Demand Forecasting and Planning

Introduction

Demand forecasting is the systematic estimation of future customer requirements for a product, service, resource, or capacity. In operations management, a forecast is not an assertion that the future is known; it is a structured estimate based on information, assumptions, and judgement. Forecasts support decisions about procurement, workforce, capacity, inventory, budgets, distribution, and customer service.

The central managerial problem is that demand is uncertain, while many operating decisions must be made before actual demand is observed. A manufacturer may need to purchase materials several months before production. An airline must allocate seats and crews before passengers book. A hospital must schedule staff before patient arrivals are known. Forecasting therefore connects market intelligence with operational planning.

A useful forecast is not necessarily the most statistically sophisticated forecast. It is one that is sufficiently accurate, timely, explainable, economical, and aligned with the decision it supports. Forecasting should consequently be treated as a process of learning and coordination rather than merely a calculation.

Learning Objectives

After studying this chapter, you should be able to:

1. Explain the role, scope, and assumptions of demand forecasting.
2. Distinguish qualitative, time-series, causal, and judgemental methods.
3. Develop and interpret moving-average, exponential-smoothing, trend, seasonal, and regression forecasts.
4. Measure forecast accuracy, bias, and uncertainty.
5. Select methods according to data, decision horizon, and business context.
6. Translate demand forecasts into aggregate plans, inventory decisions, capacity plans, and collaborative actions.
7. Critically evaluate forecasting limitations, ethical risks, and the use of analytics and artificial intelligence.

1. The Nature of Demand and Forecasting

Demand is the quantity customers are willing and able to purchase under specified conditions. It differs from sales: observed sales may be constrained by stockouts, capacity limits, distribution problems, or pricing decisions. If a retailer sells 1,000 units because only 1,000 were available, sales are 1,000 but unconstrained demand may have been higher.

Forecasting decisions should begin by defining:

• the item or demand family;
• the customer, market, or channel;
• the time horizon and forecast frequency;
• the unit of measurement;
• whether demand is independent or dependent;
• whether the forecast is for unconstrained demand or expected sales.

Independent demand arises primarily from external customers, such as demand for finished products. Dependent demand is derived from another item or activity, such as component requirements calculated from a production schedule. Dependent demand is usually planned through material requirements planning rather than forecast independently, although the finished product still requires a forecast.

Forecast horizons are commonly classified as short term, medium term, and long term. Short-term forecasts support scheduling and replenishment; medium-term forecasts support workforce, inventory, and aggregate capacity decisions; long-term forecasts support facilities, technology, product portfolios, and strategic investments. Accuracy generally declines as the horizon increases because more uncertainty accumulates.

A demand series may contain several components:

A general additive representation is:

where Y_t is observed demand at time t, L_t is level, T_t is trend, S_t is seasonal effect, C_t is cyclical effect, and I_t is irregular variation. This representation is appropriate when seasonal fluctuations have approximately constant absolute size. A multiplicative representation, Y_t = L_t × T_t × S_t × C_t × I_t, is more suitable when seasonal variation grows in proportion to the level of demand. In practice, the components may interact and need not be directly observable.

2. Forecasting as a Management Process

A robust forecasting process has seven linked stages:

1. Define the decision and demand measure. A forecast for monthly production capacity differs from one for daily dispatching.
2. Collect and prepare data. Check units, dates, missing values, stockouts, returns, promotions, and changes in product definitions.
3. Explore demand patterns. Examine plots, summary statistics, seasonality, outliers, structural breaks, and relationships with explanatory variables.
4. Select candidate methods. Use more than one plausible method when uncertainty is material.
5. Estimate and validate. Test forecasts using historical holdout periods or rolling-origin evaluation.
6. Incorporate informed judgement. Document assumptions and adjustments rather than hiding them in an unexplained number.
7. Communicate, monitor, and revise. Compare forecasts with actuals and establish ownership for corrective action.

Forecasting should be integrated with sales and operations planning (S&OP), in which sales, marketing, finance, operations, procurement, and logistics agree on a feasible plan. The statistical forecast is often a starting point, not the final organisational forecast.

