3. Demand Management and Forecasting

Introduction and Context

Production planning and control begins with a difficult managerial question: what will customers require, when will they require it, and in what quantity? The answer is rarely known with certainty. Demand is influenced by price, income, competition, technology, seasonality, economic conditions, customer preferences, and the availability of substitutes. Consequently, production decisions are made under uncertainty rather than against a perfectly known order book.

Demand management is the broader managerial process through which an organisation understands, influences, prioritises, and communicates demand. Forecasting is one important component of this process. A forecast is not a prediction of the future in a deterministic sense; it is a structured estimate based on information, assumptions, and analytical methods. Its purpose is to support decisions concerning capacity, procurement, inventory, workforce, scheduling, distribution, and customer service.

In production planning and control, the value of a forecast is determined not only by its statistical accuracy but also by how effectively it is converted into action. A highly accurate forecast that is not integrated with planning systems may be less useful than a moderately accurate forecast that is understood, regularly revised, and linked to operating decisions.

Learning Objectives

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

1. Explain the meaning, scope, and role of demand management in production systems.

2. Distinguish between independent and dependent demand and between different forecasting horizons.

3. Select and apply qualitative, time-series, and causal forecasting methods.

4. Calculate and interpret common measures of forecast error.

5. Identify forecast bias and understand its managerial consequences.

6. Translate uncertain demand into capacity, inventory, workforce, and scheduling decisions.

7. Evaluate the use of digital analytics and artificial intelligence in forecasting.

8. Recognise ethical, organisational, and practical limitations of forecasting systems.

Demand Management as a Planning Discipline

Demand management includes the activities used to sense demand, shape it where possible, and reconcile it with organisational resources. Demand sensing involves collecting information from sales orders, market data, customer interactions, point-of-sale systems, and supply-chain partners. Demand shaping involves actions such as pricing, promotion, product substitution, delivery commitments, and allocation rules. Demand communication ensures that sales, marketing, operations, finance, procurement, and senior management use a consistent set of assumptions.

A useful distinction is between independent demand and dependent demand. Independent demand arises directly from customers or the external market, such as demand for finished automobiles or packaged food. Dependent demand is derived from the requirements of another item. For example, the demand for tyres depends partly on the production schedule for vehicles. Independent demand is commonly forecast, whereas dependent demand is often calculated through material requirements planning once the production plan and bill of materials are known.

Demand may also be classified as:

• Known demand, represented by confirmed customer orders or contractual requirements.
• Uncertain demand, estimated from market behaviour and historical patterns.
• Regular demand, displaying relatively stable behaviour.
• Intermittent demand, characterised by many zero-demand periods and occasional transactions, as in spare parts.
• Derived or lumpy demand, in which demand occurs in irregular batches.

These distinctions matter because a method suitable for high-volume retail products may be inappropriate for low-volume maintenance items.

Forecasting Horizons

Forecasts are usually prepared for three overlapping horizons. The short-term horizon supports operational decisions such as weekly schedules, replenishment, and overtime. The medium-term horizon supports aggregate production planning, workforce levels, subcontracting, and inventory targets. The long-term horizon informs strategic decisions concerning facilities, technology, product portfolios, and supply-network design.

The appropriate level of detail changes with the horizon. Short-term forecasts may be prepared for individual stock-keeping units and locations. Long-term forecasts are often more useful when expressed in product families, market segments, or capacity units. Forecast uncertainty generally increases as the horizon extends, although short-term forecasts can also be unstable when demand is intermittent or affected by sudden events.

The Forecasting Process

Forecasting should be treated as a managerial process rather than as a single calculation. A practical process contains the following connected activities:

1. Define the decision to be supported, including the item, location, time bucket, horizon, and required accuracy.

2. Clarify the demand concept: sales, shipments, consumption, orders, or potential market demand.

3. Gather and clean relevant data, identifying missing values, returns, stock-outs, promotions, and exceptional events.

4. Explore the pattern for level, trend, seasonality, cycles, intermittence, and structural changes.

5. Select an appropriate method and establish a transparent baseline.

6. Generate the forecast and document assumptions.

7. Obtain informed managerial input without allowing unsupported optimism or pessimism to dominate.

8. Measure errors, investigate bias, and revise the method when necessary.

9. Translate the forecast into planning inputs and communicate uncertainty to decision-makers.

A crucial issue is the difference between observed sales and true demand. If a product was unavailable for part of a month, recorded sales may understate what customers wanted. Similarly, a promotion may temporarily inflate sales without changing the underlying baseline. Data must therefore be interpreted in its operational context.

