8. Multi-Echelon Inventory Optimization

The earlier inventory models treated one item at one stocking location. Real networks can have plants, regional hubs, distribution centers, and stores, each holding stock for the same customer flow. Optimizing every location independently can protect local targets while increasing total inventory, hiding shortages, or amplifying order swings upstream.

Multi-echelon inventory optimization (MEIO) coordinates where buffers sit across connected stocking points. It can improve service, reduce working capital, or trade one outcome for the other. It is not a guarantee of a fixed percentage savings: the result depends on the network, data quality, costs, service definition, and current operating maturity.

Represent the network

A supply-chain network has nodes (stocking or processing locations) connected by edges (supplier-to-customer relationships). A layer or stage is an echelon. The first stage supplied from outside the modeled network is the supply side; the stage facing final customer demand is the demand side.

Network type	Relationship pattern	Example
Serial	Each intermediate node has one supplier and one downstream customer node	Plant → regional hub → distribution center
Assembly	A node combines supply from multiple upstream nodes	Several component suppliers feed one assembly plant
Distribution	One upstream node supplies multiple downstream nodes	One central warehouse supplies several stores
General	Nodes may have multiple suppliers and multiple downstream customers	A branched, multi-source network

A general network may be acyclic or include cycles such as return flows. A supplier can manage inventory at a customer location under vendor-managed inventory (VMI). If the supplier also owns stock physically held at the customer, that arrangement is often called consignment. Control, physical location, and ownership are distinct attributes; record each explicitly in a network model.

Local decisions can conflict with the network

A locally optimized node chooses its own reorder point, batch, review interval, or safety stock using only its own objectives and observed orders. The upstream node then treats those orders as demand. If every stage independently forecasts and reacts to those orders, a small change at the final customer can become a larger, more irregular signal upstream. This amplification is called the bullwhip effect.

Common contributors include:

• Order forecasting: each stage forecasts its customer's orders rather than sharing final demand and inventory context.
• Order batching: infrequent or large orders create lumpy signals.
• Price changes and promotions: buyers shift order timing or build forward stock.
• Shortage gaming: customers inflate orders when supply is constrained, hoping to receive a larger allocation.

Practical countermeasures are better end-demand and inventory visibility, coordinated replenishment, shorter review and lead times where feasible, transparent allocation rules during shortage, and aligned promotion planning. A global model helps only when its data and operating rules are used consistently.

Decide where buffers belong

Three forces often pull inventory in different directions:

Driver	Often favors	Why
Lead-time pooling	Downstream	A downstream buffer can cover a longer combined replenishment exposure, pooling variation over more periods and potentially reducing total safety stock.
Demand pooling	Upstream	An upstream node may serve several locations; aggregation can smooth demand when branch variation is not too strongly positively correlated.
Value and holding cost	Upstream	Earlier-stage material may be cheaper to hold than finished goods stored near expensive customer locations.

These are tendencies, not a universal placement rule. Correlation, transport, capacity, product value, service promises, and ownership can change the trade-off. MEIO compares the whole network's cost and service rather than assuming one driver always wins.

A practical serial-network model: guaranteed service

Two broad modelling approaches help frame the trade-off. Stochastic-service models (SSM) represent how upstream stockouts make downstream replenishment times random; this can be realistic but computationally demanding. The guaranteed-service model (GSM) simplifies the problem by assuming each node protects demand up to a chosen bound and quotes a fixed service time; demand above the bound is handled by an exceptional response such as expediting, overtime, or a delayed customer promise.

This is an explicit service assumption, not magic capacity. A GSM answer is useful only if the organization can execute the exception plan when demand exceeds its modeled bound.

Risk-period allocation

For a serial network with aligned review period $R$ and deterministic incoming lead times $L_i$, the total risk-period to cover is:

$$T=R+\sum_{i=1}^{n}L_i$$

Let $x_i$ be the number of periods of demand protected by safety stock at node $i$. In the simplified serial framework used here:

$$\sum_i x_i=T,\qquad x_i\ge0$$

For independent Normal demand with per-period standard deviation $\sigma_d$ and service factor $z$, the safety stock at node $i$ is:

$$SS_i=z\sigma_d\sqrt{x_i}$$

If per-period mean demand is $\mu_d$, the corresponding local order-up-to target is $S_i=\mu_dx_i+SS_i$. Local targets can then be converted into echelon targets for coordinated ordering.

The simplified serial GSM has an all-or-nothing structure at each candidate stocking point: a node either covers its eligible incoming risk segment or receives zero risk-period allocation. The demand-facing node covers the final review interval. For three nodes, the worked candidate allocations below are four feasible cases; they are not a license to use the same enumeration for every general network.

In this simplified serial setup, the candidate count grows as $2^{n-1}$ for $n$ nodes because the demand-facing node must cover the final review interval. This is manageable for a short chain, but the assumption and search do not transfer directly to arbitrary assembly or general networks.

Worked example: compare safety-stock cost

Consider a three-node serial chain with incoming lead times $[4,3,2]$ periods and common review interval $R=1$. Total risk-period is $T=10$. Demand has per-period mean $\mu_d=100$ and standard deviation $\sigma_d=25$; target cycle service is 95% ($z=1.645$). Holding costs are $[1,2,4]$ per unit per period for upstream, middle, and downstream stock.

