Demand Analysis and Inventory Forecasting

Introduction

Inventory decisions are fundamentally decisions about uncertain future demand. A warehouse must hold enough stock to protect customer service, yet not so much that capital, space, obsolescence, damage and insurance costs become excessive. Demand analysis provides the evidence needed to understand how customers consume products; inventory forecasting converts that evidence into an informed estimate of future requirements.

Forecasting does not eliminate uncertainty. Its purpose is to reduce avoidable uncertainty, quantify the uncertainty that remains, and support decisions such as replenishment, safety-stock setting, capacity planning, purchasing and workforce scheduling. A professional forecast is therefore not simply a number. It is a documented process involving data, assumptions, statistical methods, managerial judgment, collaboration and continuous review.

Learning Objectives

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

1. Distinguish independent and dependent demand and identify common demand patterns.
2. Prepare demand data for forecasting and recognize data-quality problems.
3. Select and apply qualitative, time-series and causal forecasting methods.
4. Calculate and interpret forecast-error measures.
5. Explain the relationship between forecasts, reorder points and safety stock.
6. Apply collaborative planning principles across supply-chain partners.
7. Evaluate the limitations, ethical issues and governance requirements of forecasting systems.

1. Demand in the Supply Chain

Demand is the quantity of a product or service that customers require during a specified period. It should be distinguished from sales. Sales may be lower than true demand when an item is out of stock, subject to allocation, unavailable in a particular channel or constrained by a credit policy. If lost sales are treated as zero demand, the forecasting system may systematically underestimate future requirements.

Independent and dependent demand

Independent demand originates outside the inventory system and is influenced by customers, markets, competitors and economic conditions. Finished goods sold to customers normally have independent demand.

Dependent demand is derived from the requirement for another item. For example, the demand for tyres, batteries and packaging may depend on the planned production of vehicles or consumer products. Dependent demand is usually forecast through a master production schedule, bill of materials or material-requirements-planning system rather than by forecasting each component independently.

Demand, inventory position and demand interval

A demand observation must be associated with a time interval and a unit of measure. Daily units, weekly cases and monthly pallets are not interchangeable. The forecast horizon should cover the period for which decisions are being made, including supplier lead time where replenishment is involved.

Inventory position is commonly defined as:

The equation assumes that on-order inventory is expected to arrive and that backorders represent committed requirements. Inventory position, rather than physical stock alone, is often the relevant quantity for replenishment decisions.

2. Demand Patterns and Data Preparation

Demand analysis begins with a time-ordered demand history. The analyst should inspect the data before selecting a method.

Principal demand patterns

Random variation occurs in all series. It is not itself a pattern that can be forecast reliably. A useful decomposition is conceptual:

This additive form is most suitable when the size of seasonal fluctuations is approximately constant. A multiplicative interpretation is more suitable when seasonal variation increases in proportion to the level:

These are modelling perspectives rather than universal identities. The analyst must examine the data and validate the chosen representation.

Data preparation

Important preparation activities include:

• defining the product, location, channel and time bucket;
• correcting duplicate transactions, returns and cancelled orders;
• separating genuine demand from stockout-constrained sales;
• recording promotions, price changes, product substitutions and product launches;
• aligning calendars, time zones and fiscal periods;
• identifying discontinued, obsolete and newly introduced items;
• reconciling units of measure, pack sizes and product hierarchies; and
• preserving an audit trail of every adjustment.

Outliers should not be deleted automatically. A one-time emergency order may be a special event, whereas a recurring peak may be valid seasonal demand. Each adjustment requires a documented business explanation.

Aggregation and disaggregation

Forecasting at SKU-location level can be noisy, while forecasting only at total-company level can conceal important local patterns. A practical hierarchy may include product family, product, location and customer channel. Forecasts should be coherent where possible: the sum of lower-level forecasts should reconcile with the higher-level forecast. This is known as hierarchical forecast reconciliation.

