Demand Analysis and Forecasting

Introduction

Inventory decisions are decisions under uncertainty. A purchasing manager must decide how much to order before knowing exactly how much customers will buy; a warehouse manager must allocate storage capacity before receiving future replenishment; and a production planner must schedule labour and equipment before demand is observed. Demand analysis and forecasting provide structured ways to reduce this uncertainty, although they can never eliminate it.

A forecast is not a prediction of what must happen. It is a conditional estimate of future demand based on information, assumptions, and a specified horizon. Its managerial value depends not only on statistical accuracy but also on whether it improves replenishment, service, capacity, and financial decisions. A highly accurate forecast may have little value if it arrives too late or is not connected to an inventory policy.

This chapter examines qualitative methods, time-series models, causal forecasting, forecast accuracy and bias, seasonality, intermittent demand, and the economic value of forecasting.

Learning Objectives

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

1. Explain the role of demand analysis in inventory and warehousing decisions.
2. Select qualitative, time-series, and causal forecasting methods appropriately.
3. Calculate and interpret moving averages, exponential smoothing, seasonal indices, and forecast errors.
4. Diagnose bias, seasonality, trend, and intermittent demand.
5. Evaluate forecasts using accuracy measures while recognising their limitations.
6. Connect forecast quality with safety stock, service levels, capacity, and managerial value.
7. Design a practical forecasting process incorporating technology, judgement, and governance.

1. Demand, Sales, and the Forecasting Problem

Demand is the quantity customers are willing and able to purchase under specified conditions. Sales are the transactions actually observed. The two are not always equal. Stockouts, credit restrictions, lost orders, rationing, delivery failures, and competitor actions can make observed sales lower than unconstrained demand. For inventory planning, analysts should distinguish:

• Observed sales: recorded transactions.
• Lost demand: demand that could not be fulfilled.
• Independent demand: demand for finished goods or service items driven by external customers.
• Dependent demand: demand calculated from the requirements of another item, as in components in a bill of materials.
• Order demand: demand represented by customer orders, which may be lumpy, delayed, or cancelled.

The demand horizon should match the decision. A daily forecast may support labour and picking schedules; a weekly forecast may support replenishment; a monthly or annual forecast may support supplier contracts, facility capacity, and capital investment. Forecast uncertainty usually increases as the horizon lengthens.

A useful decomposition is:

where Y_t is demand at time t, L_t is the level, T_t is trend, S_t is seasonal effect, and R_t is irregular or random variation. This is an additive representation and is appropriate when seasonal fluctuations are approximately constant in absolute size. A multiplicative representation is:

This is more suitable when seasonal variation grows in proportion to the level. The components are conceptual and need not be uniquely identifiable without modelling assumptions.

2. Forecasting as a Management Process

Forecasting is more than selecting an algorithm. A robust process includes:

1. Define the item, location, unit, horizon, and purpose.
2. Collect sales, inventory, price, promotion, calendar, and external-driver data.
3. Correct data problems and identify censored demand caused by stockouts.
4. Explore level, trend, seasonality, outliers, and intermittency.
5. Select candidate methods and establish a simple benchmark.
6. Estimate parameters using historical data.
7. Validate using a time-ordered holdout period or rolling-origin testing.
8. Generate the forecast and its uncertainty estimate.
9. Apply documented managerial adjustments where justified.
10. Monitor errors, bias, service outcomes, and forecast value.

The forecast should be made at the level at which decisions are taken. Forecasting at a highly aggregated level can be more stable, while forecasting at an excessively detailed level can create noise. Forecasts can subsequently be reconciled so that warehouse, regional, product-family, and total forecasts are logically consistent.

3. Qualitative Forecasting Methods

Qualitative methods are essential when historical data are limited, structural change is expected, or experts possess information not represented in the data. They are particularly useful for new products, market entry, technology substitution, regulatory changes, and major promotions.

Expert judgement and executive opinion

Managers estimate future demand using experience, customer knowledge, and business intelligence. This method is fast but vulnerable to optimism, anchoring, political pressure, and overconfidence. Judgements should be documented separately from the statistical baseline so that their incremental value can be evaluated.

Sales-force composite

Sales representatives provide estimates for their territories or accounts. Aggregating these estimates can reveal customer-level intelligence, but incentives may distort submissions. For example, sales personnel may understate demand to obtain easier targets or overstate it to secure inventory.