Qualitative Forecasting Methods

Qualitative methods are especially valuable when historical data are limited, the product is new, or the environment has undergone a structural change. They depend on informed judgement rather than solely on numerical history.

The executive opinion method obtains estimates from senior managers. It is fast and useful for strategic decisions, but power relationships and dominant personalities may suppress alternative views. The sales-force composite collects estimates from sales personnel who are close to customers. It can incorporate local knowledge, although sales staff may overestimate demand to secure inventory or underestimate it to reduce performance pressure.

The jury of expert opinion uses a panel of specialists. A more structured form is the Delphi method, in which experts provide opinions independently and receive controlled, anonymised feedback over several rounds. Delphi reduces direct group pressure, but it remains dependent on the quality and diversity of the panel.

Market research, including customer surveys, interviews, concept tests, and pilot selling, is useful for new products and changes in customer preferences. Stated intentions, however, do not always become actual purchases. Analogy forecasting estimates demand for a new product by comparing it with the adoption pattern of a similar existing product.

Qualitative judgement should not be dismissed as unscientific. It becomes more reliable when assumptions are explicit, evidence is documented, estimates are separated from targets, and forecasts are compared with actual outcomes. Judgement is most effective when used to explain or adjust a statistical baseline rather than to replace analysis without justification.

Time-Series Forecasting

Time-series methods assume that useful information about future demand is contained in the historical sequence of observations. A time series may contain four components:

• Level, the underlying average quantity.
• Trend, a persistent upward or downward movement.
• Seasonality, a systematic pattern repeating within a year, week, or other fixed period.
• Cycle or irregular movement, variation associated with broader economic or exceptional events.

Naïve and Moving-Average Methods

The naïve forecast sets the next period’s forecast equal to the latest actual observation:

\[ F_{t+1}=A_t \]

It is simple, transparent, and a useful benchmark. If a sophisticated method cannot outperform it consistently, its additional complexity should be questioned.

A simple moving average uses the mean of the most recent observations:

\[ F_{t+1}=\frac{A_t+A_{t-1}+\cdots+A_{t-n+1}}{n} \]

A larger value of n smooths random variation but responds slowly to changes. A smaller value responds more quickly but may retain noise. A weighted moving average assigns greater weights to observations considered more relevant, with the weights normally summing to one.

Exponential Smoothing

Simple exponential smoothing updates the previous forecast using the latest error:

\[ F_{t+1}=F_t+\alpha(A_t-F_t) \]

where α lies between zero and one. A high α gives substantial weight to recent observations and makes the forecast responsive. A low α produces greater smoothing. Simple exponential smoothing is appropriate where demand has a relatively stable level but no strong trend or seasonality.

For data with trend, Holt’s method estimates both level and trend. For data with trend and seasonality, Holt–Winters methods estimate level, trend, and seasonal indices. Additive seasonality is appropriate when seasonal fluctuations remain approximately constant in absolute size; multiplicative seasonality is more suitable when seasonal variation changes in proportion to the level of demand.

Trend and Seasonal Analysis

Trend can be estimated through methods such as regression on time or by smoothing. Seasonal indices compare demand in a season with the underlying average. For example, an index of 1.20 indicates demand approximately 20 per cent above the average level, subject to the assumptions of the model. Seasonality should be interpreted carefully when product availability, promotions, or calendar changes distort historical patterns.

Intermittent Demand

Traditional averages often perform poorly for spare parts and other intermittent items. Croston-type approaches separately estimate the size of a demand occurrence and the interval between occurrences. More recent methods, including probabilistic models, seek to estimate both the probability and magnitude of future demand. The planning objective is often not a single precise forecast but an appropriate service level and replenishment policy.

