Demand, Lead Time and Forecasting Inputs

Introduction

Inventory decisions are often presented as decisions about quantities: how much to order, when to order, and how much safety stock to hold. In practice, however, the quality of those decisions depends first on the quality of the inputs. Demand is not a single number; it is a time-dependent and often uncertain pattern. Lead time is not merely a supplier promise; it is a random interval shaped by production, transportation, customs, receiving, and information delays. Forecasts are not objective facts but structured estimates whose errors may be systematic or random.

This chapter develops the foundations needed to characterize demand and lead-time behavior and to prepare forecasting inputs for inventory control. It distinguishes independent and dependent demand, explains demand patterns such as seasonality and intermittency, evaluates forecast accuracy and bias, and shows how demand and lead-time information should be translated into inventory-relevant measures. The emphasis is on managerial judgment supported by quantitative analysis rather than on mechanical forecasting.

Learning Objectives

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

1. Characterize demand using level, trend, seasonality, variability, and intermittency.
2. Distinguish independent demand from dependent demand and select appropriate planning logic.
3. Define lead time consistently and diagnose its sources of variability.
4. Explain the difference between forecast error, forecast accuracy, and forecast bias.
5. Calculate common error measures and interpret their limitations.
6. Prepare clean, appropriately aggregated demand and lead-time data for inventory decisions.
7. Evaluate forecasting practices under seasonal, intermittent, and uncertain demand.
8. Identify managerial, technological, ethical, and governance issues in forecast-driven inventory systems.

1. Demand as an Inventory Input

Demand is the quantity requested, ordered, consumed, or shipped during a specified time interval. These alternatives are not interchangeable. Customer orders may be cancelled, shipments may be constrained by stockouts, and recorded consumption may be delayed. Consequently, analysts must define the demand event before modeling it.

Let D_t denote demand observed during period t. The period may be an hour, day, week, or month. The choice should correspond to the decision being made. Daily replenishment requires daily data, whereas a monthly production plan may require monthly data. Units must also be consistent: cases, eaches, kilograms, litres, or monetary value should not be mixed without an explicit conversion.

A useful decomposition represents observed demand as:

where L_t is the underlying level, T_t is trend, S_t is seasonal or calendar-related variation, I_t represents intermittent or event-related effects, and ε_t is irregular random variation. This is a conceptual additive decomposition, not a universal law. It is most suitable when the components are expressed in the same units and seasonal amplitude is approximately constant. A multiplicative model may be more appropriate when variability grows with the level:

The multiplicative form requires non-negative, generally positive observations and is problematic when demand contains many zeros.

Demand characteristics

Level is the typical magnitude of demand. Trend is a persistent upward or downward movement. Seasonality is a repeatable pattern linked to a calendar cycle, such as weekly, monthly, or annual behavior. Cyclicality is a longer and less regular fluctuation, often associated with economic conditions. Promotion effects arise from temporary commercial actions. Intermittency means that nonzero demand occurs infrequently, producing many zero periods. Lumpiness combines intermittent occurrence with variable nonzero quantities.

These dimensions matter because they imply different inventory risks. A stable high-volume item may be forecast with relatively low uncertainty, while a low-volume spare part may have many zeros and occasional large requirements. Treating both items with the same averaging method can create systematic overstock or stockout exposure.

Demand versus sales history

Observed sales are not always unconstrained demand. If an item was unavailable, sales may be zero even though customers wanted it. Thus:

A robust data set should identify stockout periods, allocation rules, substitutions, returns, cancellations, promotions, price changes, and unusual events. Where lost sales can be estimated, the analyst may create a censored-demand adjustment, but such estimates should be documented as assumptions rather than treated as observed facts.

2. Independent and Dependent Demand

Independent demand

Independent demand is driven primarily by external customers or end-user requirements. Examples include finished goods sold to consumers, repair parts requested by field technicians, and retail products. The demand for one finished good is not normally calculated directly from the demand for another item. Forecasting, customer intelligence, market information, and statistical analysis are therefore central.

