Constrained Inventory Optimization

Introduction and Context

Inventory decisions are rarely made in isolation. A firm may know the economically attractive replenishment quantity for each item, yet be unable to implement all of those quantities because cash, warehouse space, supplier capacity, transport capacity, labor, or purchasing policy is limited. Constrained inventory optimization addresses this problem by selecting a coordinated portfolio of inventory decisions rather than optimizing each stock-keeping unit (SKU) independently.

The central managerial question is not simply, “What is the best order quantity for item i?” It is: “Which combination of items produces the greatest total economic contribution while respecting all relevant constraints?” This change in perspective is important in retailing, manufacturing, healthcare, spare-parts management, and humanitarian supply chains.

The chapter develops a framework based on marginal analysis, mathematical programming, scenario analysis, and simulation. It also emphasizes the difference between a mathematically feasible plan and an operationally robust plan. A solution that uses every dollar of budget and every cubic metre of storage may be optimal under average assumptions but dangerously fragile when demand or supply conditions change.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Explain why independent SKU optimization can be infeasible or economically inferior under shared constraints.
2. Formulate multi-item inventory decisions as linear, integer, or mixed-integer programs.
3. Use marginal value, shadow prices, and opportunity cost to prioritize scarce resources.
4. Incorporate budget, storage, supplier, minimum-order, and capacity restrictions.
5. Evaluate solutions through scenario analysis and Monte Carlo simulation.
6. Interpret model outputs critically and recognize data, behavioral, and governance limitations.

1. The Logic of Constrained Inventory Decisions

For an unconstrained item, a classical economic order quantity model balances annual ordering and holding costs:

Here, D_i is annual demand in units per year, S_i is ordering cost per order, h_i is annual holding cost per unit, and Q_i^* is the unconstrained order quantity in units. The model assumes relatively stable demand, replenishment without shortages, constant costs, and no binding shared resource. It is useful as a benchmark, not as a complete portfolio decision rule.

If every SKU is assigned its own unconstrained quantity, the resulting plan may exceed total budget or warehouse capacity. Moreover, the item with the largest absolute profit is not necessarily the best use of a scarce resource. The relevant comparison is often contribution per unit of the binding resource.

Suppose item i produces contribution c_i per unit and consumes a_i units of a scarce resource. Its contribution per resource unit is:

This ratio is useful when there is one binding resource, divisible decisions are permitted, and other constraints do not alter the ranking. It is not a universal replacement for optimization: multiple constraints, integer quantities, service requirements, and nonlinear costs can make simple ranking unreliable.

Important distinctions

• Order quantity: Units purchased in one replenishment.
• Order frequency: Number of replenishment orders over a planning horizon.
• Cycle stock: Inventory created by ordering in batches.
• Safety stock: Inventory held to protect against uncertainty.
• Available-to-promise inventory: Stock that can be committed to customers after considering existing obligations.
• Decision variable: A quantity controlled by the model, such as order quantity or allocation.
• Constraint: A restriction that a feasible solution must satisfy.
• Objective function: The measure being maximized or minimized.

2. Economic Measures for the Objective Function

The objective must reflect the decision’s economic consequences. Common measures include contribution margin, total relevant cost, expected profit, or service-adjusted cost.

For a planning horizon, a simplified contribution objective is:

where:

• Z = total contribution or profit;
• p_i = selling price per unit of item i;
• v_i = variable procurement and handling cost per unit;
• x_i = units acquired or allocated;
• F_i = fixed cost incurred if item i is ordered;
• y_i = binary variable equal to 1 if an order is placed and 0 otherwise;
• h_i = relevant holding cost per unit for the planning period;
• I_i = inventory carried during the period;
• P = shortage, late-delivery, or service penalty.

The formula is not appropriate unless the terms are measured over the same horizon and in consistent monetary units. If demand is uncertain, x_i may not equal sales. Expected sales may instead be modeled as min(D_i, x_i), or through scenarios. In a service-sensitive environment, lost-sales penalties and customer-priority weights may be more appropriate than accounting profit alone.

A cost-minimization formulation can be preferable when demand and service targets are prescribed:

The choice between profit maximization and cost minimization is a governance decision as well as a technical decision. Managers should document whether the model values revenue, margin, customer service, resilience, or some weighted combination.

3. Translating Inventory Restrictions into a Mathematical Program

3.1 Budget constraint

If available purchasing budget is B and unit acquisition cost is c_i, then:

Both sides are monetary values. If budget is based on average inventory rather than purchases, the coefficients must represent the appropriate inventory exposure rather than order value.

