G. P. & Fisher, M. (2000).
Supply chain inventory management and the value of shared information.
Management Science 46(8): 1032–1048. doi:10.1287/mnsc.46.8.1032.12029

Back to the retail domain

Multi-echelon inventory sits in the Place × Strategic / Tactical cell of the retail decision matrix — the structural lever that decides where in the network to hold safety stock.

Open Retail Landing
Educational solver · iid-stores assumption with uniform correlation · validate against your own demand covariance and lead-time variability before policy decisions.