Inventory under Uncertainty

Introduction

Inventory decisions are made before managers know exactly how much customers will buy, when replenishment will arrive, or whether supply conditions will remain stable. This uncertainty is especially important for materials managers because shortages interrupt production and customer service, whereas excess inventory consumes capital, space and working capital. Inventory under uncertainty therefore concerns the deliberate design of protection against variability rather than the elimination of variability.

The central decisions are:

• how much safety stock to hold;
• when to reorder;
• what stock position or target level to restore;
• what service level is economically justified; and
• how aggregation and risk pooling can reduce total uncertainty.

The chapter develops these decisions through probability distributions, continuous-review and periodic-review systems, the newsvendor model and risk-pooling principles.

Learning objectives

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

1. distinguish demand uncertainty, lead-time uncertainty and supply uncertainty;
2. use probability distributions and service-level measures in inventory decisions;
3. calculate safety stock and reorder points under alternative assumptions;
4. determine target stock in periodic-review systems;
5. apply the newsvendor model to single-period inventory decisions;
6. explain how risk pooling reduces aggregate safety stock;
7. evaluate the limitations, governance requirements and managerial implications of stochastic inventory models.

1. The nature of inventory uncertainty

Demand during a replenishment cycle is a random variable. Lead time may also be random, and the quantity actually received may differ from the quantity ordered. Let demand per period be D_t, lead time be L, and demand during lead time be D_L. In the simplest setting, D_L is uncertain because demand varies while lead time is fixed. In a more advanced setting, both demand and lead time vary.

Uncertainty should be distinguished from variability. Variability is observed fluctuation in a process; uncertainty is the inability to know the future outcome exactly. Forecasting reduces uncertainty but does not remove it. Safety stock is a response to residual uncertainty, not a substitute for sound forecasting, supplier development or process control.

Important sources include:

• Demand variability: random customer requirements, promotions, seasonality and forecast error.
• Lead-time variability: transport delays, production queues, customs and supplier unreliability.
• Supply variability: defects, short shipments, capacity constraints and disruptions.
• Parameter uncertainty: imperfect estimates of average demand, standard deviation or lead-time distribution.
• Structural uncertainty: changes in markets, technology, regulations or product design.

A stochastic inventory policy must specify the review system, replenishment lead time, demand period, unit of measurement and service objective. A calculation based on weekly demand cannot be combined directly with a lead time expressed in months.

2. Probability distributions and service concepts

2.1 Distributional representation

A probability distribution describes possible values and their probabilities. The normal distribution is often used for aggregate demand because the sum of many relatively independent effects can be approximately normal. It is convenient, but not universally appropriate. Intermittent demand may require Croston-type methods or count distributions; highly seasonal demand requires time-dependent models; and demand bounded below by zero may be poorly represented by a normal distribution.

For a normal random variable with mean μ and standard deviation σ, a protection quantity can be expressed as μ + zσ, where z is the standard-normal quantile associated with the desired probability. The use of z presumes that the relevant demand distribution is reasonably approximated by a normal distribution and that its parameters are estimated from representative data.

2.2 Cycle service level and fill rate

Two service measures must not be confused.

• Cycle service level (CSL): the probability of experiencing no stockout during a replenishment cycle.
• Fill rate: the proportion of total demand satisfied immediately from available stock.

A 95% cycle service level does not necessarily mean a 95% fill rate. A single stockout can be small or very large, so the frequency of stockouts and the quantity short are different dimensions of performance.

If demand during lead time is normally distributed with mean μ_LT and standard deviation σ_LT, and a continuous-review system seeks cycle service level α, then the reorder point is:

Here, ROP is the reorder point in units, z_α satisfies P(D_LT ≤ ROP) = α, and D_LT is lead-time demand. The first term is expected lead-time demand; the second is safety stock.

The units are consistent: both μ_LT and σ_LT are units of material, so SS is also units of material. If a discrete distribution is used, the reorder point should be selected from its cumulative distribution rather than by blindly applying a normal quantile.