Customer intention surveys

Customers are asked about planned purchases. This is useful for durable goods, business-to-business contracts, and planned capital expenditure, but stated intentions may not become purchases. Survey design, sampling, non-response, and changing economic conditions affect reliability.

Delphi method

A panel of experts provides estimates independently through several rounds. A facilitator summarises responses without identifying individuals, enabling the panel to revise estimates after reviewing group information. Delphi reduces direct conformity pressure but can be slow and remains dependent on expert selection.

Analogy and market research

Demand for a new item may be inferred from a similar product, another market, pilot stores, or test launches. The analogy must be adjusted for price, distribution, customer segment, and product maturity; otherwise apparent similarity may be misleading.

Qualitative methods are strongest when structured, evidence-based, and combined with a statistical baseline. They should not be treated as automatically superior to data-driven methods merely because they include managerial knowledge.

4. Time-Series Forecasting

Time-series methods forecast future demand from the historical pattern of the variable itself. They are appropriate when the future is expected to resemble the past sufficiently and when causal data are unavailable or difficult to forecast.

Naïve forecast

The simplest benchmark sets the next forecast equal to the latest observation:

For seasonal demand, a seasonal naïve forecast uses the observation from the same season in the previous cycle:

where m is the seasonal period and h is the forecast horizon. Benchmarks are important: a complex model should improve on a simple, transparent alternative.

Moving average

A simple moving average of the last n observations is:

All observations receive equal weight. A larger n smooths random variation but responds more slowly to level changes. The method assumes that recent observations are reasonably representative and does not explicitly model trend or seasonality.

A weighted moving average is:

where weights are non-negative and sum to one. The most recent observation is usually assigned the greatest weight when recent information is more relevant.

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

The forecast is lower than the latest observation because the average smooths the upward movement.

Simple exponential smoothing

Simple exponential smoothing updates the forecast using the latest error:

where 0 < α ≤ 1 is the smoothing constant, Y_t is actual demand, and F_t is the forecast for period t. The units of F and Y are the same. A high α responds rapidly to changes; a low value produces greater smoothing. The method is intended for demand with a relatively stable level and no systematic trend or seasonality.

Worked example. If Y_t = 150, F_t = 135, and α = 0.30, then:

The next forecast is 139.5 units, often rounded only when an operational decision requires whole units.

Trend-adjusted exponential smoothing

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

Here L_t is the estimated level, T_t the estimated trend, 0 < α ≤ 1, 0 ≤ β ≤ 1, and h the number of periods ahead. Initial values must be specified, and the method may over-project a temporary trend.

Seasonal forecasting

Seasonality is a repeating pattern associated with calendar or operating cycles. Seasonal indices can be estimated using historical ratios to a suitable level or trend estimate. In a multiplicative model, an index above 1 indicates demand above the average level; an index below 1 indicates demand below it. For a complete seasonal cycle, the average seasonal index is approximately 1, subject to estimation and rounding.

A forecast using a deseasonalised baseline is:

For additive seasonality:

The choice depends on whether seasonal amplitude is constant or proportional to the level. Promotions, moving holidays, stockouts, and changing assortment can make apparent seasonality difficult to estimate.

ARIMA and advanced time-series models

Autoregressive integrated moving-average models represent dependence on previous observations and previous errors. In practice, model identification, stationarity, differencing, parameter estimation, residual diagnostics, and out-of-sample validation are essential. More complex models are not automatically more accurate. For many inventory items, the main benefit of ARIMA or machine-learning models is achieved only when data quality, item segmentation, and deployment discipline are strong.

5. Causal Forecasting

Causal models explain demand through drivers such as price, advertising, income, weather, competitor activity, distribution, or service level. A linear regression model may be written as:

where Y_t is demand, X_jt are explanatory variables, β_j are coefficients, β₀ is the intercept, and ε_t is unexplained error. Coefficients describe conditional associations under the model; they should not automatically be interpreted as causal effects unless the research design and assumptions support causality.

A price coefficient is often negative, but the estimate may be confounded by promotions, product availability, or managers lowering prices when demand is weak. Multicollinearity makes it difficult to separate highly correlated drivers. Relationships can also change after competitor entry or policy change.

Causal forecasts are valuable when future drivers can themselves be forecast or are controlled by management. They are especially useful for promotion planning, scenario analysis, capacity planning, and evaluating price changes. Analysts should examine residual patterns, parameter stability, outliers, structural breaks, and forecast performance on periods not used for estimation.