3. Qualitative and Judgemental Forecasting

Qualitative methods are particularly useful when historical data are limited, the product is new, markets are changing, or experts possess information not yet visible in numerical records. Common methods include:

• Executive opinion: senior managers form a forecast, often through discussion. It is fast but vulnerable to hierarchy, optimism, and anchoring.
• Sales-force composite: sales personnel estimate customer or territory demand. This uses local knowledge but may reflect sales incentives or deliberate sandbagging.
• Market research: surveys, interviews, focus groups, test markets, and customer experiments investigate preferences and purchase intentions. Stated intention does not always become actual purchase.
• Analogy: demand for a new product is inferred from a similar product, market, or technology. The quality of the analogy is decisive.
• Delphi method: experts provide anonymous forecasts in successive rounds, receiving an aggregate summary between rounds. Anonymity reduces conformity and dominance effects, although consensus does not guarantee accuracy.
• Scenario analysis: alternative coherent futures—such as high growth, constrained supply, and recession—are developed. Scenarios are not probabilities unless probabilities are explicitly assigned.

Judgement should be structured. Experts should state the evidence used, assumptions made, confidence level, and conditions that would invalidate the forecast. A forecast committee should distinguish new information from pressure to meet a target. Combining independent expert estimates can reduce individual error, but correlated opinions and common biases limit the benefit.

Important judgemental biases include anchoring on last year's budget, overconfidence, confirmation bias, availability bias, optimism about new products, and political manipulation. A disciplined process uses pre-defined adjustment rules, anonymous inputs where suitable, forecast logs, and post-period review.

4. Time-Series Forecasting

Time-series methods infer future demand primarily from the historical behaviour of the same series. They are most useful when patterns are reasonably stable and causal data are unavailable or unnecessary.

4.1 Naïve methods

The simplest forecast for the next period equals the latest observation:

where F_t+1 is the forecast for period t+1 and A_t is actual demand in period t. A seasonal naïve forecast uses demand from the corresponding previous season, such as the same month last year. Naïve methods are important benchmarks. A complex model that cannot outperform a naïve benchmark out of sample may not be useful.

4.2 Moving average

A simple moving average of length n is:

Here, n is the number of equally weighted recent observations. The method smooths random variation but lags when demand has a trend. A larger n produces more smoothing and slower response; a smaller n responds more quickly but retains more noise. The method assumes that recent observations are representative and that no strong trend or changing seasonal pattern is being ignored.

Worked example. Weekly demand for the last three weeks was 120, 135, and 150 units. A three-week moving-average forecast for week 4 is:

The forecast is in units per week. It does not imply that 135 is a confidence limit or a guaranteed requirement.

4.3 Weighted moving average

A weighted moving average gives greater importance to selected observations:

The weights must satisfy w_i ≥ 0 and ∑w_i = 1. If demand for the last three weeks is 150, 135, and 120 units and weights are 0.5, 0.3, and 0.2 respectively, the forecast is (0.5 × 150) + (0.3 × 135) + (0.2 × 120) = 139.5 units. Weights may be selected through judgement or estimated by minimising historical forecast error.

4.4 Exponential smoothing

Single exponential smoothing updates the forecast as follows:

where α is the smoothing constant, 0 ≤ α ≤ 1, A_t is actual demand, and F_t is the forecast for period t. Both demand and forecast must use the same units and time interval. A high α responds strongly to recent changes; a low value produces greater smoothing. The method assumes no systematic trend or seasonality. An initial forecast F₁ must be specified, commonly using the first observation or an average of early observations.

Worked example. Let α = 0.30, actual demand in period 1 be 100 units, and the forecast for period 2 be 110 units. If actual demand in period 2 is 120 units, the forecast for period 3 is:

4.5 Trend-adjusted smoothing

When demand has a trend, Holt's method estimates level and trend. One common formulation is:

Here, L_t is the smoothed level, T_t is the estimated trend, α and β are smoothing constants between 0 and 1, and m is the number of periods ahead. Initial level and trend values are required. Holt's method is inappropriate if an apparent trend is merely a temporary shock or if seasonality is material but unmodelled.