Causal and Advanced Forecasting Approaches

Time-series methods extrapolate historical patterns. Causal forecasting explains demand through variables believed to influence it, such as price, advertising expenditure, income, weather, competitor activity, or the number of active customers. A basic regression model may be expressed as:

\[ D_t=\beta_0+\beta_1X_{1t}+\beta_2X_{2t}+\varepsilon_t \]

where D is demand, X variables are explanatory factors, β coefficients measure estimated relationships, and ε represents unexplained variation.

Causal models are useful when management expects an important driver to change. For instance, a price increase may reduce volume, while a promotional campaign may increase it. However, correlation does not establish causation. A model may fit historical data yet fail when customer behaviour, competitors, or regulations change. Analysts should assess the logic of the relationship, data quality, coefficient stability, multicollinearity, residual behaviour, and out-of-sample performance.

Econometric models, market-response models, simulation, and system-dynamics models can represent more complex relationships. Machine-learning models such as random forests, gradient boosting, and neural networks may improve performance where large, granular datasets exist. Their advantages can be offset by limited interpretability, data leakage, unstable relationships, and excessive dependence on historical patterns.

A sound forecasting function commonly combines methods. Statistical forecasts provide consistency, while sales, operations, and market specialists contribute information about promotions, tenders, disruptions, and product launches. The combination must be governed carefully: every override should have an identifiable reason, owner, and subsequent review.

Measuring Forecast Error and Bias

A forecast should be evaluated against actual demand using a defined error convention. Let e_t = A_t − F_t. Positive error means actual demand exceeded the forecast; negative error means the forecast was higher than actual demand.

The mean error is:

\[ ME=\frac{\sum e_t}{n} \]

It indicates bias but may conceal large offsetting errors. The mean absolute deviation (MAD) is:

\[ MAD=\frac{\sum |e_t|}{n} \]

It expresses average error in the original units. The mean squared error (MSE) gives greater weight to large errors:

\[ MSE=\frac{\sum e_t^2}{n} \]

The root mean squared error (RMSE) is the square root of MSE and is also expressed in demand units. The mean absolute percentage error (MAPE) is:

\[ MAPE=\frac{100}{n}\sum\left|\frac{e_t}{A_t}\right| \]

MAPE is intuitive but becomes unstable or undefined when actual demand is zero or very small. The weighted absolute percentage error and scale-independent measures such as MASE can be more appropriate across products with different volumes.

No single measure is universally best. A production manager may prefer MAD for inventory planning, RMSE where large misses are particularly costly, and a percentage measure for comparing product families. The cost of error should influence the evaluation: under-forecasting may cause lost sales and overtime, whereas over-forecasting may create obsolescence and excess inventory.

Detecting Bias

Bias is a systematic tendency to over-forecast or under-forecast. Persistent positive values of e_t indicate repeated under-forecasting; persistent negative values indicate over-forecasting. A tracking signal can be constructed as:

\[ TS=\frac{RSFE}{MAD} \]

where RSFE is the running sum of forecast errors. A sustained extreme tracking signal suggests that the forecasting process should be investigated. Exact control limits vary by organisation and application; the purpose is to prompt diagnosis, not to produce an automatic managerial decision.

Bias may arise from optimistic sales targets, deliberate safety margins, obsolete parameters, product cannibalisation, stock-outs in the historical data, or failure to remove promotional effects. Correcting bias requires addressing its cause rather than simply adding a permanent adjustment to every forecast.

From Forecast to Production Plan

A forecast is a planning input, not a production order. The organisation must decide how to respond to uncertainty. It may build to stock, build to order, assemble to order, postpone final configuration, maintain safety stock, reserve flexible capacity, or use a combination of these approaches.

The aggregate plan converts forecasts for many products into decisions about production rate, workforce, inventory, overtime, subcontracting, and backlog. The master production schedule then expresses the plan in more specific end-item quantities and time periods. Material requirements planning translates the schedule into component and purchasing requirements. Forecasts therefore connect market information with the broader hierarchy of production planning and control.

