Inventory under Uncertainty

Introduction

Inventory decisions are made before managers know exactly how much customers will buy, when replenishment will arrive, or whether supply will be disrupted. This uncertainty distinguishes deterministic inventory control from realistic inventory management. A policy that is economical when demand is stable may generate severe stockouts when demand or lead time fluctuates. Conversely, excessive protection against rare events ties up working capital, increases obsolescence, and consumes warehouse capacity.

The central managerial task is therefore not to eliminate uncertainty, which is usually impossible, but to select an economically and strategically appropriate protection level. This chapter develops the logic of safety stock, reorder points, service levels, shortage costs, probability distributions, lead-time variability, and risk pooling. It also explains how these ideas are implemented in continuous-review and periodic-review systems.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Distinguish cycle stock, safety stock, reorder point, and service-level measures.
2. Model demand and lead-time uncertainty using probability distributions.
3. Calculate safety stock and reorder points under common assumptions.
4. Analyse the combined effect of demand and lead-time variability.
5. Relate service levels to shortage costs and inventory investment.
6. Explain risk pooling and quantify its potential benefit.
7. Evaluate practical, technological, governance, and implementation issues.

1. The Nature of Inventory Uncertainty

Uncertainty has several sources:

• Demand uncertainty: actual demand differs from the forecast.
• Lead-time uncertainty: supplier or internal replenishment time varies.
• Supply uncertainty: quantity, quality, or availability differs from the order.
• Forecast-parameter uncertainty: estimated means and standard deviations are themselves imperfect.
• Operational uncertainty: recording errors, late transactions, shrinkage, and inaccurate inventory balances.

A useful distinction is between variability and uncertainty. Variability describes an observable distribution of outcomes; uncertainty also includes incomplete knowledge about that distribution. Statistical safety-stock formulas generally assume that the relevant distribution and its parameters are sufficiently stable.

During the replenishment lead time, inventory position declines as demand occurs. The reorder point must therefore cover expected demand during lead time plus a protection quantity. In a continuous-review system, the basic relationship is:

where ROP is the reorder point in units, μ_LT is mean demand during lead time, and SS is safety stock. The expression is conceptually general, although the method used to calculate SS depends on the uncertainty model.

Important inventory quantities

Inventory position, rather than physical on-hand stock alone, should normally trigger an order. Ignoring outstanding orders can cause systematic over-ordering.

2. Service Levels and the Meaning of Protection

“Service level” is not a single universally defined metric. Two measures are especially important.

Cycle service level (CSL) is the probability that no stockout occurs during a replenishment cycle. If demand during lead time is a random variable D_LT, then:

This is a probability and is often expressed as a percentage. A 95% CSL means that approximately 95% of replenishment cycles are expected to avoid a stockout, under the model assumptions.

Fill rate is the proportion of total demand satisfied immediately from available stock. A 95% fill rate does not necessarily mean that 95% of cycles have no stockout. A few stockout cycles can involve large shortages, while many cycles can involve only small shortages.

For a normally distributed lead-time demand, the reorder point is frequently written as:

Here, σ_LT is the standard deviation of lead-time demand and z is the standard-normal quantile associated with the desired CSL. Safety stock is therefore:

Selected z-values are approximately 1.282 for 90%, 1.645 for 95%, 2.054 for 98%, and 2.326 for 99% CSL. These values apply to a normal model and a one-sided upper protection limit.

A high service target has a nonlinear cost. Increasing CSL from 90% to 95% may require a manageable increase in safety stock, whereas moving from 98% to 99.9% can require a disproportionately large increase. Managers should therefore avoid treating service level as an arbitrary moral target; it is an economic and strategic decision.

3. Probability Models for Demand During Lead Time

3.1 Normal approximation

The normal distribution is useful when demand is the aggregate of many relatively independent effects and when negative demand is not a material concern. It is computationally convenient because the standard deviation and z-values directly determine the protection quantity.

If daily demand has mean μ_d and standard deviation σ_d, and lead time is a fixed L days, then under independent daily demand:

The square-root relationship follows because variances of independent observations add:

Thus, the fixed-lead-time reorder point is:

Units must be consistent: if μ_d and σ_d are units per day, L is in days, and ROP is in units.

