Stochastic Inventory, Demand, and Supply Networks

Introduction

Inventory decisions are made under uncertainty. A firm must decide how much to order, when to order, where to position stock, and how to respond when demand, replenishment time, supplier capacity, or transportation conditions vary. These decisions are stochastic because their relevant inputs are random variables rather than known constants.

Inventory is simultaneously an operational asset, a financial commitment, and a risk buffer. Excess inventory creates holding cost, obsolescence, markdowns, and working-capital requirements. Insufficient inventory creates shortages, lost sales, backorders, emergency transportation, and damaged customer relationships. Stochastic modeling provides a disciplined way to quantify these trade-offs.

This chapter develops models for uncertain demand and lead times, safety stock and service levels, the newsvendor problem, reorder-point and base-stock policies, multi-echelon networks, disruption risk, and resilient supply-chain design. The emphasis is not only on formulas but also on assumptions, interpretation, data requirements, and managerial judgment.

Learning Objectives

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

1. Represent demand and replenishment uncertainty using appropriate probability models.
2. Calculate reorder points, safety stock, cycle stock, and service levels.
3. Explain and apply the newsvendor critical-fractile model.
4. Compare reorder-point, order-up-to, and base-stock policies.
5. Analyze pooling, multi-echelon inventory, and risk propagation across networks.
6. Evaluate disruption scenarios and design practical resilience responses.
7. Critically assess model assumptions, data quality, ethics, and governance.

1. Inventory as a Stochastic Control Problem

A periodic inventory position can be represented as

where IP_t is inventory position at time t, OH_t is on-hand inventory, OO_t is on-order inventory, and BO_t is backordered demand. All quantities use the same physical unit. Inventory position differs from on-hand inventory because outstanding orders and backorders affect future availability.

A policy maps observed information into an order decision. In a stochastic setting, the decision may depend on current stock, outstanding orders, forecasts, observed demand, and the state of suppliers. The principal performance measures are:

• Cycle service level (CSL): probability of no stockout during a replenishment cycle.
• Fill rate: proportion of demand units supplied immediately from stock.
• Expected shortage: average units unavailable during a specified period.
• Inventory turnover: annual demand or cost of goods sold divided by average inventory.
• Total relevant cost: ordering, holding, shortage, purchasing, transportation, and sometimes obsolescence cost.

CSL and fill rate are not interchangeable. A single stockout can produce a low cycle service level even if only a few units are short, whereas fill rate measures the volume of unmet demand.

2. Modeling Demand and Lead-Time Uncertainty

2.1 Demand representation

Demand may be modeled as a time series, a count process, a conditional distribution, or a forecast-error distribution. Common choices include:

• Normal models for aggregated, relatively stable demand;
• Poisson models for low-volume count demand;
• negative binomial models when count variance exceeds the mean;
• intermittent-demand methods such as Croston-type approaches;
• empirical or bootstrap distributions when data are plentiful but non-normal;
• scenario distributions when structural changes or disruptions dominate historical patterns.

For a planning horizon of L periods, let demand in period t be D_t. Cumulative demand is

If demands are independent with means μ_t and variances σ_t², then

and

If demands are correlated, the general variance is

Positive correlation increases uncertainty relative to the independent case. This matters during promotions, weather events, price changes, or macroeconomic shocks.

2.2 Random lead time

Let lead time be the random variable L. Demand during lead time is D_L. If demand per period and lead time are independent, with constant mean μ and variance σ_D², then conditioning on lead time gives

and

The first term reflects demand variability; the second reflects lead-time variability. This decomposition is useful because it identifies whether improvement should focus on forecasting demand or stabilizing replenishment.

If demand and lead time are dependent, the independence formula is invalid. For example, congestion may increase lead time precisely when demand surges. Simulation, empirical joint distributions, or conditional models are then more appropriate.

2.3 Forecast error rather than raw demand

Inventory buffers should generally protect against forecast error, not simply against historical demand variation. If the forecast for lead-time demand is F_L and actual demand is D_L, define error as

The protection quantity should be based on the distribution of e_L, conditional on the information available when the order is placed. Using an unconditional historical standard deviation can overstate or understate risk.