3. Qualitative Forecasting Methods

Qualitative methods are especially useful when historical data are limited, the product is new, structural change is occurring or reliable external knowledge is important.

Executive opinion and expert judgment

Senior managers may provide a rapid forecast based on market knowledge. Its advantage is speed and access to strategic information. Its weaknesses include optimism bias, anchoring, political influence and inconsistent assumptions. A documented assumptions log and independent challenge improve reliability.

Sales-force composite

Sales representatives estimate demand in their territories, and the estimates are aggregated. Sales personnel may understand customer intentions well, but they can also understate demand to create achievable targets or overstate it to obtain inventory and production support. Incentives should therefore be designed carefully.

Market research and customer surveys

Interviews, pilot sales, conjoint studies and surveys can reveal preferences and purchase intentions. Stated intention is not always actual purchase behaviour, so survey evidence should be combined with transaction data where possible.

Delphi method

The Delphi method obtains independent estimates from a panel of experts through several anonymous rounds. A coordinator summarizes the responses and returns the summary for reconsideration. Anonymity reduces the influence of dominant personalities. The method is useful for long-range or technologically uncertain decisions, but it can be time-consuming and remains dependent on the quality of the panel.

Scenario analysis

Scenario analysis develops internally consistent views such as base case, high-demand and low-demand outcomes. It is not a substitute for a probability forecast, but it helps managers prepare capacity, inventory and supplier responses for plausible conditions.

Qualitative judgment should not be treated as inherently superior to statistical forecasting. A sound process normally begins with a statistical baseline and applies documented judgment only where there is credible information not captured in the data.

4. Time-Series Forecasting

Time-series methods use the historical relationship between a variable and time. They are most useful when the future is expected to resemble the past sufficiently for historical patterns to remain informative.

Naïve forecast

For a stable series, the next-period forecast may equal the most recent observation:

where F_t+1 is the forecast for period t+1 and A_t is actual demand in period t. The method is simple and provides an important benchmark. For seasonal demand, a seasonal naïve forecast may use the observation from the corresponding previous season:

where m is the number of periods in one seasonal cycle and h is the forecast horizon.

Simple moving average

A moving average of the last n observations is:

All observations receive equal weight. A larger n smooths random variation but responds more slowly to genuine change. The method assumes that recent history is representative and that no strong trend or seasonality requires explicit treatment.

Weighted moving average

A weighted moving average assigns different weights:

The weights must satisfy w_i ≥ 0 and ∑w_i = 1. Greater weight on recent observations increases responsiveness but may increase sensitivity to noise.

Simple exponential smoothing

Simple exponential smoothing updates the forecast as follows:

where 0 < α < 1, A_t is actual demand and F_t is the forecast for period t. A high value of α reacts quickly to new observations; a low value produces more smoothing. Simple exponential smoothing is appropriate for level demand without systematic trend or seasonality.

#### Worked example: exponential smoothing

Suppose actual demand in period 1 is 120 units, the period-1 forecast is 100 units, and α = 0.30. The period-2 forecast is:

If actual demand in period 2 is 110 units, the period-3 forecast is:

The result may be retained as 107.2 for calculation or rounded according to the inventory system’s policy. Rounding should not create a false impression of precision.

Trend-adjusted smoothing

When a series has trend, a level-only method tends to lag. Holt’s method separately estimates level and trend. One common formulation is:

Here L_t is the estimated level, T_t is the estimated trend, α and β are smoothing constants between 0 and 1, and h is the number of periods ahead. The method assumes that a reasonably stable trend exists and may perform poorly when a trend reverses abruptly.

Seasonal forecasting

Seasonal indices express typical demand in a season relative to an overall level. In a multiplicative model, a seasonal forecast can be represented as:

The indices should normally average approximately 1 across a complete seasonal cycle. Seasonal factors must be estimated from sufficient historical cycles and reviewed when market conditions change.