3.2 Non-normal demand

Normality should not be assumed simply because a spreadsheet makes it convenient. Intermittent demand, new products, promotions, capacity limits, and highly skewed demand may require alternative approaches, such as empirical quantiles, Poisson or negative-binomial models, Croston-type methods for intermittent demand, or simulation.

For a discrete demand distribution, the reorder point can be selected as the smallest quantity r satisfying:

This quantile method is often more defensible than forcing a normal approximation onto sparse or highly skewed data. If demand is lumpy, the resulting service level may jump in substantial increments because inventory is measured in indivisible units.

4. Variable Lead Time and Combined Uncertainty

When both demand and lead time vary, multiplying average demand by average lead time gives expected lead-time demand, but not its variability. Under independence between daily demand and lead time, with mean demand μ_d, demand standard deviation σ_d, mean lead time μ_L, and lead-time standard deviation σ_L, the variance of lead-time demand is:

Equivalently:

The first term represents demand variability accumulated over the average lead time. The second represents demand exposure created by lead-time variability.

The corresponding normal-approximation policy is:

The formula assumes independent daily demand, independent lead time, a stable process, and a reasonable normal approximation for total lead-time demand. If demand and lead time are correlated, a covariance term is required:

The covariance expression is an approximation based on random-sum reasoning. Positive correlation—such as suppliers taking longer precisely when demand surges—increases protection requirements. In such cases, historical simulation or a joint probabilistic model is preferable.

Worked Example 1: Fixed lead time

A distributor sells 100 units per day on average. Daily standard deviation is 20 units. Lead time is fixed at 4 days. The target CSL is 95%.

Mean lead-time demand:

Standard deviation of lead-time demand:

With z = 1.645:

Therefore:

The system should place a replenishment order when inventory position reaches approximately 466 units. The result does not mean that demand will never exceed 466 units during lead time; it means that, under the normal model, the probability of avoiding a stockout is approximately 95%.

Worked Example 2: Variable lead time

Suppose daily demand has μ_d = 100 units and σ_d = 20 units. Mean lead time is 4 days and lead-time standard deviation is 1 day. Demand and lead time are assumed independent. For a 95% CSL:

Safety stock is:

Thus:

Compared with the fixed-lead-time example, the reorder point rises by about 111 units. Lead-time variability is substantial because average daily demand is high. Reducing lead-time variation may be more economical than holding all of the additional stock.

5. Periodic Review Systems

In a periodic-review system, inventory is checked every T time units and replenishment arrives after lead time L. Protection must cover the protection interval T + L, because the next opportunity to order may be T units away.

With fixed review interval and lead time, independent daily demand, and normal demand:

The order-up-to level S is the target inventory position after review. If current inventory position is IP, the order quantity is:

For variable lead time, replace the standard deviation term with the appropriate standard deviation of demand over the protection period. Periodic review generally requires more safety stock than continuous review because uncertainty accumulates during the review interval. It can nevertheless be attractive when ordering or physical counting is expensive, or when suppliers operate on scheduled delivery cycles.

6. Shortage Costs and the Economic Service Level

A service target should reflect the consequences of shortage. Relevant shortage costs can include:

• lost contribution margin;
• emergency transportation or procurement;
• production stoppage;
• customer penalties and compensation;
• lost future business and reputation;
• substitution or expediting costs;
• safety, health, or legal consequences.

The critical fractile provides a foundational single-period relationship. Let Cu be the underage or shortage cost per unit and Co be the overage cost per unit. The economically preferred order quantity under a single-period model satisfies:

Thus the target quantile is Cu ÷ (Cu + Co). This is the newsvendor logic and applies when excess inventory has a specified end-of-period cost and replenishment after the selling period is not relevant. It should not be mechanically applied to continuous replenishment systems, where holding, ordering, and future-period effects matter.

For a continuous-review item, a more complete analysis compares the marginal cost of one additional unit of protection with the expected reduction in shortage-related costs. Exact optimization may require a backorder, lost-sales, or multi-echelon model. The managerial principle remains: increase safety stock while the marginal benefit of improved service exceeds its marginal carrying and obsolescence cost.

Cycle service level versus fill rate

A manager who reports only CSL may conceal the magnitude of shortages. Fill rate is often more appropriate for high-volume commodities, while CSL may be more informative for critical spare parts where any stockout is highly consequential. Both should be reported with their time horizon, demand definition, and treatment of backorders clearly specified.