2.3 Cost-based service levels

Service should be selected economically, not automatically maximized. The marginal cost of increasing protection should be compared with the marginal expected cost of shortages, including expediting, lost contribution, downtime, penalties and reputational damage. Holding costs include capital, storage, insurance, deterioration and obsolescence.

A high service level may be justified for a critical aircraft component but not for a low-value, easily substituted office item. Criticality, substitutability, demand volatility and replenishment responsiveness should therefore supplement purely statistical analysis.

3. Safety stock with demand and lead-time variability

3.1 Fixed lead time and variable demand

Suppose demand per period is independent, with mean μ_D and standard deviation σ_D, and lead time is fixed at L periods. Lead-time demand has:

The square-root relationship assumes independent demand across periods with constant variance. It follows because the variance of a sum of independent demands is the sum of their variances:

Thus the reorder point is:

3.2 Variable lead time and fixed-period demand

If demand per period is constant at d but lead time is random with mean μ_L and standard deviation σ_L, then:

This special case treats demand as deterministic and lead-time uncertainty as the only source of variation.

3.3 Both demand and lead time variable

Under independence between demand and lead time, and using the common approximation based on mean lead time:

The corresponding reorder point is:

The formula assumes demand observations are independent across periods, lead time is independent of demand, demand and lead-time parameters are stable, and a normal approximation is acceptable. If demand and lead time are correlated, the variance requires a covariance term. A more exact random-sum model may also be required:

under the same independence assumptions. The expression is therefore not a universal rule; it is a model-dependent result.

3.4 Worked example: continuous review

A component has average weekly demand of 500 units and weekly standard deviation of 120 units. Supplier lead time averages 3 weeks with a standard deviation of 1 week. Demand and lead time are assumed independent. The desired cycle service level is 97.5%, for which z = 1.96.

Expected lead-time demand is:

μ_LT = 500 × 3 = 1,500 units.

Lead-time-demand standard deviation is:

σ_LT = √(3 × 120² + 500² × 1²)

σ_LT = √(43,200 + 250,000) = √293,200 ≈ 541.5 units.

Safety stock is:

SS = 1.96 × 541.5 ≈ 1,061.3 units.

Therefore:

ROP = 1,500 + 1,061.3 ≈ 2,561.3 units.

A practical policy would round upward to 2,562 units, subject to pack-size and operational constraints. The calculation shows that lead-time variability is substantial: eliminating it would reduce the standard deviation to 120√3 ≈ 207.8 units. Supplier reliability may therefore reduce inventory more effectively than simply increasing the forecast.

4. Reorder points, order-up-to levels and target stock

4.1 Continuous-review, fixed-order-quantity systems

In a continuous-review system, inventory position is monitored continuously. The inventory position is:

An order of fixed quantity Q is placed when IP ≤ ROP. The reorder point protects against demand during lead time. The economic order quantity, when its assumptions apply, is:

where D is annual demand in units per year, S is ordering cost per order, and H is annual holding cost per unit. In uncertain environments, Q* and the reorder point solve different problems: Q* addresses replenishment-cycle economics, while ROP addresses protection against uncertainty.

4.2 Periodic-review systems

In a periodic-review system, stock is reviewed every P periods. Since the next review may occur just after a replenishment decision, protection is needed over the protection interval P + L. The order-up-to level is:

For fixed lead time L and independent demand:

The order quantity at review is:

where IP is the inventory position at review. This is often called an order-up-to or target-stock policy. It differs from a continuous-review reorder point because it covers both review-period demand and lead-time demand.

4.3 Worked example: target stock

Daily demand has mean 80 units and standard deviation 20 units. Reviews occur every 5 days, lead time is 4 days, and the desired cycle service level is 95%, so z = 1.645. The protection interval is 9 days.

Expected demand is:

μ₉ = 9 × 80 = 720 units.

Standard deviation is:

σ₉ = 20√9 = 60 units.

Therefore:

S = 720 + 1.645 × 60 = 818.7 units.

The target stock is approximately 819 units. If inventory position at review is 310 units, the order is 819 − 310 = 509 units, before considering minimum order quantities or rounding.

5. The newsvendor model

The reorder-point model is suitable for repeated replenishment. The newsvendor model addresses a single selling opportunity or a short life cycle, such as fashion goods, newspapers, fresh food or seasonal promotional products.