Intermittent demand

For spare parts and slow-moving items, many periods may have zero demand. A simple average can produce poor timing and biased replenishment decisions. Intermittent-demand methods, such as Croston-type approaches and their variants, separately estimate the non-zero demand size and the interval between non-zero demands. These methods require careful implementation because variants differ in bias and treatment of obsolescence. For highly intermittent or obsolete items, lifecycle policy, criticality and service consequences may matter more than a point forecast.

5. Causal Forecasting

Causal methods relate demand to explanatory variables such as price, promotions, weather, economic activity, installed equipment, competitor actions or customer population. They are valuable when external drivers are known and can be forecast or controlled.

A basic linear model is:

where D_t is demand, X_t is an explanatory variable, a is the intercept, b is the estimated change in demand for a one-unit change in X_t, and ε_t is the unexplained error. Units must be consistent: if X_t is price in currency per unit, b has units of product units per currency unit.

With several drivers:

A coefficient indicates association after accounting for the other variables; it does not automatically prove causation. Analysts should consider omitted variables, reverse causality, multicollinearity, autocorrelated errors and structural breaks. A promotion variable may be correlated with high-demand periods because managers promote products when they already expect demand to rise.

Regression is useful only when future explanatory variables are available or can be estimated. A model that explains history accurately may still forecast poorly if the future price, weather or economic environment is unknown.

6. Forecast Accuracy and Error Measurement

Let the forecast error be:

Positive error means actual demand exceeded the forecast; negative error means the forecast exceeded actual demand. The sign convention must be stated because some organizations define error in the reverse direction.

Common measures

Here n is the number of evaluated periods. MAE and RMSE have the same units as demand. MAPE is undefined when A_t = 0 and can become unstable when actual demand is very small. Alternatives include weighted absolute percentage error or scaled measures such as mean absolute scaled error, depending on the application.

#### Worked example: error measures

Actual demand for three periods is 100, 120 and 80 units. Forecast demand is 90, 130 and 100 units. Errors are 10, −10 and −20 units. Therefore:

The negative ME indicates net overforecasting under the stated convention, while MAE describes typical absolute error. No single measure is sufficient for every decision.

Bias and tracking

A forecast may have acceptable average absolute error but persistent bias. Cumulative error can be monitored through a tracking signal, commonly expressed as:

where MAD is mean absolute deviation, often calculated as MAE over the monitoring window. The measure is useful as a warning indicator, not as an automatic proof that a model must be replaced. Control limits should be established according to the organization’s process and demand characteristics.

Validation and model selection

Forecasts should be tested using a time-ordered holdout period or rolling-origin evaluation. Randomly mixing future observations into the training data creates leakage and gives an unrealistically favourable result. Models should be compared against a naïve benchmark and assessed using measures relevant to the inventory decision. A small statistical improvement may not justify greater complexity, poor explainability or higher maintenance cost.

7. From Forecasts to Inventory Decisions

A demand forecast is an input to, not a replacement for, inventory policy. For a continuous-review system, a basic reorder point is:

If average demand per period is d and lead time is L periods, expected lead-time demand is dL, assuming demand and lead time are sufficiently stable and independent for this approximation.

When lead-time demand is approximately normal and a cycle-service target is selected, a common safety-stock approximation is:

where z is the standard-normal value associated with the desired cycle-service level and σ_LT is the standard deviation of demand during lead time. If independent demand has standard deviation σ_d per period and lead time is a fixed L periods:

If both demand and lead time vary and are treated as independent, a commonly used approximation is:

where σ_L is the standard deviation of lead time. This approximation assumes the stated distributions, independence and stable parameters. It may be unsuitable for correlated demand and lead time, promotions, intermittent demand or severe non-normality. Service level must also be defined carefully: cycle service and fill rate are not identical.