7. Risk Pooling and Inventory Aggregation

Risk pooling means combining uncertain demand streams so that positive and negative deviations partially offset one another. If n independent locations each have demand standard deviation σ, aggregate standard deviation is:

If each location separately holds zσ units of safety stock, total safety stock is nzσ. If demand is pooled centrally, safety stock is z√nσ. The theoretical reduction ratio is:

For n = 4, the potential reduction is 50%, assuming equal demand, independent variation, identical service targets, and no additional centralisation costs.

With correlation ρ among n equal-demand locations, aggregate standard deviation becomes:

When ρ = 0, diversification benefits are strong. When ρ = 1, all locations move together and there is no variability reduction beyond scale economies. Positive correlation caused by common weather, economic cycles, or promotions limits risk-pooling benefits.

Risk pooling can be achieved through:

• centralised inventory;
• delayed product differentiation or postponement;
• shared service parts;
• virtual pooling through inventory visibility and lateral transshipments;
• common components used across multiple products.

Pooling is not free. Transportation distance, response time, facility congestion, disruption concentration, and customer-service expectations may offset the inventory benefit. The appropriate decision is a total-cost and resilience analysis, not a safety-stock calculation alone.

8. A Practical Procedure for Setting Safety Stock

A robust implementation process is as follows:

1. Define the service measure. Decide whether the target is CSL, fill rate, or a contractual metric.
2. Segment items. Use value, criticality, demand pattern, substitution, shelf life, and supply risk—not value alone.
3. Clean the data. Correct stockouts, returns, promotions, one-time projects, unit conversions, and calendar effects.
4. Estimate demand and lead-time distributions. Calculate means, standard deviations, quantiles, and correlations where relevant.
5. Select the model. Use fixed- or variable-lead-time formulas, empirical quantiles, simulation, or a multi-echelon model as appropriate.
6. Set the reorder point. Include expected demand over the correct protection period and the selected safety stock.
7. Validate operationally. Check minimum order quantities, case packs, supplier calendars, shelf life, and review frequency.
8. Back-test. Apply the policy to historical data and measure stockouts, fill rate, inventory, and expedite costs.
9. Monitor and revise. Re-estimate parameters after structural changes rather than allowing obsolete settings to persist.

A useful policy dashboard includes forecast error, supplier lead-time percentiles, inventory accuracy, stockout frequency, fill rate, excess inventory, ageing, emergency orders, and parameter overrides.

9. Technology, Analytics, and AI

Enterprise resource planning and warehouse-management systems can calculate inventory position in near real time, but system speed cannot compensate for inaccurate master data. Important controls include unit-of-measure governance, event timestamps, supplier calendars, transaction discipline, and reconciliation of physical and system stock.

Advanced analytics can estimate empirical lead-time distributions, detect regime changes, and simulate correlated demand and supply disruption. Machine-learning forecasts may improve the mean forecast, but safety stock must still be based on the distribution of forecast errors and the operational lead-time process. A more accurate point forecast does not automatically imply a lower safety stock if residual uncertainty remains high.

AI-generated recommendations require governance. Managers should be able to inspect data sources, forecast horizons, service assumptions, overrides, and exception rules. Automated policies should not silently reduce stock for critical medical, safety, or maintenance items. Human accountability, audit trails, access controls, and scenario testing are essential.

10. Illustrative Case: MedEquip Service Parts

Background

MedEquip Services supports hospital diagnostic machines. A replacement sensor is expensive, has no practical substitute, and a stockout can keep a machine unavailable for several days. The firm currently holds identical safety stock at four regional depots.

Problem

Each depot experiences average daily demand of 10 sensors with a daily standard deviation of 4. Lead time is fixed at 5 days. Management targets a 95% cycle service level. Demand among depots is approximately independent because hospitals have different maintenance schedules.

Analysis

For one depot:

Safety stock per depot is:

Four separate depots therefore hold approximately 60 sensors of safety stock, excluding cycle stock.