Let underage cost C_u be the cost of ordering one unit too few, and overage cost C_o be the cost of ordering one unit too many. The critical fractile is:

where F(Q) is the cumulative distribution function of demand and Q* is the optimal order quantity. Equivalently, choose Q* such that the probability demand is no greater than Q* equals the critical fractile.

For normally distributed demand with mean μ and standard deviation σ:

where z is the quantile corresponding to C_u/(C_u + C_o). The model assumes one ordering opportunity, known costs, no replenishment during the selling period and a specified treatment of leftover stock. If salvage value v exists, overage cost commonly becomes C_o = c − v, where c is purchase cost, while underage cost may include lost margin and other consequences.

Worked example: seasonal material

A firm must order a seasonal packaging material. Demand is normal with mean 10,000 units and standard deviation 2,000 units. The underage cost is ₹30 per unit and the overage cost is ₹10 per unit. The critical fractile is:

30/(30 + 10) = 0.75.

The 75th percentile of the standard normal distribution is approximately z = 0.674. Thus:

Q* = 10,000 + 0.674 × 2,000 = 11,348 units.

The result is not a 75% fill rate. It is the quantity at which the marginal expected cost of under-ordering equals the marginal expected cost of over-ordering under the model.

6. Risk pooling and inventory aggregation

Risk pooling combines uncertain demand across locations, products or time periods. If demands are independent, aggregation reduces relative variability because standard deviations grow more slowly than means. For n identical independent locations, each with mean μ and standard deviation σ, total mean and standard deviation are:

A decentralized system with safety stock nzσ can therefore require more total safety stock than a centralized system with safety stock zσ√n. The risk-pooling benefit is:

This ideal result assumes independence, identical distributions, equal service targets and no additional centralisation cost. With correlation ρ among identical location demands:

When ρ = 0, pooling is strongest; when ρ = 1, demand moves together and pooling provides no variance reduction relative to simply adding location-level standard deviations. Negative correlation can produce even greater benefits, though it may be difficult to sustain.

Risk pooling can be physical, informational or virtual. A central warehouse is physical pooling; postponement and common components are product pooling; inventory visibility and flexible allocation are informational pooling. Pooling can increase transport distance, response time, handling cost and vulnerability to a central disruption, so total landed cost and resilience must be evaluated.

7. Applied case: pooling service parts in an industrial network

Background

The following is an illustrative case. An equipment manufacturer supplies a common service part from three regional warehouses. Each region has weekly mean demand of 400 units and standard deviation of 100 units. Lead time from the factory is two weeks, assumed fixed. Each warehouse independently targets a 95% cycle service level, with z = 1.645.

Problem

Management is considering a central distribution centre. The objective is to estimate the statistical inventory effect while recognising that customer response, transport cost and disruption exposure may change.

Analysis

For one warehouse, two-week demand has mean:

μ_LT = 2 × 400 = 800 units.

Its standard deviation is:

σ_LT = 100√2 ≈ 141.4 units.

Safety stock per warehouse is:

SS = 1.645 × 141.4 ≈ 232.6 units.

Three warehouses therefore hold approximately 3 × 232.6 = 697.8 units of safety stock.

If demand is pooled centrally and independent across regions, aggregate two-week demand has standard deviation:

σ_pool = 141.4√3 ≈ 244.9 units.

Central safety stock is:

SS_pool = 1.645 × 244.9 ≈ 402.9 units.

The statistical reduction is approximately 697.8 − 402.9 = 294.9 units, or about 42.3% of decentralized safety stock.

Approach and solution

The firm should not implement centralisation solely on this calculation. It should compare the inventory saving with incremental transport, emergency delivery, facility, handling and information-system costs. It should test regional correlations, demand segmentation and critical-part requirements. A hybrid policy may be preferable: common, predictable parts are pooled centrally, while critical or urgent parts remain regionally protected.

Outcome and learning points

The analysis demonstrates the square-root effect and the importance of correlation. It also shows why service level must be measured at the customer-response point, not only at the warehouse. Pooling is a system design decision involving inventory, logistics and resilience rather than a purely mathematical reduction in stock.