Forecast error should be included in safety-stock analysis. Using only historical demand variability may understate uncertainty when the forecasting process itself is inaccurate. The appropriate error variance depends on the forecasting model, lead time and review policy.

8. Collaborative Planning and Forecasting

Supply-chain partners often possess different information. A retailer sees point-of-sale demand, a manufacturer understands production constraints and a supplier knows capacity and material risks. Collaborative planning seeks to combine these perspectives through structured processes such as collaborative planning, forecasting and replenishment (CPFR), vendor-managed inventory and sales-and-operations planning.

A robust collaborative process typically includes:

1. agreeing on product definitions, calendars and data ownership;
2. sharing point-of-sale, inventory, promotion and shipment data;
3. producing an agreed statistical baseline;
4. documenting events, assumptions and overrides;
5. reconciling forecasts across product, location and time hierarchies;
6. translating the consensus forecast into replenishment and capacity plans; and
7. reviewing forecast accuracy, bias and service outcomes.

Collaboration does not mean that every partner receives unrestricted data. Commercial confidentiality, cybersecurity, access controls and data-retention policies are necessary. Partners should also agree how forecast changes, allocation decisions and exceptional demand will be communicated.

The distinction between demand and supply plans is important. A consensus demand forecast represents expected customer requirements; a constrained supply plan represents what can actually be produced, transported and delivered. Treating a supply constraint as a reduction in demand can conceal service problems and distort future forecasts.

9. Technology, AI and Governance

Enterprise resource planning, warehouse-management and advanced-planning systems can automate data collection, forecast generation, exception alerts and replenishment calculations. Machine-learning methods may detect nonlinear relationships, promotion effects and high-dimensional patterns. They are not automatically superior to simpler methods. Performance depends on data quality, stable processes, appropriate validation and operational explainability.

AI-supported forecasting should be governed through:

• documented data lineage and model purpose;
• protection against leakage and inappropriate use of confidential data;
• monitoring for drift, bias and degradation;
• human review of unusual forecasts and overrides;
• version control, audit trails and reproducible calculations; and
• clear accountability for inventory and service decisions.

Ethical issues include unfair allocation of scarce products, discriminatory service outcomes caused by biased data, excessive surveillance of customers or employees and the environmental cost of overproduction and expedited transport. A forecast is not neutral when it determines who receives limited inventory.

10. Illustrative Case: Meridian Medical Supplies

Background

Meridian Medical Supplies distributes sterile procedure kits to hospitals. Demand is weekly. The company historically used a three-week moving average for every item. A major hospital contract introduced periodic volume surges, while occasional supplier delays created stockouts.

Problem

The warehouse reported frequent stockouts during contract-related peaks and excess stock immediately afterward. Management initially proposed increasing the moving-average window, but this would have made the forecast even less responsive.

Analysis

The analyst first reconstructed unconstrained demand by identifying weeks in which orders were rationed. Contract start dates and scheduled procedures were added as event information. The baseline forecast used exponential smoothing for ordinary weeks, while contract volumes were planned from the hospital’s procedure schedule. Forecasts were evaluated through rolling-origin testing against a seasonal-naïve benchmark. Bias and MAE were reported separately for ordinary and event weeks.

The replenishment review showed that lead time varied from two to four weeks. Therefore, safety stock based only on demand variation was inadequate. Supplier lead-time data were incorporated, and critical kits received a higher service target than non-critical ancillary products.

Approach and solution

Meridian introduced a monthly collaborative meeting with the hospital and supplier. The hospital shared its procedure calendar and revisions; Meridian separated demand forecasts from supply constraints; and the supplier shared capacity risks. Forecast overrides required a written reason and expiry date. The warehouse also established an exception report for unusually high forecasts, negative inventory, zero-demand streaks and large forecast errors.

Outcome

The new process reduced the tendency to interpret stockouts as weak demand, improved visibility of contract peaks and aligned safety stock with both demand and lead-time uncertainty. The case does not imply that one method is universally best. Its central lesson is that data correction, segmentation, collaboration and policy design were as important as the forecasting algorithm.