3.2 Storage constraint

If each unit occupies v_i cubic metres and capacity is V, then:

The use of I_i rather than x_i is important when goods flow through the warehouse during the period. If the constraint concerns receiving or peak on-hand stock, the model must represent that specific exposure.

3.3 Supplier capacity

For supplier s with capacity K_s and supplier set S_s, a capacity restriction is:

This can represent production capacity, allocation quotas, import restrictions, or contractual limits. If capacity varies by period, use a separate constraint for each period.

3.4 Minimum-order quantities and fixed charges

A supplier may require a minimum order m_i whenever an item is purchased. Introduce binary variable y_i:

Here, M_i is a valid upper bound, and y_i ∈ {0,1}. If y_i = 0, the item is not ordered; if y_i = 1, the quantity lies between its minimum and maximum. The upper bound should be as tight as practically possible. An unnecessarily large M_i can slow mixed-integer optimization and produce weak computational relaxations.

3.5 Warehouse or handling capacity

If item i uses w_i units of handling labor and available labor is W, then:

A model may contain several simultaneous resource constraints. In that case, ranking products by one ratio is generally invalid because improving one resource utilization may worsen another.

3.6 Service-level constraints

Let SL_i denote the achieved service level and α_i the required minimum. Then:

A cycle-service approximation for normally distributed demand during lead time is:

where SS_i is safety stock, z_i is the standard-normal quantile associated with the target service level, and σ_L,i is the standard deviation of lead-time demand. This approximation assumes an appropriate demand distribution and a reliable estimate of lead-time variability. It should not be used mechanically for intermittent or highly skewed demand.

4. Marginal Analysis and Shadow Prices

Marginal analysis asks how much objective value is gained by relaxing a constraint by one unit. In a linear program, the associated shadow price is the marginal improvement in the objective, within the range over which the current basis remains valid.

For example, if the shadow price of budget is $0.18 per dollar, an additional $1,000 of budget is worth approximately $180 in objective improvement, provided the change remains within the sensitivity range. This is not necessarily the market price of capital and should not be extrapolated indefinitely.

When only one resource is binding and quantities are divisible, products can be considered in descending order of contribution per resource unit. With two or more binding resources, use linear programming or mixed-integer programming. A product with lower contribution per dollar may still be selected because it uses little storage, supplier capacity, or labor.

Reduced cost provides another useful interpretation. For a product excluded from an optimal linear-programming solution, reduced cost indicates how much its contribution coefficient would need to improve before it becomes attractive, subject to the model’s current structure. In integer programs, reduced-cost interpretations are more limited and should be treated cautiously.

5. Worked Example: Budget and Storage Constraints

A distributor is selecting quantities of three products for a monthly promotion. The available budget is $50,000 and storage capacity is 400 cubic metres. The following data apply:

Let x_A, x_B, and x_C be quantities purchased. The linear program is:

subject to:

Contribution per dollar is $0.40 for A, $0.50 for B, and $0.60 for C. Contribution per cubic metre is $40, $60, and $30 respectively. The rankings conflict: C is best on budget, while B is best on storage. Consequently, a single ratio cannot identify the optimum.

Consider the feasible solution (x_A,x_B,x_C)=(0,500,500). It uses $50,000 of budget and 650 cubic metres, so it is infeasible. The storage restriction is decisive in that combination. A feasible candidate is (0,400,250), using $34,000 and 400 cubic metres, with contribution $18,000. The unused budget indicates that storage is binding while budget is slack.

A better candidate can be found by filling storage with B and C while checking budget. If all 500 units of B are selected, the remaining storage is 150 cubic metres, allowing 187.5 units of C; the budget used is $37,500 + $7,500 = $45,000 and contribution is $19,500. The remaining budget cannot be used without exceeding storage unless product A replaces some units.

Compare replacing one unit of B by one unit of A: storage rises by 0.5 cubic metre, contribution rises by $10, and cost rises by $40. Compare replacing one unit of C by one unit of A: storage rises by 0.2 cubic metre, contribution rises by $16, and cost rises by $60. The second replacement is more attractive on both contribution per additional storage and contribution per additional dollar. Integer optimization should be used because actual units are indivisible.

The example illustrates three points. First, an apparently attractive product can be excluded because it consumes the wrong scarce resource. Second, unused budget does not imply a poor solution when storage is binding. Third, manual marginal comparisons are useful for interpretation but are not a substitute for a solver when several constraints and integer requirements interact.