4.6 Seasonal forecasting

For seasonal demand, the analyst may use seasonal indices. In a multiplicative model:

A seasonal factor above 1 indicates demand above the average level; a factor below 1 indicates demand below it. Seasonal factors should normally average 1 across a complete seasonal cycle. For additive seasonality, the seasonal effect is added rather than multiplied. Seasonal indices must be recalculated when customer behaviour, product mix, or calendar effects change.

5. Causal and Regression Forecasting

Causal methods relate demand to explanatory variables such as price, income, advertising, weather, competitor activity, web traffic, or economic indicators. A simple linear regression model is:

where Y_t is demand, X_t is an explanatory variable, a is the intercept, b is the expected change in demand for a one-unit change in X_t, and ε_t is the unexplained error. The units of b are units of demand per unit of X.

For several explanatory variables:

Regression estimates are useful only when the predictors are available, or can themselves be forecast, at the time the demand forecast is produced. Association is not proof of causation. Important assumptions for ordinary least squares inference include an appropriate functional form, independent or suitably modelled errors, stable relationships, and no severe multicollinearity. Forecasting may still be useful when some classical inference assumptions are imperfect, but validation is essential.

A coefficient should be interpreted conditionally. If b = −2.5 for price measured in currency units and demand measured in units, a one-currency increase in price is associated with an estimated decrease of 2.5 units, holding other included variables constant. This does not establish that changing price alone will produce exactly that result, particularly when price and promotion are jointly determined.

Regression can incorporate interventions through indicator variables, for example a promotion dummy equal to 1 during promotion periods and 0 otherwise. Analysts should avoid leakage: a predictor that becomes known only after the forecast date must not be used to evaluate a real-time forecast.

6. Measuring Forecast Error, Bias, and Uncertainty

Let actual demand in period t be A_t and the forecast be F_t. Define forecast error as:

Under this convention, positive error means demand exceeded the forecast. Some organisations use the opposite sign, so the convention must be stated.

Common accuracy measures

These measures assume comparable periods and a clear treatment of zero or near-zero actuals. MAPE is undefined when A_t = 0 and unstable when actual demand is very small. WAPE can be dominated by high-volume items and may hide poor performance on low-volume critical items. RMSE is useful when large errors create disproportionate cost. No single measure is universally best; the measure should reflect the decision's economic consequences.

Worked error example. Suppose actual demand for three periods is 100, 120, and 80 units, while forecasts are 90, 130, and 100 units. Errors are 10, −10, and −20 units. Thus:

The negative ME indicates underforecasting? No. With e_t = A_t − F_t, negative ME means forecasts were, on average, higher than actual demand: overforecasting. Sign interpretation must follow the stated convention.

Tracking signal and bias monitoring

A tracking signal compares cumulative error with MAD:

It is a diagnostic, not a universal statistical test. Persistent positive or negative values may indicate bias, a changed demand process, or a data problem. Organisations should establish limits appropriate to the forecast process rather than applying a threshold mechanically.

Accuracy is not the same as uncertainty. A point forecast gives one central estimate, whereas a prediction interval expresses a range of plausible future demand. Intervals widen with horizon and should account for residual variability, parameter uncertainty, and potential structural change. For inventory decisions, the relevant question is often the cost-weighted distribution of demand during replenishment lead time, not merely the average forecast.

7. Selecting and Combining Forecasts

Method selection should consider:

• forecast horizon and update frequency;
• amount and quality of historical data;
• trend, seasonality, intermittence, and structural breaks;
• availability of causal predictors;
• cost of underforecasting versus overforecasting;
• explainability and governance requirements;
• computational and organisational capability.

Forecasts should be evaluated out of sample. A time-ordered split preserves the forecasting problem: earlier observations train the model and later observations test it. Rolling-origin evaluation repeatedly trains on an expanding or moving historical window and forecasts one or more future periods. Randomly shuffling time-series observations can create unrealistic leakage.

Forecast combinations often improve robustness. A combined forecast can be written as:

where F_jt is forecast j, w_j ≥ 0, and ∑w_j = 1. Equal weights are transparent; performance-based weights may be estimated using historical validation. Combining statistical and judgemental forecasts can be valuable, but the judgemental adjustment should be recorded separately so that its value can be audited.