Learning points

• Forecast constrained sales cautiously.
• Segment ordinary, seasonal, event-driven and critical demand.
• Evaluate models using time-ordered evidence.
• Link uncertainty measures to service and lead-time policy.
• Govern judgmental overrides rather than eliminating them.

11. Implementation Best Practices and Limitations

An effective implementation should begin with a clear decision calendar: when forecasts are produced, frozen, reviewed and converted into orders. Products should be segmented by value, volume, criticality, lifecycle and demand pattern. The method should be proportionate to the decision; a highly complex model is not justified for a stable, low-value item if a transparent baseline performs equally well.

Forecasts should be communicated with a point estimate, horizon, assumptions, confidence or scenario information and known limitations. Forecast accuracy should be measured after the relevant lead time, not only immediately after publication. Managers should distinguish forecast error from execution error: a supplier delay or warehouse picking failure is not necessarily a forecasting failure.

Limitations remain substantial. Structural breaks, new products, promotions, substitution, competitor actions, weather extremes, geopolitical disruption and bullwhip effects can invalidate historical relationships. Aggregated forecasts may conceal local shortages, while highly granular forecasts may be too noisy. Forecasts also influence behaviour: customers may place larger orders when they expect scarcity, and suppliers may build inventory in response to a shared forecast. Consequently, forecasting must be integrated with scenario planning, inventory policy, capacity management and human judgment.

Chapter Summary

Demand analysis distinguishes customer requirements from observed sales and separates independent from dependent demand. Demand may exhibit level, trend, seasonality, cyclicality, intermittency or event-driven variation. Reliable forecasting requires clean, correctly timed and appropriately aggregated data.

Qualitative methods contribute expert and market knowledge when data are limited or structural change is important. Time-series methods include naïve forecasts, moving averages, exponential smoothing, trend-adjusted models and seasonal approaches. Causal models link demand to explanatory variables but require careful attention to assumptions, data availability and the difference between association and causation.

Forecast error should be measured using more than one indicator. ME reveals bias, MAE provides average error in units, RMSE penalizes large errors and percentage measures require caution when actual demand is zero or small. Forecasts become useful inventory inputs only when translated into reorder points, safety stock, service targets and capacity plans. Collaborative planning improves decisions by combining partner information, while technology and AI require validation, transparency, cybersecurity and human accountability. No method removes uncertainty; professional practice manages it explicitly.

Key Terms

Demand: Quantity of a product or service required during a specified period.

Independent demand: Demand originating from external customers or markets.

Dependent demand: Demand derived from the requirement for another item.

Forecast: An estimate of future demand over a defined horizon.

Demand pattern: A recurring or systematic characteristic of demand over time.

Seasonality: A regular demand pattern associated with a calendar or operating cycle.

Intermittent demand: Demand containing many zero observations and occasional non-zero values.

Naïve forecast: A forecast based on a recent or corresponding historical observation.

Moving average: An average of a specified number of recent observations.

Exponential smoothing: A method that updates forecasts using a weighted combination of actual demand and the previous forecast.

Causal forecast: A forecast based on explanatory variables believed to influence demand.

Forecast error: The difference between actual and forecast demand.

Bias: Persistent tendency to overforecast or underforecast.

MAE: Mean absolute error, expressed in demand units.

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

Safety stock: Inventory held to protect against demand or supply uncertainty.

Reorder point: Inventory position at which replenishment is initiated.

CPFR: Collaborative planning, forecasting and replenishment among supply-chain partners.

Forecast reconciliation: Adjustment of forecasts so that related hierarchical forecasts are coherent.

Demand sensing: Use of recent, often near-real-time signals to update a demand estimate.

Review Questions

Short Answer Questions

1. What is the difference between independent and dependent demand?
2. Why can observed sales underestimate true demand?
3. State two assumptions of simple exponential smoothing.
4. What information does MAE provide that ME does not?
5. Why should forecast overrides be documented?