6. Integer and Mixed-Integer Programming

Inventory decisions often contain indivisibilities:

• case-pack quantities;
• pallet multiples;
• minimum order quantities;
• supplier activation decisions;
• quantity discounts;
• product substitution choices;
• facility or transport-mode selection.

A mixed-integer linear program (MILP) combines continuous and discrete variables. A generic form is:

Here, x represents continuous quantities, y represents binary decisions, c and f are objective coefficients, A and G are constraint matrices, and b is the vector of resource limits. The formulation assumes linear relationships. Nonlinear storage, quantity discounts, or risk functions may require piecewise-linear approximations, nonlinear programming, or specialized algorithms.

A solver may report an optimal solution, a feasible solution with a known optimality gap, or no feasible solution. Managers should distinguish these outcomes. A solution with a 0.5% optimality gap may be operationally preferable to a mathematically exact solution that depends on unreliable forecasts.

7. Scenario Analysis and Robustness

A single forecast hides uncertainty. Scenario analysis evaluates the inventory plan under alternative assumptions, such as:

• base demand, high demand, and low demand;
• normal and disrupted supplier capacity;
• different exchange rates or purchase prices;
• alternative lead times;
• changes in warehouse capacity;
• different service-level priorities.

Let scenario s have probability p_s and objective value Z_s. Expected value is:

A plan optimized only for expected demand may perform badly in a disruption scenario. Two-stage stochastic programming can represent an initial order decision followed by recourse, such as emergency purchasing or substitution. A robust optimization approach instead seeks a solution that remains acceptable across a defined uncertainty set. Robustness generally has a cost: the firm may sacrifice expected profit to reduce severe downside outcomes.

A practical workflow is:

1. Define uncertain parameters and plausible ranges.
2. Establish scenarios with documented assumptions.
3. Optimize or evaluate the same candidate plan under each scenario.
4. Compare expected contribution, worst-case performance, service failure, and resource violations.
5. Select a plan using explicit risk preferences rather than an unexplained preference for the average case.

8. Simulation for Service and Risk Evaluation

Optimization identifies a recommended decision; simulation tests how that decision behaves under repeated uncertainty. In a Monte Carlo inventory simulation, demand and lead time are sampled repeatedly from specified distributions, inventory is updated period by period, and performance statistics are recorded.

For replication r, total cost may be represented as:

After R replications, estimated mean cost is:

The standard error of the estimated mean is:

where s_TC is the sample standard deviation of replication costs. Independent replications, adequate warm-up treatment for continuing systems, and a sufficiently large R are required for credible inference. Simulation output is not automatically accurate merely because many trials are run; biased demand distributions or incorrect replenishment logic remain problematic.

Simulation can estimate fill rate, cycle service level, stockout frequency, average inventory, expedite frequency, and tail losses. It is particularly valuable when demand is intermittent, lead times are correlated, or replenishment policies are nonlinear.

9. Applied Case: Northstar Medical Distribution

Background

Northstar Medical Distribution supplies hospitals with three categories of sterile products. Product A has high margin but occupies substantial refrigerated space. Product B has moderate margin and reliable domestic supply. Product C is compact but imported, with a supplier allocation limit. Hospitals require a 98% fill rate for emergency products and a 95% target for routine products.

Problem

For the next quarter, Northstar faces a $1.2 million purchasing budget, 9,000 refrigerated pallet positions, supplier limits, and minimum pallet orders. The independent forecasts would require $1.45 million and 10,600 pallet positions. Management must decide whether to reduce quantities, change the product mix, or purchase emergency supply later at a premium.

Analysis

The team formulates a MILP with order quantities, binary supplier-activation variables, inventory positions by month, and shortage variables. Budget and pallet capacity are hard constraints. Emergency procurement is permitted as a recourse decision with a high penalty. Product-specific service constraints are imposed only for emergency-class items; applying a 98% target to every SKU would consume excessive capacity with limited clinical benefit.

The team then tests the solution using 10,000 simulated demand and lead-time paths. Supplier disruption is modeled in a separate scenario rather than mixed invisibly into the base distribution. Results show that the initial MILP plan has the highest expected contribution, but its emergency-service failure rate is unacceptable under a two-week import disruption. A slightly lower-margin plan increases domestic Product B and reduces dependence on C. It lowers expected contribution by 1.8% but reduces the 95th-percentile shortage cost by 27%.

Approach and outcome

Northstar adopts the diversified plan, negotiates an additional 500 pallet positions for the peak month, and establishes a trigger for emergency orders when projected coverage falls below a defined threshold. The model is rerun weekly with updated forecasts. The decision is not justified by expected profit alone; it reflects service criticality, disruption exposure, and the asymmetric consequences of hospital shortages.