Dependent demand

Dependent demand is derived from the requirements of another item or planned activity. If one finished product requires four components, then the component requirement depends on the production schedule and the bill of material. Under a deterministic bill of material:

If 500 assemblies are planned and each requires 3 bearings, gross bearing requirement is 1,500 bearings. Net requirements must then consider on-hand inventory, scheduled receipts, safety stock, scrap, and available inventory.

Material requirements planning (MRP) is principally a dependent-demand system. It explodes a master production schedule through bills of material and offsets requirements by lead times. Forecasting the component independently can be inappropriate because it may conflict with the actual production plan and create duplicated safety stock.

The distinction is not absolute. A component can have dependent demand in production and independent demand in the aftermarket. A service organization may need to forecast replacement parts independently even when the original equipment bill of material is known. The correct classification depends on the decision context and the source of demand.

3. Lead Time: Definition and Behavior

Lead time is the elapsed time between two explicitly defined events. A purchasing lead time might run from purchase-order release to receipt at the warehouse. A customer lead time might run from order receipt to customer delivery. A manufacturing lead time may include queueing, setup, processing, movement, inspection, and release.

A practical decomposition is:

where L is total lead time and the components represent administrative, waiting, processing, transportation, and receiving intervals. The equation is additive because the component times are measured in the same time units. Some activities overlap; if they do, the total is not the simple sum and the process definition must reflect actual elapsed time.

Lead time should be characterized by at least its mean, dispersion, minimum, maximum, and relevant percentiles. Let observed lead times be L₁, L₂, ..., L_n. The sample mean is:

and the sample standard deviation is:

Here n is the number of observations. The divisor n−1 is used for the usual unbiased estimate of population variance when observations are treated as a sample. These statistics describe historical performance; they do not guarantee future performance.

Lead-time distributions are often right-skewed: most orders arrive near the expected date, but a minority experience severe delays. In such cases, the mean and standard deviation may conceal operational risk. Percentiles, such as the 90th or 95th percentile, can be more informative for service-risk discussions. Analysts should also segment lead time by supplier, lane, order quantity, season, transport mode, and incoterm. A pooled average can hide materially different processes.

Lead-time demand

If demand per period has mean μ_D and lead time is a constant L periods, expected demand during lead time is:

This formula assumes a stable mean and consistent time units. If demand and lead time are both random, then:

provided the expected demand rate is appropriately defined and the relationship between demand rate and lead time does not introduce dependence that changes the calculation. More generally, demand during lead time should be measured or simulated directly.

A commonly used approximation for the variance of demand during random lead time is:

This assumes independent, identically distributed period demand, independent lead time, and a lead time measured in whole periods. Here σ_D² is demand variance per period and σ_L² is lead-time variance. If demand and lead time are correlated, or if demand is autocorrelated, the approximation can be materially wrong. In those situations, empirical lead-time-demand observations or Monte Carlo simulation are preferable.

4. Forecasting Inputs and Data Preparation

Forecasting begins with data design, not algorithm selection. A forecasting table should normally contain a time stamp, item or location identifier, demand quantity, unit of measure, transaction status, price or promotion indicators where relevant, stock availability, and lead-time or supply-event fields. A data dictionary should define every field and its source.

Data-quality checks

Important checks include:

• duplicate transactions and duplicate time buckets;
• missing periods versus genuine zero demand;
• unit conversions and pack-size changes;
• returns, cancellations, substitutions, and transfers;
• stockout and allocation periods;
• unusual promotion, launch, discontinuation, or disruption events;
• changes in customer, supplier, or transportation structure;
• time-zone, fiscal-calendar, and daylight-saving inconsistencies.

A missing observation is not automatically a zero. A zero is evidence that no demand was recorded under a defined process; a missing value means the process is unknown. Imputing missing values without preserving an indicator of imputation can contaminate forecast-error analysis.

Aggregation and the forecasting hierarchy

Demand exists at multiple levels: SKU–customer, SKU–warehouse, product family, region, and total business. Aggregation can reduce random variation but can also conceal local constraints. Forecasts at different levels should be coherent where decisions require it. For example, warehouse forecasts should reconcile sensibly with the total product forecast, while avoiding double counting transfers.