Forecast uncertainty should be represented explicitly. A point forecast can be accompanied by a prediction interval, demand scenarios, or a range such as most likely, optimistic, and pessimistic. Capacity planning may use a central forecast for normal resources and a high-demand scenario for testing bottlenecks. Inventory policy can incorporate forecast error through safety stock, although safety stock should reflect both variability and the required service level.

Managers should distinguish forecast accuracy from decision quality. An unexpected demand shock may produce a large error even when the method was reasonable. Conversely, an accurate forecast may lead to poor results if procurement is slow, capacity is unavailable, or the organisation cannot execute the plan.

Digital Forecasting, Analytics, and Artificial Intelligence

Enterprise resource planning systems, advanced planning and scheduling systems, customer relationship platforms, point-of-sale feeds, and internet-connected equipment provide increasingly detailed demand data. Automated pipelines can clean observations, detect anomalies, update forecasts, and distribute revised planning signals. Dashboards make forecast performance visible across products, customers, and locations.

Artificial intelligence can identify nonlinear relationships and interactions that are difficult to specify manually. It is particularly useful when demand depends on many variables, such as weather, digital traffic, promotions, and location-specific behaviour. Nevertheless, AI does not eliminate the need for domain knowledge. Models can reproduce historical bias, fail during structural breaks, and generate apparently precise outputs that users cannot explain.

Good governance requires version control, data lineage, access controls, documented assumptions, monitoring for drift, and human review of material overrides. Forecasts used to allocate scarce products or prioritise customers should be examined for unfair or commercially inappropriate outcomes. Privacy obligations also apply when customer-level data are used.

Illustrative Case: Forecasting at Meridian Home Appliances

Background

Meridian Home Appliances, a fictional mid-sized manufacturer, produces three families of countertop appliances for national retailers. Production capacity is shared across the families, and component suppliers require six to eight weeks’ notice. Historically, the planning team used the sales department’s monthly estimates with limited statistical review.

Situation and Problem

During one year, Meridian experienced frequent stock-outs of a popular blender, excess inventory of a premium mixer, and repeated schedule changes. The blender had been promoted in several months, while the mixer faced a competitor’s product launch. Recorded sales also understated blender demand during stock-out periods. The sales estimates were consistently higher than actual mixer sales but lower than the eventual market requirement for the blender.

Analysis

The team first separated confirmed orders, ordinary sales, promotional sales, and lost-sales estimates. It plotted monthly demand and identified seasonality before comparing a naïve forecast, a three-month moving average, exponential smoothing, and a regression model including price and promotion indicators. Forecast error was evaluated using MAD and bias measures rather than MAPE alone, because some regional observations were very small.

The analysis showed that the moving average reacted too slowly after promotions, while simple exponential smoothing performed reasonably for the stable mixer. A causal model improved the blender forecast when promotion intensity was known in advance. However, the model was not used mechanically: marketing provided the promotion calendar, and operations identified months in which supply constraints had censored observed sales.

Approach and Solution

Meridian introduced a statistical baseline for each product family, followed by a monthly cross-functional demand review. Overrides required a documented reason, such as a confirmed promotion, new distribution agreement, or competitor action. The firm adopted different methods by demand pattern: exponential smoothing for the stable mixer, a seasonal causal model for the blender, and a judgement-supported scenario approach for a newly launched appliance.

The resulting forecasts were converted into a central production plan and a high-demand capacity scenario. Critical components were purchased against the central plan with selected flexibility clauses, while final packaging was postponed where feasible. A dashboard tracked MAD, bias, stock-outs, forecast overrides, and forecast value added by method.

Outcome

The principal improvement was not a claim of perfect accuracy. Rather, sales and operations developed a shared explanation of demand, promotions were communicated earlier, and capacity decisions were based on visible scenarios. The organisation reduced avoidable schedule changes and discovered that some apparent forecasting errors were actually data-quality and availability problems.