Learning points

1. A constraint can be physical, financial, contractual, or ethical.
2. Service requirements should reflect customer and clinical criticality.
3. Scenario analysis can reveal weaknesses hidden by expected values.
4. A slightly inferior nominal solution may be superior on risk-adjusted performance.
5. Model governance requires documented assumptions and regular review.

10. Implementation, Governance, and Technology

A constrained inventory model should be integrated with enterprise resource planning, warehouse-management, purchasing, and supplier-management systems. Data interfaces should identify unit conversions, pack sizes, lead-time definitions, obsolete items, and timing conventions. A model that treats a case as one unit while the warehouse treats it as 24 eaches can generate apparently precise but operationally invalid recommendations.

Artificial intelligence and machine learning can improve demand forecasting, anomaly detection, lead-time prediction, and scenario generation. They do not eliminate the need for optimization logic. A highly accurate forecast can still produce a poor decision if budget, minimum-order, or service constraints are omitted. AI-generated recommendations should be auditable, bounded by business rules, and reviewed for bias—for example, systematically deprioritizing low-volume rural customers.

Governance should address:

• ownership of model assumptions;
• approval of objective functions and service priorities;
• version control and reproducibility;
• monitoring of forecast error and constraint violations;
• human override authority;
• privacy and security of supplier and customer data;
• documentation of exceptional decisions.

Ethically, optimization should not allocate scarce goods solely according to margin when safety, equity, or contractual obligations are relevant. In healthcare and humanitarian operations, priority weights may need to reflect need and urgency rather than profitability. Regulatory requirements may also impose minimum stock, traceability, expiry, or recall obligations.

11. Limitations and Best Practices

The main limitations are often managerial rather than mathematical. Forecasts may be biased, costs may omit disruption consequences, and constraints may be measured incorrectly. Linear models can conceal nonlinear effects such as congestion, expiry, or volume discounts. Shadow prices can be misinterpreted outside their valid sensitivity ranges. Simulation results are conditional on the assumed distributions and policy logic.

Best practices include:

1. Begin with a clearly defined planning horizon and unit system.
2. Separate hard constraints from desirable targets.
3. Validate inputs with purchasing, warehouse, finance, and commercial teams.
4. Use tight bounds and appropriate integer variables.
5. Compare independent-SKU, optimized, and robust alternatives.
6. Report binding constraints, shadow prices, infeasibilities, and optimality gaps.
7. Stress-test disruption, forecast error, and lead-time variability.
8. Review recommendations against practical ordering calendars and supplier relationships.
9. Monitor realized service, inventory, and shortage outcomes after implementation.
10. Recalibrate the model when assumptions or business priorities change.

Chapter Summary

Constrained inventory optimization coordinates multiple SKU decisions when resources are limited. The objective may maximize contribution, minimize relevant cost, or balance service and risk. Budget, storage, supplier capacity, minimum-order quantities, fixed charges, labor, and service requirements can be represented as mathematical constraints. Marginal contribution and shadow prices clarify the value of scarce resources, but simple ratios are reliable only in limited special cases. Linear programming handles divisible decisions, while integer and mixed-integer programming represent case packs, minimum orders, and activation choices. Scenario analysis and simulation evaluate uncertainty, disruption exposure, and service performance. Effective practice combines mathematical rigor with data governance, operational validation, ethical judgment, and continuous monitoring.

Key Terms

Constrained optimization: Selecting the best decision subject to specified restrictions.

Objective function: Mathematical expression representing the goal of the model.

Feasible solution: A decision that satisfies every model constraint.

Binding constraint: A constraint satisfied exactly at the solution and limiting improvement.

Shadow price: Marginal objective improvement from relaxing a constraint within a valid sensitivity range.

Reduced cost: Improvement required in an excluded variable’s objective coefficient for it to become attractive in a linear program.

Mixed-integer linear program: Optimization model containing continuous and discrete variables with linear relationships.

Minimum-order quantity: Smallest permitted order when an item is purchased.

Fixed ordering cost: Cost incurred when an item or supplier is activated, regardless of quantity.

Cycle stock: Inventory resulting from replenishment lot size.

Safety stock: Inventory held to protect against uncertainty.

Service level: Measure of the probability or proportion of demand supplied from available stock.

Fill rate: Proportion of demand quantity supplied without delay.

Scenario analysis: Evaluation of decisions under alternative assumptions.