3. Reorder-Point Policies and Safety Stock

A continuous-review (Q,r) policy orders quantity Q whenever inventory position reaches reorder point r. A standard reorder point is

where SS is safety stock. If lead-time demand is approximately normal with mean μ_L and standard deviation σ_L, and the target cycle service level is α, then

and therefore

Here z_α is the standard-normal quantile satisfying P(Z≤z_α)=α. This formula applies when the normal approximation is reasonable, demand and lead-time uncertainty have been correctly combined, and the desired service measure is a no-stockout probability.

For constant lead time L and independent, identically distributed demand, μ_L=Lμ and σ_L=σ_D√L. Thus,

If lead time is random and independent of demand, use

rather than simply σ_D√E[L].

Worked Example: Reorder point with uncertain lead time

A distributor sells 100 units per day on average. Daily demand has standard deviation 20 units. Lead time has mean 4 days and standard deviation 1.5 days. Assume independence and approximate lead-time demand as normal. The target cycle service level is 95%, so z_0.95≈1.645.

Mean lead-time demand is

μ_L=100×4=400 units.

Its variance is

σ_L²=20²×4+100²×1.5²=1,600+22,500=24,100.

Thus σ_L≈155.24 units, safety stock is

SS=1.645×155.24≈255.37 units,

and the reorder point is

r≈400+255.37=655.37,

or approximately 656 units when whole units are required. The large contribution from μ² Var(L) shows that lead-time uncertainty is the dominant risk source.

Cost-based safety stock

A service target is not always exogenous. If each shortage unit costs p and each unit held for the relevant period costs h, a simplified critical-ratio logic may be used, but periodic and continuous-review settings differ. Managers should avoid selecting a high service level merely because it sounds desirable; the marginal cost of additional protection should be compared with the marginal shortage cost.

4. The Newsvendor Model

The newsvendor model applies to a single ordering opportunity or a product whose value declines sharply after the selling season. Examples include fashion goods, newspapers, event catering, seasonal produce, and capacity reserved for uncertain demand.

Let Q be the order quantity and D random demand. Let selling price be p, unit acquisition cost be c, and salvage value be v, with p>c≥v. The underage cost is

and the overage cost is

The optimal order quantity under a continuous demand distribution satisfies

The ratio is called the critical fractile. It is the probability that demand should not exceed the chosen quantity under the marginal economic trade-off. The formula assumes one order opportunity, known costs, a specified salvage value, and no replenishment during the selling horizon.

Worked Example: Seasonal product

Suppose p=80, c=50, and v=20. Then C_u=30 and C_o=30, so

F(Q^*)=30/(30+30)=0.5.

The firm should order the median of the demand distribution. If demand is normal with mean 1,000 and standard deviation 200, the median equals the mean, giving Q^*=1,000 units. This result does not mean demand is known; it reflects symmetric economics, not certainty.

If salvage value increases, overage becomes less costly and the optimal quantity rises. If shortage cost rises, the optimal quantity also rises. Quantity discounts, capacity constraints, lost goodwill, and multiple replenishment opportunities require extensions of the basic model.

5. Base-Stock and Order-Up-To Policies

A base-stock policy replenishes inventory position to a target level after each review. In a periodic-review system with review interval R and lead time L, the protection period is generally R+L. A common order-up-to level is

If demand over the protection period is normal, then

The order at review time t is

where IP_t is inventory position immediately before ordering. This policy responds to both demand and outstanding orders, avoiding the mistake of ordering solely from on-hand stock.

In a continuous-review (Q,r) policy, the order quantity is often fixed and the reorder point protects lead-time demand. In an order-up-to policy, order quantities vary and the target protects the entire review-plus-lead-time interval. The choice depends on ordering costs, supplier constraints, review technology, and operational cadence.

6. Pooling and Multi-Echelon Inventory

Risk pooling occurs when demand from several locations is aggregated. If two independent locations each have standard deviation σ, their aggregate standard deviation is σ√2, not 2σ. Consequently, safety stock per unit of aggregate demand can decline. The benefit is weaker when demands are positively correlated.

For locations with standard deviations σ_i and pairwise correlations ρ_ij, aggregate variance is

Pooling is not free. It may increase delivery distance, response time, transportation cost, and vulnerability to a central-facility failure. A sound design compares total landed cost and service performance rather than safety stock alone.