8. Translating Forecasts into Plans

A forecast becomes operationally useful only when translated into decisions. The main links are:

• Aggregate planning: determine production, workforce, overtime, subcontracting, inventory, and backlogs over a medium-term horizon.
• Master production scheduling: disaggregate the aggregate plan into end-item quantities and timing.
• Material requirements planning: convert end-item schedules into component and raw-material requirements using bills of material and lead times.
• Capacity planning: compare forecast workload with available machine, labour, warehouse, transport, and supplier capacity.
• Inventory planning: set order quantities, reorder points, safety stock, and service targets.
• Revenue and financial planning: align volume assumptions with prices, margins, working capital, and cash flow.

Forecasts should be expressed at the level that supports the decision. A high-level product-family forecast may be more accurate than a highly disaggregated SKU forecast, but purchasing needs SKU-level information. Hierarchical forecasting reconciles forecasts across levels so that category totals and item-level forecasts are coherent. Forecast aggregation often reduces random variation, while disaggregation may expose intermittent demand and greater error.

A useful governance principle is to separate demand ownership from supply feasibility. Sales and marketing may provide market intelligence, while operations tests whether the resulting demand plan can be supplied. When supply constraints exist, the organisation should distinguish unconstrained demand, feasible supply, and expected shipments rather than silently replacing one with another.

9. Illustrative Case: Forecasting at Meridian Appliances

Background. Meridian Appliances, a fictional mid-sized manufacturer, sells countertop air purifiers through retailers and its online channel. Demand has grown, but monthly forecasts are frequently revised. The factory experiences overtime in high-demand months and excess finished-goods inventory after promotions.

Problem. The firm used the previous month's sales as the forecast. A stockout in March reduced recorded sales, so April's forecast was too low. Marketing promotions were also communicated after production decisions had been made. The operations director needs a forecast for the next six months and a process that reduces both bias and avoidable cost.

Analysis. The team first reconciles shipments, cancelled orders, stockouts, returns, and promotional periods. It finds an upward trend, strong seasonality before summer, and a temporary demand spike during a competitor's product recall. A simple moving average reacts too slowly to the trend. A seasonal naïve benchmark, Holt-Winters-type seasonal smoothing, and a regression model using price, promotion, and web searches are compared through rolling-origin validation. The regression model performs well in months where promotion plans are known, but poorly when marketing changes are made late. The seasonal model is more stable operationally.

Approach and solution. Meridian adopts a three-layer process. First, a statistical baseline is generated using cleaned unconstrained demand and a seasonal trend model. Second, marketing submits a calendar of planned promotions and their expected uplift, with assumptions and confidence ratings. Third, an S&OP meeting reviews capacity, supplier lead times, and alternative scenarios. The company reports forecast error using WAPE for portfolio monitoring and MAD for high-volume production items; MAPE is not used for low-volume SKUs with frequent zero demand. Forecast adjustments are logged and reviewed after each cycle.

The production plan uses the central forecast, while selected components with long lead times are planned against an upper-demand scenario. Procurement is given a probability-informed range rather than a single number. A stockout flag prevents constrained sales from being treated as unconstrained demand.

Outcome. In the illustrative case, the process reduces systematic overreaction to single-month changes, makes promotion assumptions visible, and improves cross-functional accountability. It does not eliminate forecast error. Instead, it enables Meridian to make explicit trade-offs between inventory, overtime, lost sales, and supplier flexibility.

Learning points. Data preparation can matter more than model complexity. Forecasts must reflect the information available at decision time. Scenario ranges are valuable when capacity and lead times make errors costly. Finally, a forecast should be evaluated by the decisions it improves, not only by a statistical score.

10. Technology, AI, and Contemporary Developments

Modern forecasting systems may use cloud data platforms, automated model selection, gradient-boosting algorithms, neural networks, probabilistic forecasts, and demand-sensing signals such as point-of-sale data, search activity, weather, and mobility. These tools can detect nonlinear relationships and process many series, but they do not remove the need for causal reasoning, data governance, or managerial accountability.