Learning Points

The case demonstrates that forecasting quality depends on demand definition, data interpretation, method selection, cross-functional governance, and operational flexibility. A forecast should be evaluated in relation to the decision it supports and the cost of being wrong.

Challenges, Risks, and Common Mistakes

Common mistakes include forecasting sales when the decision requires customer consumption, ignoring stock-outs, using a single method for all products, selecting a method because it is sophisticated, and judging performance on one unusually favourable period. Another error is confusing a target with a forecast. A target expresses what management wants to achieve; a forecast estimates what is likely under stated assumptions. Combining them without disclosure can create unrealistic production plans.

Forecasts can also be distorted by anchoring, groupthink, optimism bias, recency bias, and incentives tied to sales commitments. Excessive manual adjustment reduces accountability and makes it impossible to learn from errors. At the opposite extreme, blind reliance on an algorithm ignores information about product launches, strikes, regulation, or competitor action.

Intermittent demand, short product life cycles, promotions, cannibalisation, changing channels, and structural breaks require special care. Forecasts should be reviewed at an appropriate frequency, but constant revision without a stable process can create nervousness and amplify the bullwhip effect throughout the supply chain.

Best Practices for Managers

Effective demand management generally follows these principles:

• Begin with the decision and planning horizon, not with a favourite technique.
• Establish a transparent baseline, often using a naïve or simple statistical method.
• Segment products by demand pattern, value, volatility, and lifecycle stage.
• Correct data for stock-outs, returns, promotions, and one-time events.
• Separate statistical forecasts, managerial overrides, and performance targets.
• Measure both accuracy and bias using metrics suited to the decision.
• Evaluate forecasts out of sample and compare them with simple benchmarks.
• Use scenarios and ranges when uncertainty is material.
• Assign ownership for assumptions, overrides, and forecast review.
• Link forecast performance to inventory, service, capacity, and financial outcomes.

Chapter Summary

Demand management coordinates the sensing, interpretation, shaping, and communication of market requirements. Forecasting supports production planning but cannot remove uncertainty. Independent demand is generally forecast, while dependent demand is derived from planned production and product structure. Qualitative methods are valuable for new products and structural change; time-series methods exploit level, trend, and seasonality; causal models relate demand to explanatory variables. Forecasts must be assessed through measures such as ME, MAD, MSE, RMSE, MAPE, and bias indicators, with attention to the cost of error and the problem of zero demand. Forecasts become useful only when translated into aggregate plans, master schedules, inventory policies, capacity decisions, and procurement actions. Digital tools and AI can improve scale and responsiveness, but they require sound data, human judgement, transparency, and governance.

Key Terms

Demand management: The process of sensing, influencing, prioritising, and communicating customer demand.

Forecast: A structured estimate of future demand based on information and assumptions.

Independent demand: Demand arising directly from external customers or markets.

Dependent demand: Demand derived from the requirements of another item or production schedule.

Forecast horizon: The future period covered by a forecast.

Demand sensing: The collection and interpretation of current signals about market requirements.

Demand shaping: Actions intended to influence the timing, quantity, or form of demand.

Time series: A sequence of observations recorded over successive time periods.

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

Naïve forecast: A forecast that uses the latest actual observation as the next forecast.

Exponential smoothing: A method that updates a forecast by weighting the latest forecast error.

Causal forecasting: Forecasting based on explanatory variables believed to influence demand.

Forecast error: The difference between actual demand and forecast demand.

MAD: The mean of absolute forecast errors, expressed in demand units.

RMSE: The square root of mean squared error, giving extra influence to large errors.

Forecast bias: A systematic tendency to over-forecast or under-forecast.

Intermittent demand: Demand containing numerous zero periods and occasional occurrences.

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

Prediction interval: A range indicating the uncertainty surrounding a forecast.

Forecast override: A documented managerial adjustment to a statistical forecast.