Multi-echelon systems

A multi-echelon network contains several inventory stages, such as suppliers, plants, regional distribution centers, and stores. Inventory at one stage can substitute partly for inventory at another, but lead times and capacity constraints determine whether that substitution is operationally useful.

The echelon inventory at stage i includes physical inventory at that stage plus inventory downstream that it is responsible for serving. Echelon policies simplify coordination by assigning target levels to stages. Optimization may use stochastic dynamic programming, stochastic programming, simulation, or approximate policies.

Important design questions include:

• Where should uncertainty be absorbed?
• Which stage has the shortest response time to customers?
• Can inventory be repositioned laterally?
• Are supplier minimum-order quantities or production setups binding?
• Does centralization create a single point of failure?

The classical intuition that upstream inventory is always preferable is incomplete. Upstream stock is often more flexible across markets, but downstream stock may be essential when customer lead-time promises are strict or transportation is unreliable.

7. Disruption Risk and Resilient Supply-Chain Design

Routine variability and disruption risk should be modeled separately. Routine demand variation may be represented by a probability distribution estimated from many observations. A port closure, cyberattack, pandemic, or supplier bankruptcy is a low-frequency, high-impact event and may not be adequately represented by historical averages.

A scenario model assigns disruption scenario s probability p_s and cost or performance outcome C_s. Expected cost is

However, expected cost alone can hide severe tail outcomes. A risk-averse analysis may use a service constraint, a worst-case scenario, a percentile, or conditional value at risk. For loss C, value at risk at level β is a quantile of the loss distribution. Conditional value at risk is the expected loss in the worst (1−β) fraction of cases, subject to the precise convention used.

Resilience strategies include:

1. Redundancy: multiple suppliers, routes, or facilities.
2. Flexibility: substitutable materials, cross-trained labor, and adaptable production.
3. Visibility: accurate, timely information about inventory, orders, and capacity.
4. Buffering: safety stock, spare capacity, and time buffers.
5. Collaboration: shared forecasts, risk registers, and coordinated recovery plans.
6. Recovery capability: tested contingencies and prioritized customer allocation.

Resilience is not identical to robustness. Robustness maintains performance during shocks; resilience includes the ability to recover and adapt. Redundancy can improve resilience but also increase cost, complexity, and environmental impact.

8. Illustrative Case: Medical Device Network Resilience

Background

Consider MedSupply, an illustrative manufacturer of sterile diagnostic cartridges. It serves hospitals from one national distribution center and two regional warehouses. Demand is moderately seasonal. A critical plastic component is sourced from one overseas supplier with an average lead time of 28 days and a standard deviation of 10 days.

Problem

The firm historically set safety stock using demand variability alone. During a recent port disruption, lead time increased to 70 days, causing hospital backorders. Management must decide whether to hold more inventory, qualify a second supplier, or redesign the network.

Analysis

The previous policy used a reorder point based on approximately 28 days of demand. It ignored the variance contribution from lead time. Moreover, the national distribution center created a concentration risk: a warehouse outage could interrupt all regions.

The analytics team develops three disruption scenarios: normal operations, port disruption, and supplier failure. It estimates scenario probabilities from external risk research and internal incident data, then simulates demand, lead times, supplier capacity, and hospital priority rules. The team evaluates expected cost, 95th-percentile shortage, fill rate, working capital, and recovery time.

Approach and solution

The recommended design is hybrid:

• retain moderate central safety stock for common components;
• qualify a second supplier in a different geographic region;
• maintain a smaller regional buffer for emergency hospital demand;
• use a dual-sourcing allocation rule that shifts volume when disruption indicators rise;
• establish a hospital-priority allocation policy before a shortage occurs;
• monitor lead-time distributions rather than only average lead time;
• conduct quarterly disruption simulations.

The second supplier has a higher unit price, so it is not used for all demand. Its option value lies in preserving supply during severe events. The regional buffer is deliberately limited because excessive decentralization would duplicate inventory.

Outcome

In simulation, the hybrid design produces higher routine procurement cost than the single-source design but substantially reduces extreme shortages and recovery time. Its business case depends on the value assigned to avoided hospital disruption, regulatory penalties, and patient-care consequences. The result illustrates why a purely expected-cost comparison can undervalue resilience.