Long Answer Questions

1. Explain the main demand patterns found in inventory systems and discuss the forecasting implications of each.
2. Compare moving averages, exponential smoothing and causal forecasting in terms of data requirements, responsiveness and limitations.
3. Explain how data preparation affects forecast quality and inventory decisions.
4. Discuss how forecast accuracy measures should be selected for a warehouse environment.
5. Explain how collaborative planning can reduce inventory distortion across supply-chain partners.

Analytical / Application Questions

1. A product has actual weekly demand of 200, 240 and 220 units. Forecasts were 210, 230 and 250 units. Calculate ME and MAE, and interpret the bias.
2. A warehouse uses simple exponential smoothing with α = 0.25. The current forecast is 400 units and actual demand is 440 units. Calculate the next forecast and explain the effect of increasing α.
3. A critical item has average demand of 50 units per week, fixed lead time of four weeks, weekly demand standard deviation of 8 units and a service factor z = 2.05. Using ROP = dL + zσ_d√L, calculate the reorder point and state the assumptions behind the calculation.

Multiple Choice Questions with Answers

Question 1

Which statement best describes independent demand?

A. Demand derived from a bill of materials

B. Demand generated by external customers or markets

C. Demand created solely by supplier lead time

D. Demand caused by warehouse picking errors

Correct Answer: B. Demand generated by external customers or markets

Question 2

A moving average with a larger number of periods generally does what?

A. Responds more quickly to recent changes

B. Removes the need for demand data

C. Produces more smoothing but may respond more slowly

D. Guarantees an unbiased forecast

Correct Answer: C. Produces more smoothing but may respond more slowly

Question 3

If actual demand is 120 units, the forecast is 100 units, and error is defined as actual minus forecast, what is the error?

A. −20 units

B. −10 units

C. 10 units

D. 20 units

Correct Answer: D. 20 units

Question 4

Which measure is expressed in the same units as demand and penalizes large errors more heavily?

A. RMSE

B. MAPE

C. ME

D. Tracking signal

Correct Answer: A. RMSE

Question 5

Simple exponential smoothing is most appropriate for which situation?

A. Demand with no observations at all

B. Stable level demand without systematic trend or seasonality

C. Demand determined entirely by a bill of materials

D. Demand with a known changing seasonal cycle only

Correct Answer: B. Stable level demand without systematic trend or seasonality

Question 6

In the equation F_t+1 = αA_t + (1 − α)F_t, what happens when α is increased?

A. The forecast gives more weight to the latest actual demand

B. The forecast ignores the latest actual demand

C. The forecast becomes independent of all previous forecasts

D. The forecast must equal zero

Correct Answer: A. The forecast gives more weight to the latest actual demand

Question 7

Why can MAPE be problematic for intermittent demand?

A. It always has demand units

B. It cannot be calculated when actual demand is zero

C. It penalizes large errors more than RMSE

D. It measures only forecast bias

Correct Answer: B. It cannot be calculated when actual demand is zero

Question 8

For fixed lead time L and independent demand standard deviation σ_d, which expression gives the standard deviation of lead-time demand under the stated approximation?

A. σ_d ÷ L

B. Lσ_d

C. σ_d√L

D. σ_d + L

Correct Answer: C. σ_d√L

Question 9

What is the principal purpose of a statistical baseline in collaborative forecasting?

A. To prevent partners from sharing any information

B. To provide a transparent starting point before documented judgment is added

C. To replace all inventory policies

D. To convert supply constraints into demand

Correct Answer: B. To provide a transparent starting point before documented judgment is added

Question 10

A forecast for three periods has errors 10, −10 and −20 units. What is the MAE?

A. −6.67 units

B. 10 units

C. 13.33 units

D. 40 units

Correct Answer: C. 13.33 units