The appropriate time bucket depends on lead time and review frequency. If lead time is two days, monthly aggregation may be too coarse for replenishment. If demand is highly intermittent, daily aggregation may produce mostly zeros and obscure the meaningful demand process. A useful practice is to examine several granularities before selecting one.

5. Forecasting Under Different Demand Patterns

Stable and trending demand

For stable demand, a moving average may estimate the local level:

where k is the number of recent periods and F_t+1 is the forecast for the next period. A larger k smooths noise but responds slowly to change. A smaller window responds quickly but is more volatile.

Simple exponential smoothing updates the level recursively:

where 0 < α < 1 is the smoothing constant. This method is appropriate when demand has no systematic trend or seasonality. It should not be interpreted as a causal model; it is a weighted updating rule.

For trend, methods such as Holt’s linear exponential smoothing may be used. For recurring seasonal patterns, seasonal indices, regression with calendar variables, or seasonal exponential-smoothing models may be suitable. A method should be selected through time-based validation rather than in-sample fit alone.

Seasonality

Seasonality can arise from weather, holidays, school calendars, pay cycles, production shutdowns, or replenishment schedules. A seasonal index for period or season s can be conceptualized as:

For a multiplicative model, indices are often normalized so their average across a complete cycle equals 1. An index above 1 indicates demand above the overall average. Seasonal estimation requires enough historical cycles and careful treatment of changing trend. A holiday that moves between calendar weeks may require event-based variables rather than fixed week numbers.

Seasonality also affects lead time. Supplier shutdowns and port congestion may create predictable seasonal delays. Forecasting demand seasonality while assuming constant lead time can understate inventory exposure.

Intermittent demand

Intermittent demand contains many zero observations. Standard methods may produce biased or unstable forecasts because the timing of demand and the size of nonzero demand are separate processes. Croston-type approaches estimate these two components separately, while related methods modify the update procedure to address bias. Such methods are not universally superior; they should be evaluated using inventory-relevant outcomes, including stockouts and holding cost.

For very slow-moving items, it may be more useful to forecast a demand distribution or probability of demand occurrence than a precise point forecast. Classification by demand frequency and variability can guide method selection, but arbitrary thresholds should be tested against actual service and cost outcomes.

6. Forecast Error, Accuracy, and Bias

Let e_t denote forecast error. A common convention is:

Positive error means actual demand exceeded the forecast; negative error means the forecast exceeded actual demand. Some organizations use the reverse sign. The convention must be stated because bias conclusions depend on it.

Mean error, also called mean forecast error under this convention, is:

A positive ME indicates underforecasting on average; a negative ME indicates overforecasting. Bias is a directional tendency, not simply a large error.

Mean absolute error is:

It is expressed in demand units and is easy to interpret. Mean squared error is:

and root mean squared error is RMSE = √MSE. Squaring gives greater weight to large errors. These measures are appropriate when the same scale and volume are being compared.

Mean absolute percentage error is:

It is undefined when D_t = 0 and unstable when actual demand is close to zero. Weighted absolute percentage error can be written as:

WAPE is more suitable for aggregated portfolios but can be dominated by high-volume items. No single metric is universally best. Bias, service impact, cost impact, and forecast value should be considered together.

Worked example: error measures

Suppose actual demand for four weeks is 100, 120, 80, and 100 units, while forecasts are 90, 130, 100, and 95 units. Errors under the stated convention are 10, −10, −20, and 5 units.

The mean error is:

The mean absolute error is:

The squared errors are 100, 100, 400, and 25, so:

The negative ME indicates modest overforecasting on average, while MAE and RMSE describe the typical and large-error magnitudes. A manager should not conclude that the forecast is good merely because bias is near zero: positive and negative errors can cancel.

Forecast value and bias correction

A forecast can be statistically accurate but operationally poor if it ignores promotions, capacity, or service priorities. Conversely, a planner’s adjustment may improve performance when it incorporates credible market intelligence, but unrecorded overrides prevent learning and create accountability problems. Every override should record its reason, owner, date, and subsequent outcome.