If a central service-parts hub can replenish all regions without reducing effective response service, aggregate daily demand has mean 40 and standard deviation:

For five-day lead time:

Central safety stock is:

The theoretical safety-stock reduction is about 30 sensors. However, the firm must add transportation time, emergency courier capacity, and a contingency for hub disruption. If the hub adds one day of effective lead time, demand exposure and safety stock rise. If hospitals require immediate local availability, pure centralisation may violate the service objective.

Approach and solution

MedEquip adopts a hybrid design: central stock for routine demand, small local quantities for critical hospitals, and lateral transshipment between depots. It uses separate service targets based on hospital criticality, tracks actual lead-time percentiles, and reviews the correlation assumption every quarter.

Outcome

The inventory model reduces duplicated protection stock while preserving response capability. The benefit comes not from the square-root formula alone but from redesigning the service network, data controls, and escalation process.

Learning points

1. Pooling works best when demand streams are not highly correlated.
2. The relevant constraint is customer response time, not merely physical location.
3. A theoretical inventory reduction must be tested against disruption and transportation risk.
4. Criticality-based segmentation is superior to a single service target for all items.

11. Limitations, Risks, and Governance

Safety-stock formulas are decision aids, not guarantees. Common limitations include:

• nonstationary demand caused by product launches or market decline;
• biased data because stockouts censor observed demand;
• supplier lead times that are dependent on order size or capacity;
• correlations that change during disruptions;
• ignored minimum order quantities and shelf-life constraints;
• inappropriate normality assumptions;
• inaccurate inventory records;
• service metrics that reward local performance while harming total supply-chain cost.

Governance should specify who owns parameters, how overrides are approved, when models are recalibrated, and how critical items are protected. Safety-related and regulated products may require documented validation, traceability, segregation, and compliance with industry-specific requirements. Ethical inventory management also involves fair allocation during scarcity, especially for essential goods. A mathematically optimal policy for a profitable customer segment may be socially or contractually unacceptable if it systematically denies essential supplies to vulnerable users.

Best-Practice Checklist

• Match the model to the demand pattern and replenishment process.
• Use inventory position, not on-hand stock alone.
• Separate forecast uncertainty from lead-time uncertainty.
• Measure both CSL and fill rate where appropriate.
• Segment service targets by business criticality.
• Treat lead-time reduction and reliability as alternatives to holding stock.
• Validate normal assumptions and use empirical or simulation methods when needed.
• Include correlation and disruption scenarios in network decisions.
• Back-test policies and monitor parameter drift.
• Govern automated recommendations with auditability and human accountability.

Chapter Summary

Inventory under uncertainty requires protection against uncertain demand, uncertain replenishment, and imperfect information. The reorder point equals expected demand over the relevant protection period plus safety stock. Under a normal approximation and fixed lead time, safety stock is zσ_LT, where z represents the chosen cycle-service quantile. With variable lead time and independent demand, lead-time-demand variance includes both demand variability and lead-time variability: σ_LT² = μ_Lσ_d² + μ_d²σ_L².

Cycle service level measures the probability of avoiding a stockout in a cycle, whereas fill rate measures the proportion of demand supplied immediately. These measures should not be confused. Economic service levels depend on shortage, holding, obsolescence, expediting, and strategic costs. Risk pooling reduces aggregate variability when demand streams are not perfectly correlated, but centralisation can introduce transportation, responsiveness, and disruption risks. Effective practice combines statistical modelling, data governance, item segmentation, back-testing, and managerial judgement.

Key Terms

Safety stock: Inventory held to protect against uncertainty.

Reorder point: Inventory-position threshold that triggers replenishment.

Lead-time demand: Demand occurring during replenishment lead time.

Cycle service level: Probability of no stockout during a replenishment cycle.

Fill rate: Proportion of demand supplied immediately from available inventory.

Inventory position: On-hand plus on-order inventory minus backorders.

Protection period: Time for which inventory must cover uncertain demand.

Demand variability: Dispersion of demand around its expected value.

Lead-time variability: Dispersion of replenishment time around its mean.

Standard deviation: Measure of dispersion used in safety-stock calculations.

Safety factor (z): Standard-normal quantile corresponding to a target probability.

Shortage cost: Cost caused by unavailable inventory.

Overage cost: Cost of inventory remaining after the relevant demand opportunity.

Critical fractile: Underage-cost ratio used to determine a single-period target quantile.

Risk pooling: Combining demand streams to reduce relative uncertainty.