Review Questions

Short Answer Questions

1. Distinguish between independent and dependent demand.

2. What is the purpose of a naïve forecast?

3. Explain the effect of increasing alpha in simple exponential smoothing.

4. Why can MAPE be problematic when actual demand is zero?

5. What does persistent forecast bias indicate?

Long Answer Questions

1. Explain the forecasting process from demand definition through integration with production planning.

2. Compare qualitative, time-series, and causal forecasting methods, including their appropriate applications.

3. Discuss the role of level, trend, seasonality, and intermittence in time-series forecasting.

4. Explain how forecast accuracy and forecast bias should be measured and interpreted by production managers.

5. Discuss how digital analytics and artificial intelligence can improve forecasting while creating new governance risks.

Analytical / Application Questions

1. A product recorded demands of 90, 100, and 110 units in the last three months. Calculate the three-month moving-average forecast for the next month and explain one advantage and one limitation of the method.

2. A forecast for four periods is 120, 150, 130, and 160 units, while actual demand is 130, 140, 150, and 170 units. Calculate the mean error and MAD using e = actual minus forecast. What does the result suggest about bias?

3. A spare-parts item has demand in only four of the last twelve months, with highly variable quantities. Recommend a forecasting and inventory-planning approach, explaining why a standard moving average may be unsuitable.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes demand management from forecasting?

A. Demand management is limited to calculating statistical averages.

B. Forecasting eliminates uncertainty, whereas demand management creates it.

C. Demand management includes sensing, influencing, prioritising, and communicating demand; forecasting is one of its components.

D. Forecasting applies only to dependent demand.

Correct Answer: C. Demand management includes sensing, influencing, prioritising, and communicating demand; forecasting is one of its components.

Question 2

A three-month moving average is changed to a one-month moving average. What is the most likely effect?

A. Greater smoothing and slower response to changes.

B. Faster response but greater sensitivity to random variation.

C. Automatic removal of seasonality.

D. Elimination of forecast bias.

Correct Answer: B. Faster response but greater sensitivity to random variation.

Question 3

In simple exponential smoothing, a high value of alpha generally makes the forecast:

A. More responsive to the latest observation.

B. Independent of recent demand.

C. Less responsive to changes in demand.

D. Suitable only for intermittent demand.

Correct Answer: A. More responsive to the latest observation.

Question 4

If actual demand is 200 and the forecast is 240, using error = actual minus forecast, the error is:

A. 40

B. −40

C. 440

D. 0.83

Correct Answer: B. −40

Question 5

Which measure gives relatively greater weight to large forecast errors?

A. Mean error

B. MAD

C. MSE

D. Median demand

Correct Answer: C. MSE

Question 6

A persistent positive mean error, where error equals actual minus forecast, most likely indicates:

A. Systematic under-forecasting.

B. Systematic over-forecasting.

C. Perfect forecast accuracy.

D. Absence of seasonality.

Correct Answer: A. Systematic under-forecasting.

Question 7

Which situation most strongly supports the use of a causal forecasting model?

A. Demand has a stable level and no known influencing variables.

B. Management expects a major price and promotion change that historically affects demand.

C. All historical observations are missing.

D. The item has no relationship with market conditions.

Correct Answer: B. Management expects a major price and promotion change that historically affects demand.

Question 8

Why should recorded sales during a stock-out be interpreted cautiously?

A. They may overstate potential customer demand.

B. They may understate demand because some customers could not purchase.

C. They always represent dependent demand.

D. They automatically remove forecast bias.

Correct Answer: B. They may understate demand because some customers could not purchase.

Question 9

Which is the most appropriate managerial response to a statistical forecast override?

A. Allow unlimited undocumented adjustments.

B. Reject all judgement regardless of market information.

C. Document the reason, owner, and assumptions and review the outcome later.

D. Convert the override directly into a production order.

Correct Answer: C. Document the reason, owner, and assumptions and review the outcome later.

Question 10

For highly intermittent spare-parts demand, which approach is generally most appropriate?

A. Use a method that considers occurrence intervals and demand size, together with service-level policy.

B. Assume demand is constant every period.

C. Use only annual average demand and ignore zero periods.

D. Replace the forecast with the sales target.

Correct Answer: A. Use a method that considers occurrence intervals and demand size, together with service-level policy.