Stochastic programming: Optimization incorporating probabilistic uncertainty and recourse.

Robust optimization: Optimization seeking acceptable performance across an uncertainty set.

Monte Carlo simulation: Repeated random sampling used to estimate system performance.

Optimality gap: Difference between a known feasible solution and a bound on the best possible solution.

Recourse: Corrective action available after uncertain conditions are observed.

Model governance: Controls ensuring that models are valid, transparent, approved, and monitored.

Review Questions

Short Answer Questions

1. Why can independently optimal SKU quantities be infeasible at the portfolio level?
2. What is the managerial meaning of a shadow price on warehouse capacity?
3. Distinguish cycle stock from safety stock.
4. When is contribution per unit of a scarce resource a useful prioritization rule?
5. Why are binary variables used for minimum-order quantities?

Long Answer Questions

1. Formulate a multi-item inventory model containing budget, storage, supplier-capacity, and service-level constraints. Define all variables and assumptions.
2. Explain the differences between linear programming, integer programming, and mixed-integer programming in inventory applications.
3. Discuss why expected-profit optimization may produce a fragile inventory plan.
4. Explain how Monte Carlo simulation can complement an optimization model.
5. Evaluate the ethical and governance issues that arise when constrained inventory is allocated among customers with different levels of need.

Analytical / Application Questions

1. A firm has two binding resources and three products. Explain why ranking products by contribution per dollar may produce a poor solution, and describe an appropriate analytical method.
2. Design a scenario-analysis plan for an importer facing uncertain demand, exchange rates, and supplier disruption. Specify performance measures and decision criteria.
3. A solver reports a feasible solution with a 1% optimality gap, but simulation shows a high probability of service failure. Explain how management should interpret and respond to these results.

Multiple Choice Questions with Answers

Question 1

Which condition most directly justifies ranking products by contribution per unit of a scarce resource?

A. There is one binding resource and decisions are divisible.

B. There are several binding resources and fixed order costs.

C. All products have identical resource consumption.

D. Demand is completely uncertain and intermittent.

Correct Answer: A. There is one binding resource and decisions are divisible.

Question 2

In the constraint ∑ c_i x_i ≤ B, what does B represent?

A. Total storage volume

B. Available purchasing budget

C. Supplier lead time

D. Required service level

Correct Answer: B. Available purchasing budget

Question 3

What type of variable is normally used to indicate whether an item is ordered at all?

A. Continuous variable

B. Forecast-error variable

C. Binary variable

D. Shadow-price variable

Correct Answer: C. Binary variable

Question 4

A shadow price of $0.18 per additional dollar of budget means that, locally, an extra $1,000 of budget is expected to improve the objective by approximately:

A. $18

B. $180

C. $1,800

D. $18,000

Correct Answer: B. $180

Question 5

Which model is most appropriate when case-pack quantities and supplier activation decisions must be represented?

A. Simple moving-average model

B. Mixed-integer linear program

C. Unconstrained EOQ model

D. Qualitative SWOT analysis

Correct Answer: B. Mixed-integer linear program

Question 6

What is the principal purpose of scenario analysis in constrained inventory planning?

A. To eliminate all uncertainty from the forecast

B. To evaluate performance under alternative assumptions

C. To replace every operational constraint with an average

D. To guarantee the expected-profit solution is optimal

Correct Answer: B. To evaluate performance under alternative assumptions

Question 7

If a product requires 2 cubic metres per unit and warehouse capacity is 100 cubic metres, the maximum quantity of that product alone is:

A. 20 units

B. 50 units

C. 100 units

D. 200 units

Correct Answer: B. 50 units

Question 8

Which statement about simulation is correct?

A. Simulation automatically corrects biased input distributions.

B. Simulation tests how a specified policy performs under repeated uncertainty.

C. Simulation makes integer variables unnecessary.

D. Simulation always proves global optimality.

Correct Answer: B. Simulation tests how a specified policy performs under repeated uncertainty.

Question 9

In the service-level approximation SS_i = z_i σ_L,i, what does σ_L,i represent?

A. Unit purchase price

B. Annual ordering cost

C. Standard deviation of lead-time demand

D. Supplier capacity

Correct Answer: C. Standard deviation of lead-time demand

Question 10

Why might a firm choose a slightly lower expected-profit plan?

A. It may reduce disruption exposure and severe shortage outcomes.

B. It always uses more budget and storage.

C. It removes the need for governance.

D. It guarantees demand will equal the forecast.

Correct Answer: A. It may reduce disruption exposure and severe shortage outcomes.