8. Implementation, technology and governance

A robust implementation follows a disciplined sequence:

1. define the service measure and customer consequence of shortage;
2. cleanse transaction, demand, lead-time and backorder data;
3. segment items by value, criticality, volatility and lifecycle;
4. estimate distributions and test whether normality is reasonable;
5. calculate parameters using the same time units;
6. simulate or back-test candidate policies;
7. account for minimum order quantities, pack sizes, shelf life and capacity;
8. monitor actual service, bias, stockouts, excess and parameter drift;
9. review policies when suppliers, products or market conditions change.

Enterprise resource planning and advanced planning systems can automate inventory-position calculations, quantile estimation and exception alerts. Machine-learning forecasts may improve point forecasts, but a forecast model does not automatically produce a correct safety stock. Forecast error distributions, bias, intermittent demand and changing lead times must be measured. AI-generated recommendations should be explainable, version-controlled and subject to human approval, especially for critical materials.

Governance issues include data ownership, auditability, override controls and equitable allocation during shortages. When supply is constrained, allocation rules should be transparent and aligned with contractual, safety and public-interest obligations. Deliberately starving one customer group to protect reported service elsewhere can distort performance measures and create ethical as well as commercial risk.

9. Limitations and advanced considerations

Stochastic formulas can create false precision. Standard deviations estimated from short or unstable histories may be unreliable. Promotions, product introductions and disruptions violate stationarity. A normal model can assign positive probability to negative demand and can understate tail risk. For high-consequence items, empirical quantiles, bootstrap methods, Monte Carlo simulation or scenario analysis may be more appropriate.

Demand and lead time may be correlated: congestion can delay replenishment precisely when demand rises. In such cases, the independence formula understates risk. Multi-echelon systems also require placement decisions: safety stock at a factory, central warehouse and local warehouse cannot be optimised independently because service depends on the network.

Managers should distinguish expected service from guaranteed service. A 99% target still permits occasional failures, and a service target based on historical averages may not protect against structural breaks. Sensitivity analysis should examine the effect of changes in μ, σ, lead time, correlation and shortage cost. The most valuable intervention may be process improvement—shorter and more reliable lead time—rather than holding more stock.

Chapter Summary

Inventory under uncertainty uses probability and economics to protect operations against variable demand, lead time and supply. Safety stock is the protection component of a reorder point or order-up-to level. With fixed lead time and independent demand, lead-time standard deviation grows with the square root of lead time. When both demand and lead time vary, the variance must reflect both sources and their assumptions. Continuous-review systems reorder at a reorder point, whereas periodic-review systems restore inventory position to a target covering the review period plus lead time. The newsvendor model selects a one-time order quantity using the critical fractile formed from underage and overage costs. Risk pooling reduces aggregate relative variability when demands are not perfectly correlated, but centralisation introduces logistics, resilience and governance trade-offs. Effective practice combines statistical models with segmentation, data quality, simulation, supplier improvement and continuous monitoring.

Key Terms

Safety stock: Inventory held to protect against uncertainty.

Reorder point: Inventory-position threshold that triggers replenishment.

Lead-time demand: Demand occurring while an order is being replenished.

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

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

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

Target stock: Desired inventory position after a periodic-review order.

Protection interval: Review period plus replenishment lead time.

Demand variability: Fluctuation in quantities requested over time.

Lead-time variability: Fluctuation in replenishment duration.

Normal distribution: Continuous probability distribution defined by mean and standard deviation.

Quantile: Value below which a specified cumulative probability lies.

Critical fractile: Cost-based probability ratio used in the newsvendor model.

Underage cost: Economic cost of ordering less than realised demand.

Overage cost: Economic cost of ordering more than realised demand.

Risk pooling: Aggregation of uncertainty to reduce relative variability.

Correlation: Degree to which two random quantities move together.

Continuous review: Inventory monitoring conducted continuously.

Periodic review: Inventory assessed at fixed time intervals.

Parameter uncertainty: Uncertainty about estimated means, variances or distributions.