Forecast bias can be monitored with cumulative error:

A persistently positive or negative cumulative error signals directional drift. A tracking signal is sometimes defined as TS = CFE ÷ MAD, where MAD is mean absolute deviation over a defined window. Because conventions and thresholds vary, tracking signals should be used diagnostically rather than as automatic triggers.

7. A Worked Inventory-Input Example

A distributor reviews a product with average weekly demand of 250 units and weekly standard deviation of 60 units. Supplier lead time averages 3 weeks with a standard deviation of 0.5 week. Assume independent weekly demand and lead time, no autocorrelation, and a stable demand rate.

Expected lead-time demand is:

Using the random-lead-time approximation:

Therefore:

The result shows that lead-time variability contributes substantially to exposure. If lead time were treated as fixed, the estimated standard deviation would be √3 × 60 ≈ 103.9 units, materially lower. The approximation is useful for preliminary analysis, but the firm should validate it against actual demand-during-lead-time observations and investigate whether delays coincide with high-demand periods.

8. Illustrative Case: Seasonal and Intermittent Service Parts

Background

The fictional company Northstar Mobility Services supports 4,000 industrial vehicles. It stocks filters, sensors, and electronic control modules at regional depots. Filters have regular seasonal demand; control modules are expensive, slow-moving, and intermittent.

Problem

Northstar’s planning system uses a twelve-month moving average for all parts. During the annual maintenance season, filters stock out despite adequate annual inventory. Meanwhile, control modules accumulate because occasional large requirements are averaged into a forecast that is too high for subsequent months. Supplier lead-time reports show an average of 28 days, but the 95th percentile is 52 days.

Analysis

The team first separates customer orders, emergency shipments, returns, and depot transfers. It identifies maintenance campaigns and marks periods in which parts were unavailable. For filters, weekly demand and calendar effects are analyzed by depot. For control modules, the team models occurrence frequency and nonzero demand size separately. Lead time is segmented by supplier and transport mode rather than pooled.

Forecast error is measured with MAE and bias, while service outcomes are assessed through fill rate and emergency-order frequency. MAPE is rejected for control modules because many weeks have zero demand. The analysis also reveals that demand surges during maintenance campaigns coincide with supplier congestion, violating the assumption that demand and lead time are independent.

Approach and solution

Northstar adopts seasonal forecasts for filters, with campaign indicators and depot-level aggregation. It uses a demand-distribution approach for control modules and sets differentiated review policies. Supplier scorecards report median, 90th-percentile, and 95th-percentile lead times. Stockout periods are flagged so that zero shipments are not automatically interpreted as zero demand. Planner overrides are documented and reviewed monthly.

Outcome

The company reduces filter stockouts during campaigns without applying the same inventory increase to every month. Control-module purchases become less frequent but more explicitly linked to risk, criticality, and repair commitments. The most important improvement is not a particular algorithm; it is the alignment of data definitions, demand segmentation, lead-time measurement, and inventory objectives.

Learning points

1. Seasonal demand should be modeled at the level where the seasonality occurs.
2. Intermittent demand requires metrics and methods that handle zeros.
3. Lead-time percentiles may reveal risk hidden by averages.
4. Stockouts censor observed demand.
5. Forecast evaluation should include operational consequences, not only statistical error.

9. Managerial, Governance, and Technology Implications

Forecasts influence purchasing, production, working capital, supplier commitments, and customer service. Managers should therefore distinguish a statistical forecast from a constrained operating plan. A sales forecast may represent unconstrained market demand, whereas an operations plan may reflect capacity, allocation, and financial priorities.

Data governance requires ownership of definitions, version control, audit trails, access controls, and documented transformations. Incentive systems can create bias: sales teams may overstate demand to secure inventory, while buyers may understate it to meet working-capital targets. Governance should assess whether forecast errors are being manipulated or merely reflecting genuine uncertainty.

Machine-learning systems can incorporate promotions, pricing, weather, web traffic, and external signals. They may improve predictive performance when data is abundant and stable, but they do not remove the need for causal reasoning, data-quality checks, and explainability. Automated forecasts can reproduce historical allocation bias or fail after structural breaks. Human review should focus on exceptional events and documented business knowledge, not on routinely overriding the model.