Artificial intelligence is most useful when it improves scale, speed, anomaly detection, scenario generation, or probabilistic estimation. Risks include data leakage, unstable relationships, opaque model behaviour, automation bias, discrimination in allocation, and excessive dependence on high-frequency proxies. A model trained during an unusual period may fail when conditions normalise. Human review should focus on exceptions, structural changes, and decisions with material consequences—not on routinely overriding models without evidence.

Best practice includes version-controlled data, reproducible pipelines, documented features, back-testing, monitoring for drift, access controls, model approval, and clear ownership. Forecast systems should preserve an audit trail showing the baseline forecast, overrides, rationale, and subsequent outcome. Privacy rules apply when customer-level data are used; data minimisation and appropriate aggregation are important safeguards.

11. Challenges and Limitations

Forecasting is limited by:

• structural breaks caused by technology, regulation, crises, or competitors;
• incomplete or biased historical data;
• intermittent and lumpy demand;
• new products with no demand history;
• endogeneity, where price and promotion respond to expected demand;
• supply constraints that obscure true customer demand;
• the bullwhip effect, in which small demand changes become amplified upstream;
• organisational incentives that encourage optimistic or conservative forecasts;
• model overfitting and false precision.

Forecast uncertainty should be communicated honestly. A narrow interval can create false confidence, while an overly broad range may be ignored. Managers should connect forecast ranges to actions: for example, base capacity for expected demand, flexible labour for an upper scenario, and contingency suppliers for a severe but plausible outcome.

Best-Practice Checklist

1. Define demand, horizon, unit, and decision owner.
2. Separate demand from constrained sales wherever possible.
3. Establish naïve benchmarks before adopting complex models.
4. Inspect trend, seasonality, outliers, missing values, and structural breaks.
5. Validate with time-ordered, out-of-sample tests.
6. Measure both magnitude of error and directional bias.
7. Match metrics to economic consequences and volume characteristics.
8. Record judgemental adjustments and their assumptions.
9. Reconcile forecasts across product, geographic, and organisational levels.
10. Convert point forecasts into scenarios, capacity actions, and inventory policies.
11. Monitor drift, data quality, and override behaviour.
12. Review the value created by the forecast, not merely its statistical accuracy.

Chapter Summary

Demand forecasting estimates future customer requirements so that organisations can make timely decisions under uncertainty. Demand series may contain level, trend, seasonality, cycles, random variation, and interventions. Qualitative methods are valuable when data are scarce or structural change is significant; time-series methods exploit historical patterns; causal methods use explanatory variables; and combinations can improve robustness.

Moving averages smooth demand, exponential smoothing updates forecasts recursively, trend-adjusted methods represent changing levels, seasonal models represent recurring patterns, and regression relates demand to predictors. Every method depends on assumptions and should be evaluated with information that would have been available at the forecast date.

Forecast error should be measured with clearly defined signs and suitable metrics such as ME, MAD, RMSE, MAPE, or WAPE. Accuracy, bias, and uncertainty are different concepts. Forecasts become managerial plans through S&OP, aggregate planning, master scheduling, material requirements planning, capacity planning, and inventory decisions. Effective governance requires transparent data, documented judgement, scenario thinking, ethical use of technology, and continuous monitoring.

Key Terms

Forecasting: Systematic estimation of future demand or related quantities.

Demand: Quantity customers are willing and able to purchase under specified conditions.

Independent demand: Demand arising primarily from external customers.

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

Time series: Chronologically ordered observations of a variable.

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

Trend: Persistent long-term direction in a demand series.

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

Exponential smoothing: Recursive forecasting method that weights recent observations more heavily.

Regression: Statistical method relating a dependent variable to one or more explanatory variables.

Forecast error: Difference between actual demand and forecast demand.

Bias: Persistent tendency to overforecast or underforecast.

MAD: Mean absolute deviation, the average absolute forecast error.

RMSE: Root mean squared error, sensitive to large errors.

MAPE: Mean absolute percentage error, unsuitable for zero or near-zero actuals.

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

S&OP: Cross-functional process for reconciling demand, supply, and financial plans.

Scenario analysis: Examination of several coherent possible futures.