Review Questions

Short Answer Questions

1. Distinguish cycle service level from fill rate.
2. What does safety stock protect against?
3. State the assumptions behind the square-root-of-lead-time relationship.
4. Why does a periodic-review system protect over P + L rather than only L?
5. Define the critical fractile in the newsvendor model.

Long Answer Questions

1. Explain how demand and lead-time variability jointly affect the reorder point. Include assumptions and formula interpretation.
2. Compare continuous-review and periodic-review inventory systems in terms of information requirements, protection interval and ordering logic.
3. Explain risk pooling using variance and correlation, and discuss when centralisation may fail to deliver practical benefits.
4. Discuss why service-level selection should incorporate shortage economics and item criticality rather than rely on a universal percentage.
5. Evaluate the limitations of normal-distribution safety-stock models in modern supply networks.

Analytical / Application Questions

1. Daily demand has mean 120 units and standard deviation 30 units. Lead time is fixed at 6 days and the desired cycle service level is 97.5% (z = 1.96). Calculate safety stock and reorder point.
2. A periodic-review item has mean weekly demand of 700 units, standard deviation 140 units, a review period of 2 weeks and lead time of 3 weeks. For a 95% service level (z = 1.645), calculate the target stock assuming independent demand.
3. Three independent regional demands each have standard deviation 180 units over the protection interval. Compare total decentralized safety stock with pooled safety stock at z = 2.05, ignoring differences in mean demand. Interpret the result managerially.

Multiple Choice Questions with Answers

Question 1

Which expression represents a continuous-review reorder point when lead-time demand is normally distributed?

A. ROP = μ_LT + z_ασ_LT

B. ROP = μ_D + z_ασ_D

C. ROP = Q*/H

D. ROP = C_u/(C_u+C_o)

Correct Answer: A. ROP = μ_LT + z_ασ_LT

Question 2

If weekly demand is independent with standard deviation σ_D and fixed lead time is L weeks, what is the standard deviation of lead-time demand?

A. σ_D/L

B. Lσ_D

C. σ_D√L

D. σ_D²L

Correct Answer: C. σ_D√L

Question 3

A periodic-review system with review period P and lead time L must generally protect against demand over which interval?

A. Only P

B. Only L

C. P − L

D. P + L

Correct Answer: D. P + L

Question 4

A 95% cycle service level means that:

A. 95% of all demand units will always be filled

B. The probability of no stockout in a replenishment cycle is 95%

C. Inventory turnover must be 95%

D. Lead time will be reduced by 95%

Correct Answer: B. The probability of no stockout in a replenishment cycle is 95%

Question 5

In the newsvendor model, if underage cost is ₹20 and overage cost is ₹10, the critical fractile is:

A. 0.33

B. 0.50

C. 0.67

D. 0.90

Correct Answer: C. 0.67

Question 6

Three identical independent locations each have standard deviation 100 units. The standard deviation of their aggregate demand is:

A. 100 units

B. 173.2 units approximately

C. 300 units

D. 10,000 units

Correct Answer: B. 173.2 units approximately

Question 7

Which condition produces the strongest basic risk-pooling benefit?

A. Perfect positive correlation among locations

B. Zero correlation among locations

C. Identical demand at every location on every day

D. A common disruption affecting all locations

Correct Answer: B. Zero correlation among locations

Question 8

In the inventory-position equation, backorders are:

A. Added to on-hand inventory

B. Ignored because they are not physical stock

C. Subtracted from on-order inventory only

D. Subtracted from on-hand plus on-order inventory

Correct Answer: D. Subtracted from on-hand plus on-order inventory

Question 9

For demand mean 500 units per week, fixed lead time 4 weeks, and demand standard deviation 100 units per week, the approximate lead-time-demand standard deviation is:

A. 100 units

B. 200 units

C. 400 units

D. 50 units

Correct Answer: B. 200 units

Question 10

Which statement about AI-supported safety-stock recommendations is most accurate?

A. A better point forecast automatically determines correct safety stock

B. AI eliminates the need to define a service measure

C. Forecast-error distributions, data quality and governance remain necessary

D. AI makes correlation and lead-time variability irrelevant

Correct Answer: C. Forecast-error distributions, data quality and governance remain necessary