Privacy and security are relevant where forecasts use customer-level behavior, connected-equipment data, or location information. Data minimization, lawful use, access controls, and retention policies should be applied. Models should also be monitored for drift, especially after product changes, market disruptions, or supplier network redesign.

10. Best Practices and Limitations

Effective forecasting inputs follow several principles:

• define demand and lead-time events explicitly;
• preserve units, calendars, and time granularity;
• distinguish missing data from zero demand;
• flag stockouts, promotions, launches, and disruptions;
• segment items by demand behavior and supply risk;
• evaluate both accuracy and bias;
• use out-of-sample, rolling-origin validation;
• measure lead-time distributions, not only averages;
• reconcile forecasts with inventory and capacity decisions;
• document overrides and assumptions;
• use simulation or empirical distributions when independence assumptions fail.

Forecasts remain uncertain. Historical patterns may not persist, structural breaks may invalidate models, and aggregate accuracy may conceal poor performance for critical items. A forecast is therefore an input to a decision system, not a substitute for judgment. Inventory policies should be stress-tested against plausible demand surges, supplier delays, and data failures.

Chapter Summary

Demand and lead time are foundational inventory inputs whose definitions and data quality determine the usefulness of subsequent decisions. Demand can display level, trend, seasonality, cyclicality, promotion effects, intermittency, and lumpiness. Independent demand is driven by external requirements and is commonly forecast, whereas dependent demand is derived from parent-item plans and is commonly calculated through bills of material and MRP.

Lead time must be defined by start and end events and decomposed into its operational components. Its mean alone is insufficient; dispersion, skewness, percentiles, segmentation, and dependence with demand can materially change inventory exposure. Expected lead-time demand is straightforward under stable assumptions, but random lead time and correlated demand require more careful analysis.

Forecast error is directional, while accuracy measures describe magnitude. ME identifies bias, MAE provides unit-based typical error, RMSE emphasizes large errors, and percentage measures require caution with zeros and low-volume items. Seasonal and intermittent demand require differentiated methods and evaluation criteria. Sound practice combines statistical models, operational knowledge, governance, documented overrides, and technology used with appropriate skepticism.

Key Terms

Demand: Quantity requested, consumed, ordered, or shipped during a defined period.

Independent demand: Demand driven primarily by external customers or end users.

Dependent demand: Demand derived from the requirements of another item or planned activity.

Lead time: Elapsed time between two explicitly defined operational events.

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

Level: Typical magnitude around which demand varies.

Trend: Persistent upward or downward movement in demand.

Seasonality: Repeating variation associated with a calendar or recurring cycle.

Intermittent demand: Demand with frequent zero periods and occasional nonzero requirements.

Lumpiness: Intermittent demand combined with highly variable nonzero quantities.

Forecast: An estimate of future demand based on information available at a defined time.

Forecast error: Difference between actual demand and forecast under a stated sign convention.

Bias: Persistent tendency to overforecast or underforecast.

MAE: Mean absolute error, expressed in the original demand units.

RMSE: Root mean squared error, which gives greater weight to large errors.

MAPE: Mean absolute percentage error; problematic when actual demand is zero or very small.

WAPE: Weighted absolute percentage error based on aggregate absolute error and demand.

Stockout censoring: Distortion of observed sales because demand could not be fulfilled.

Forecast hierarchy: Multiple related demand levels, such as SKU, location, region, and total.

Cumulative forecast error: Running sum of forecast errors used to diagnose directional drift.

Review Questions

Short Answer Questions

1. Why is observed sales history not always equivalent to unconstrained demand?
2. Define independent demand and dependent demand, giving one example of each.
3. What operational activities may be included in total purchasing lead time?
4. Why can MAPE be unsuitable for intermittent demand?
5. What does a negative mean error indicate when error is defined as actual demand minus forecast?

Long Answer Questions

1. Explain how level, trend, seasonality, intermittency, and lumpiness affect inventory decisions.
2. Compare statistical forecasting for independent demand with MRP logic for dependent demand.
3. Discuss why lead-time percentiles and segmentation may be more useful than a single average lead time.
4. Explain the assumptions behind the random-lead-time variance approximation and describe situations in which it may fail.
5. Evaluate the role of human judgment, data governance, and machine learning in a modern forecasting process.