Bullwhip effect: Amplification of demand variability as information moves upstream in a supply chain.

Forecast reconciliation: Adjustment of forecasts across hierarchical levels to achieve coherent totals.

Review Questions

Short Answer Questions

1. Distinguish demand from observed sales and explain why the difference matters.
2. What are the principal components of a demand time series?
3. State the assumptions and formula of single exponential smoothing.
4. Why can MAPE be misleading for low-volume or intermittent demand?
5. Explain the role of S&OP in using forecasts.

Long Answer Questions

1. Compare qualitative, time-series, and causal forecasting methods, including their assumptions, strengths, and limitations.
2. Explain how a forecasting process should move from data preparation through validation to managerial implementation.
3. Derive and illustrate the use of moving averages and exponential smoothing with suitable examples.
4. Discuss forecast error, bias, and uncertainty. How should each influence management decisions?
5. Critically evaluate the opportunities and risks of artificial intelligence in demand forecasting.

Analytical / Application Questions

1. A retailer records demand of 240, 260, 250, and 290 units over four weeks. Calculate the three-week moving-average forecast for week 5 and explain one limitation of the result.
2. Actual demand over three periods is 500, 550, and 450 units, while forecasts are 480, 570, and 500 units. Using e_t = A_t − F_t, calculate ME and MAD and interpret the signs and units.
3. A manufacturer has a seasonal product, frequent stockouts, and a planned promotion whose timing is uncertain. Design a forecasting and planning process that separates constrained sales from unconstrained demand and identifies appropriate scenarios, metrics, and governance controls.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes demand from sales?

A. Demand is always lower than sales.

B. Sales may be constrained by stockouts or capacity, while demand may be higher.

C. Sales include only future customer requirements.

D. Demand is observed only after production is completed.

Correct Answer: B. Sales may be constrained by stockouts or capacity, while demand may be higher.

Question 2

Using a three-period moving average, what is the forecast when recent demands are 120, 135, and 150 units?

A. 125 units

B. 130 units

C. 135 units

D. 140 units

Correct Answer: C. 135 units

Question 3

In single exponential smoothing, increasing α generally makes the forecast:

A. Less responsive to recent observations

B. More responsive to recent observations

C. Independent of the previous forecast

D. Automatically seasonal

Correct Answer: B. More responsive to recent observations

Question 4

With forecast error defined as e_t = A_t − F_t, a negative mean error indicates, on average, that the organisation:

A. Underforecast demand

B. Overforecast demand

C. Had zero forecast error

D. Used a seasonal model

Correct Answer: B. Overforecast demand

Question 5

Which accuracy measure is undefined when an actual demand value is zero?

A. MAD

B. RMSE

C. MAPE

D. ME

Correct Answer: C. MAPE

Question 6

A causal forecasting model is most appropriate when:

A. Historical demand is unavailable and no explanatory information exists

B. Demand is related to predictors that are available at forecast time

C. All demand is random and independent of business conditions

D. The organisation wants to avoid validating forecasts

Correct Answer: B. Demand is related to predictors that are available at forecast time

Question 7

What is the main purpose of a naïve forecast?

A. To guarantee the most accurate forecast

B. To provide a simple benchmark for comparison

C. To remove all seasonality from data

D. To estimate supplier lead time

Correct Answer: B. To provide a simple benchmark for comparison

Question 8

Which practice is most likely to create data leakage in forecast evaluation?

A. Testing on a later time period

B. Using only information available at the forecast date

C. Randomly mixing future observations into the training data

D. Comparing against a seasonal naïve benchmark

Correct Answer: C. Randomly mixing future observations into the training data

Question 9

Which process reconciles demand, supply, capacity, and financial considerations across functions?

A. S&OP

B. ABC classification

C. Preventive maintenance

D. Value-stream mapping

Correct Answer: A. S&OP

Question 10

Why should constrained sales often be adjusted before forecasting future demand?

A. They may understate what customers would have purchased without the constraint.

B. They always exaggerate the true market size.

C. They eliminate the need for safety stock.

D. They make all forecasts seasonal.

Correct Answer: A. They may understate what customers would have purchased without the constraint.