Learning points

1. Average lead time is insufficient when its variance is large.
2. A low-probability disruption can justify flexible capacity even when it raises routine cost.
3. Service priorities must be designed before shortages occur.
4. Scenario probabilities should be challenged and updated; they are not objective facts.
5. Simulation should support, not replace, managerial and clinical judgment.

9. Implementation, Data, and Technology

A reliable stochastic inventory system requires clean transaction data, explicit units, synchronized calendars, and documented assumptions. Analysts should distinguish order date, ship date, receipt date, promised date, and available-for-use date. Censored lead-time observations—orders still outstanding—must not be treated as completed observations.

Forecasting systems may combine statistical models, machine learning, causal variables, and judgment. Machine learning can improve demand prediction when relationships are nonlinear and data are rich, but forecast accuracy alone does not determine inventory performance. A forecast with biased errors or poorly calibrated prediction intervals may produce inferior safety stock.

Useful technology practices include:

• prediction intervals rather than point forecasts;
• scenario simulation for promotions and disruptions;
• optimization with explicit capacity and service constraints;
• anomaly detection for unusual demand and lead time;
• digital twins for network stress testing;
• explainable decision rules for high-consequence products.

AI governance is essential. Models should be monitored for drift, documented for auditability, and tested for systematic service disparities. If scarce medical or humanitarian goods are allocated by an algorithm, the priority rule must be transparent, ethically defensible, and subject to human oversight. Data privacy, cybersecurity, supplier confidentiality, and regulatory traceability also matter.

10. Limitations and Best Practices

Stochastic inventory formulas are approximations. Normality may fail for intermittent, bounded, lumpy, or highly seasonal demand. Demand may be endogenous: promotions change demand, and stockouts censor observed demand. Lead times may depend on order size, congestion, or supplier workload. Costs may be nonlinear, and service failures may have reputational or safety consequences that are difficult to monetize.

Best practices include:

1. Define the service metric before calculating safety stock.
2. Use forecast-error distributions at the relevant horizon.
3. Separate routine variation from disruption scenarios.
4. Model correlation across products, regions, and suppliers.
5. Validate assumptions through back-testing and out-of-sample simulation.
6. Include capacity, minimum order quantities, and transportation constraints.
7. Report sensitivity to costs, probabilities, and lead-time assumptions.
8. Review policies after structural changes rather than relying on static parameters.
9. Combine optimization with stress testing and managerial review.
10. Treat resilience, sustainability, and fairness as design objectives, not afterthoughts.

Chapter Summary

Stochastic inventory management balances the economic costs of holding stock against the operational and strategic costs of shortages. Demand and lead-time uncertainty must be modeled jointly when appropriate. Reorder points combine expected lead-time demand with safety stock, while the variance of random lead-time demand includes both demand and lead-time components. The newsvendor model selects a one-time order quantity using the critical fractile, whereas reorder-point and base-stock policies support repeated replenishment.

Risk pooling can reduce safety stock, but correlation, transportation, and concentration risk limit its benefits. Multi-echelon systems require decisions about where inventory should be positioned and how responsibility should be coordinated. Disruptions require scenario analysis, resilience investments, and attention to tail risk rather than averages alone. Finally, effective implementation depends on data quality, calibrated forecasts, transparent governance, simulation, and continuous validation.

Key Terms

Base-stock policy: A policy that replenishes inventory position to a target level.

Backorder: Customer demand promised for later delivery rather than immediately lost.

Cycle service level: Probability of avoiding a stockout during a replenishment cycle.

Critical fractile: The target demand percentile in the newsvendor model.

Echelon inventory: Inventory at a stage plus downstream inventory assigned to that stage.

Expected shortage: Average quantity of demand not immediately supplied.

Fill rate: Fraction of demand units supplied directly from available stock.

Forecast error: Difference between actual demand and the forecast made before demand occurs.

Lead time: Time between placing an order and receiving usable supply.

Newsvendor model: A one-period stochastic model for an order made before uncertain demand is realized.

Order-up-to level: Target inventory position in a periodic-review replenishment system.

Pooling: Aggregating demand or inventory across locations to reduce relative variability.

Reorder point: Inventory-position threshold that triggers replenishment.

Resilience: Ability of a supply network to withstand, recover from, and adapt to disruption.

Safety stock: Inventory held to protect against uncertainty.