Correlation: Degree to which two uncertain variables move together.

Periodic review: Inventory policy in which stock is assessed at fixed intervals.

Continuous review: Inventory policy in which inventory position is monitored continuously.

Postponement: Delaying final product differentiation to pool component demand.

Parameter drift: Change over time in the statistical characteristics of a process.

Review Questions

Short Answer Questions

1. Distinguish safety stock from cycle stock.
2. What is the difference between cycle service level and fill rate?
3. Why does the standard deviation of demand over L independent days equal √Lσ_d?
4. Identify the two sources of variability in the variable-lead-time formula.
5. State two conditions under which risk pooling may fail to produce large benefits.

Long Answer Questions

1. Explain how a manager should select between a normal approximation, an empirical quantile, and simulation for setting a reorder point.
2. Derive and interpret the fixed-lead-time reorder-point formula, including its assumptions and units.
3. Discuss how shortage costs influence an economically appropriate service level.
4. Compare continuous-review and periodic-review inventory systems.
5. Evaluate the strategic and operational trade-offs of centralising safety stock.

Analytical / Application Questions

1. Daily demand has mean 80 units and standard deviation 15 units. Lead time is fixed at 9 days and the target CSL is 98%. Calculate mean lead-time demand, safety stock, and reorder point.
2. A firm has three independent locations, each with demand standard deviation 6 units per day. Compare the safety stock required for separate stocking with pooled stocking at the same service factor. State the assumptions behind the comparison.
3. A supplier’s average lead time is 6 days with a standard deviation of 2 days. Daily demand has mean 50 and standard deviation 10 units, with independence assumed. Calculate the standard deviation of lead-time demand and explain whether reducing lead-time variability or demand variability appears more influential.

Multiple Choice Questions with Answers

Question 1

A reorder point under a normal lead-time-demand model is generally calculated as:

A. Mean lead-time demand minus safety stock

B. Mean lead-time demand plus safety stock

C. Annual demand divided by lead time

D. Safety stock multiplied by order quantity

Correct Answer: B. Mean lead-time demand plus safety stock

Question 2

If daily demand is independent with standard deviation σ_d and lead time is fixed at L days, lead-time-demand standard deviation is:

A. Lσ_d

B. σ_d ÷ L

C. √Lσ_d

D. L²σ_d

Correct Answer: C. √Lσ_d

Question 3

For a 95% cycle service level under a normal model, the commonly used z-value is approximately:

A. 0.95

B. 1.282

C. 1.645

D. 2.326

Correct Answer: C. 1.645

Question 4

Which statement best describes fill rate?

A. Probability that lead time is zero

B. Percentage of demand supplied immediately from available stock

C. Probability that every replenishment cycle has excess stock

D. Average supplier delivery time

Correct Answer: B. Percentage of demand supplied immediately from available stock

Question 5

Daily demand has mean 100 units, standard deviation 20 units, and fixed lead time of 4 days. The standard deviation of lead-time demand is:

A. 20 units

B. 40 units

C. 80 units

D. 400 units

Correct Answer: B. 40 units

Question 6

With variable lead time and independent demand, which term captures the effect of lead-time variability in the lead-time-demand variance?

A. μ_Lσ_d²

B. μ_d²σ_L²

C. μ_dσ_dσ_L

D. σ_d ÷ μ_L

Correct Answer: B. μ_d²σ_L²

Question 7

If four identical, independent demand streams are pooled, aggregate standard deviation is how many times the standard deviation of one stream?

A. 1 time

B. 2 times

C. 4 times

D. 8 times

Correct Answer: B. 2 times

Question 8

The critical fractile in a single-period inventory model is:

A. Co ÷ Cu

B. Cu ÷ (Cu + Co)

C. Cu ÷ Co

D. (Cu + Co) ÷ Cu

Correct Answer: B. Cu ÷ (Cu + Co)

Question 9

In a periodic-review system with review interval T and lead time L, protection generally must cover:

A. T − L

B. L only

C. T + L

D. Annual demand only

Correct Answer: C. T + L

Question 10

Positive correlation among regional demands generally causes risk-pooling benefits to:

A. Increase without limit

B. Remain exactly unchanged

C. Decrease compared with independent demands

D. Become unrelated to variability

Correct Answer: C. Decrease compared with independent demands