Under the simplified model, compare these four feasible risk-period allocations:

Case	Upstream $x_1$	Middle $x_2$	Demand-side $x_3$
1	4	3	3
2	0	7	3
3	4	0	6
4	0	0	10

For each case calculate $SS_i=z\sigma_d\sqrt{x_i}$ (use zero stock where $x_i=0$), then total holding cost $C_h=\sum_i h_iSS_i$. Rounded results are:

Case	Safety stock by node (units)	Total safety stock	Holding cost
1	82.3, 71.2, 71.2	224.7	509.6
2	0, 108.8, 71.2	180.0	502.5
3	82.3, 0, 100.7	183.0	485.2
4	0, 0, 130.0	130.0	520.2

Case 4 minimizes the number of safety-stock units because it pools the entire risk period downstream. Case 3 has the lowest holding cost in this example because the high downstream holding rate makes it cheaper to place part of the buffer upstream. Minimizing units is not the same as minimizing cost. The result also excludes cycle stock, transport, capacity, and the cost of exceptional responses; those omissions must be understood before using the allocation operationally.

Translate local targets into echelon orders

A local order-up-to target $S_i$ describes stock at one node. An echelon target $\mathcal{S}_i$ describes stock at that node and all downstream nodes. In a simple serial network ordered upstream to downstream:

$$\mathcal{S}_i=\sum_{j=i}^{n}S_j$$

Echelon inventory is calculated the same way from current local inventory $I_j$:

$$\mathcal{I}_i=\sum_{j=i}^{n}I_j$$

The upstream order requirement is the shortfall between echelon target and echelon inventory. Physical availability and capacity constraints may then cap what is actually shipped; never interpret an unconstrained target shortfall as proof that stock can be delivered.

For Case 1, the local targets are $S_i=\mu_dx_i+SS_i$, giving approximately $[482,371,371]$. These give echelon targets $[1224,742,371]$. Current local inventory $[300,250,100]$ gives echelon inventory $[650,350,100]$, so the raw echelon shortfalls are $[574,392,271]$ units.

Limits and next steps

The serial example assumes aligned review timing, deterministic lead times, shared demand information, bounded demand, and a simplified exception response. Stochastic-service models represent interactions between stockouts and replenishment times more explicitly, but they can be much harder to solve. Assembly and general networks may need constrained optimization or dynamic programming. Distribution systems add demand pooling and possible correlation between branches.

Use the serial GSM as a transparent starting point. For a real network, validate the demand and cost inputs, model capacity/transport constraints, make the emergency-service promise operational, and compare predicted results with simulation or observed performance.

Practice: allocate and execute

Concept check

Quiz

Network-wide inventory decisions

1 / 5

A central warehouse serves several stores. Which topology best describes this network?

Pick an answer to see if you are right.

Rank the serial GSM analysis

Order the dependent analysis steps for the stated three-node case.

Quiz

Compare serial-network allocations

1 / 1

Lead times, review period, demand mean and SD, service factor, holding costs, and four feasible allocations are given.

OrderDrag the handle to rearrange — top should be first.

1. ⋮⋮1Multiply each node’s safety stock by its holding cost and sum to get network cost.
2. ⋮⋮2Convert local targets and local inventory into echelon targets, echelon inventory, and order requirements.
3. ⋮⋮3For each allocation, calculate node safety stock z sigma sqrt(x_i).
4. ⋮⋮4Compare feasible cases by the chosen objective, then check omitted operating constraints.
5. ⋮⋮5Confirm total risk period and that each candidate allocation covers it while meeting node constraints.
6. ⋮⋮6For the selected allocation, calculate local targets S_i = mean demand times x_i plus safety stock.

Drag items into the correct order (or use ↑↓), then check.

Python practice: compare feasible allocations

Use the four risk-period cases and node holding costs from the worked example. Each row below is one candidate allocation, with periods protected at upstream, middle, and downstream nodes.

Case	$x_1$	$x_2$	$x_3$
1	4	3	3
2	0	7	3
3	4	0	6
4	0	0	10

Implement allocation_costs(z_sigma, holding_costs, allocations). Return the one-based lowest-cost case followed by each case's holding cost. The test wrapper formats values to one decimal.

Python practice: convert local inventory to echelon orders

Use upstream-to-downstream arrays. A serial echelon includes that node and all nodes downstream. Do not create a negative purchase order when echelon inventory exceeds its target.

Node	Local target	Current local inventory
Upstream	482	300
Middle	371	250
Demand-facing	371	100

Implement echelon_orders(local_targets, local_inventory) and return order requirements from upstream to downstream.

Takeaways

• MEIO chooses buffer locations across a network; local targets alone do not describe total network cost or service.
• Bullwhip can be amplified by order forecasting, batching, promotions, and shortage gaming; visibility and coordinated rules help.
• Lead-time pooling, demand pooling, and value/holding cost can favor different inventory locations.
• The serial GSM is a bounded-demand approximation with an explicit exception response, not a universal exact optimizer.
• Echelon inventory and targets include downstream nodes; cap operational orders by real availability and capacity.

Previous: Beyond Normality · Next: Distribution MEIO and serial simulation.