Service level: A measure of supply availability, defined according to a specified metric.

Shortage cost: Economic or operational cost caused by unavailable inventory.

Stochastic demand: Demand represented as a random variable or process.

Underage cost: Cost of ordering too little in a one-period decision.

Variance: A measure of dispersion equal to the expected squared deviation from the mean.

Review Questions

5 Short Answer Questions

1. Distinguish cycle service level from fill rate.
2. Why does random lead time increase lead-time-demand variance?
3. What is the purpose of safety stock?
4. Define the critical fractile in the newsvendor model.
5. State two advantages and two disadvantages of risk pooling.

5 Long Answer Questions

1. Explain how demand and lead-time uncertainty should be combined in a reorder-point calculation. State the assumptions behind the standard formula.
2. Compare continuous-review (Q,r) policies with periodic-review order-up-to policies.
3. Explain how correlation affects inventory pooling across regional warehouses.
4. Discuss why expected cost alone may be inadequate for disruption-risk decisions.
5. Evaluate the role of machine learning and simulation in stochastic inventory management, including governance concerns.

3 Analytical / Application Questions

1. Daily demand has mean 80 units and standard deviation 15 units. Lead time is constant at 5 days. Assuming independent daily demand and a 97.5% cycle service target, calculate the reorder point using z=1.96.
2. A seasonal product sells for 120, costs 70, and has salvage value 20. Calculate the critical fractile and explain how the selected demand percentile would change if salvage value increased to 50.
3. Two warehouses have demand standard deviations of 400 and 300 units, with correlation 0.4. Calculate aggregate demand standard deviation using the pooling formula, then discuss whether centralization is necessarily preferable operationally.

Multiple Choice Questions with Answers

Question 1

Which quantity includes on-hand inventory, on-order inventory, and backorders with the appropriate signs?

A. Inventory position

B. Cycle stock

C. Safety stock

D. Net demand

Correct Answer: A. Inventory position

Question 2

If daily demand has standard deviation σ_D and lead time is constant at L days, which expression gives lead-time-demand standard deviation under independent daily demand?

A. Lσ_D

B. σ_D/√L

C. σ_D√L

D. σ_D+L

Correct Answer: C. σ_D√L

Question 3

In the standard reorder-point model, safety stock is generally calculated as:

A. Mean demand divided by lead time

B. A service-factor quantile multiplied by lead-time-demand standard deviation

C. Order quantity minus annual demand

D. Selling price minus purchase cost

Correct Answer: B. A service-factor quantile multiplied by lead-time-demand standard deviation

Question 4

A product has selling price 80, cost 50, and salvage value 20. What is its newsvendor critical fractile?

A. 0.25

B. 0.375

C. 0.50

D. 0.625

Correct Answer: C. 0.50

Question 5

Which statement correctly distinguishes fill rate from cycle service level?

A. Fill rate measures the probability of no stockout in a cycle

B. Cycle service level measures the percentage of units supplied

C. Fill rate measures the proportion of demand units supplied immediately

D. They are always numerically identical

Correct Answer: C. Fill rate measures the proportion of demand units supplied immediately

Question 6

If two independent locations each have demand standard deviation σ, the standard deviation of aggregate demand is:

A. σ

B. 2σ

C. σ/2

D. σ√2

Correct Answer: D. σ√2

Question 7

In a periodic-review order-up-to policy, the protection period generally includes:

A. Only the review interval

B. Only supplier lead time

C. Review interval plus lead time

D. Annual demand divided by order quantity

Correct Answer: C. Review interval plus lead time

Question 8

Which term represents the cost of ordering too little in the newsvendor model?

A. Overage cost

B. Underage cost

C. Holding-rate cost

D. Echelon cost

Correct Answer: B. Underage cost

Question 9

If demand and lead time are independent, which term captures variability caused by random lead time when daily mean demand is μ?

A. μ²Var(L)

B. μVar(L)²

C. Var(L)/μ²

D. σ_D√Var(L)

Correct Answer: A. μ²Var(L)

Question 10

Which is a resilience strategy for disruption risk?

A. Relying exclusively on the lowest-cost supplier

B. Eliminating all safety stock

C. Using geographically differentiated suppliers

D. Ignoring low-frequency events

Correct Answer: C. Using geographically differentiated suppliers