Analytical / Application Questions

1. A product has weekly mean demand of 400 units, weekly standard deviation of 80 units, mean lead time of 2.5 weeks, and lead-time standard deviation of 0.4 week. Assuming independent demand and lead time, calculate expected lead-time demand and the approximate standard deviation of demand during lead time. Interpret the result.
2. Actual demand over five periods is 50, 0, 75, 0, and 100 units, while forecasts are 40, 20, 60, 30, and 80 units. Calculate mean error and MAE using e_t = D_t − F_t. Explain why percentage accuracy measures should be treated cautiously.
3. Design a data-preparation and forecasting approach for a spare-parts portfolio containing seasonal filters, intermittent electronic modules, and items frequently affected by stockouts. Specify demand definitions, segmentation, error measures, lead-time analysis, and governance controls.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes independent demand from dependent demand?

A. Independent demand is calculated from a bill of material, whereas dependent demand is estimated from customer surveys.

B. Independent demand is driven primarily by external requirements, whereas dependent demand is derived from another item or planned activity.

C. Independent demand always has low variability, whereas dependent demand always has high variability.

D. Independent demand occurs only in manufacturing, whereas dependent demand occurs only in retail.

Correct Answer: B. Independent demand is driven primarily by external requirements, whereas dependent demand is derived from another item or planned activity.

Question 2

If forecast error is defined as actual demand minus forecast, what does a negative mean error generally indicate?

A. Systematic underforecasting

B. Systematic overforecasting

C. High forecast variability with no directional tendency

D. A data set containing no zero-demand periods

Correct Answer: B. Systematic overforecasting

Question 3

Why is MAPE often unsuitable for intermittent demand?

A. It cannot be calculated when actual demand is zero and becomes unstable near zero.

B. It always assigns zero error to a missed demand event.

C. It is measured in kilograms rather than percentage points.

D. It gives greater weight to the largest absolute errors than RMSE.

Correct Answer: A. It cannot be calculated when actual demand is zero and becomes unstable near zero.

Question 4

A component requires 4 units per parent product. If the planned parent quantity is 600 units, what is the gross component requirement before inventory and scrap adjustments?

A. 150 units

B. 604 units

C. 2,400 units

D. 2,404 units

Correct Answer: C. 2,400 units

Question 5

Weekly demand has mean 250 units and mean lead time is 3 weeks. Under stable-demand assumptions, what is expected lead-time demand?

A. 83.3 units

B. 253 units

C. 750 units

D. 1,000 units

Correct Answer: C. 750 units

Question 6

Which measure is expressed in the original demand units and averages absolute forecast errors?

A. MAE

B. MAPE

C. Seasonal index

D. Correlation coefficient

Correct Answer: A. MAE

Question 7

What is the principal inventory implication of a right-skewed lead-time distribution?

A. The median and maximum lead time must always be identical.

B. Averages may conceal a small number of severe delays, making percentiles useful.

C. Lead-time standard deviation is necessarily zero.

D. Demand forecasting is no longer required.

Correct Answer: B. Averages may conceal a small number of severe delays, making percentiles useful.

Question 8

Which data condition most directly causes observed sales to understate customer demand?

A. A documented unit conversion

B. A period in which the item was stocked out

C. A complete and unconstrained order history

D. A stable supplier transit time

Correct Answer: B. A period in which the item was stocked out

Question 9

A moving average with a larger window generally has which characteristic?

A. More responsiveness and less smoothing

B. Less smoothing and greater sensitivity to the latest observation

C. More smoothing and slower response to structural changes

D. Guaranteed elimination of forecast bias

Correct Answer: C. More smoothing and slower response to structural changes

Question 10

Which practice is most appropriate when demand and lead time are correlated?

A. Assume independence without testing because averages are sufficient.

B. Use only MAPE, regardless of zero demand.

C. Consider empirical lead-time-demand analysis or simulation.

D. Replace all missing demand observations with zero.

Correct Answer: C. Consider empirical lead-time-demand analysis or simulation.