6. Forecast Accuracy, Error, and Bias

Let the error be:

A positive error means actual demand exceeded the forecast; a negative error means the forecast exceeded actual demand. The sign convention must be stated because some organisations define error in the opposite direction.

Common measures

For MAPE, Y_t must not be zero. Percentage metrics can also overweight small-demand items. WAPE is not an average of item-level percentages and can be dominated by high-volume items. MAE is often easier to connect to units of safety stock and capacity.

Worked accuracy example. Suppose actual demand is 100, 120, and 80 units, while forecasts are 90, 130, and 100 units. Errors are 10, −10, and −20 units.

Thus, the forecast overestimated demand on average by 6.67 units under the stated convention. The MAE is:

The MAPE is:

Bias and tracking signal

Accuracy and bias are different. A model can have a small average absolute error but a persistent directional error. Cumulative forecast error is:

A commonly used tracking signal is:

where MAD_t is a chosen estimate of mean absolute deviation, often updated over time. Control limits must be defined by the organisation; there is no universal threshold that is valid for every demand environment. A sustained positive or negative tracking signal suggests investigation, not automatic model replacement.

7. Intermittent and Lumpy Demand

Intermittent demand contains many zero-demand periods. Lumpy demand combines intermittency with substantial variation in non-zero quantities. Spare parts, emergency maintenance items, and slow-moving SKUs commonly exhibit these patterns. Standard moving averages can generate misleading forecasts because they blur the timing and size of demand.

Croston-type methods separately estimate the size of non-zero demand and the interval between non-zero demands. A forecast is then based on estimated demand size divided by estimated interval. Variants such as SBA and TSB adjust bias or incorporate the probability that an item remains active. These methods require careful implementation and should be compared with simple benchmarks and inventory outcomes.

Intermittent-demand analysis should also consider whether zero sales mean zero demand or a stockout. If an item was unavailable, treating the observation as true zero demand biases the forecast downward. Item segmentation by demand frequency, value, criticality, and lead time supports differentiated methods and review policies.

8. Forecasts and Inventory Decisions

A forecast is an input to, not a substitute for, an inventory policy. If lead-time demand is forecast as μ_L and the standard deviation of lead-time demand is σ_L, a basic reorder point is:

where safety stock under a normal-demand approximation is often represented as:

Here z is the standard-normal quantile corresponding to a chosen cycle-service probability. This formulation assumes that lead-time demand is reasonably represented by a distribution with stable parameters and that the lead time is known or its variability is incorporated into σ_L. If demand and lead time are both variable, their joint distribution should be considered rather than simply applying a demand-only formula.

Forecast error should be included in uncertainty estimates. A forecast point estimate without an error distribution can encourage false precision. Inventory managers must also consider supplier minimums, order cycles, capacity constraints, substitution, perishability, criticality, and the asymmetric costs of shortage and excess.

Forecast aggregation can improve accuracy because positive and negative errors may offset at higher levels. However, inventory is often held at item-location level, where aggregate accuracy may conceal serious SKU-level errors. Forecasts and replenishment rules should therefore be evaluated at the level of the operational decision.

9. Forecast Value and Governance

Forecast value is the improvement in business outcomes attributable to forecasting, not merely a lower statistical error. Relevant outcomes include:

• reduced stockouts and emergency shipments;
• lower average inventory and obsolescence;
• improved warehouse utilisation;
• better supplier and labour planning;
• improved customer service;
• lower expediting and substitution costs.

A useful evaluation compares a proposed process with a credible baseline. The comparison should control for changes in assortment, service targets, lead times, promotions, and supply constraints. A statistically better forecast may not create value if it leads to excessive inventory, is too complex to maintain, or is ignored by planners.

Governance should define forecast ownership, data definitions, model approval, override authority, audit trails, and performance review. Overrides should record the reason, author, date, affected horizon, and outcome. This creates a learning system rather than a hidden layer of unmeasured judgement.

10. Technology, AI, and Contemporary Practice

Modern forecasting platforms can process large numbers of item-location combinations, incorporate calendars and promotions, and automate model selection. Machine-learning methods such as gradient boosting and neural networks may capture nonlinear relationships, but they require sufficient historical data, stable feature definitions, and careful validation. They can fail under regime change, sparse data, data leakage, or unanticipated events.

AI-generated forecasts should be explainable enough for planners to understand major drivers and uncertainty. Human review is valuable when the system encounters a new product, disruption, promotion, or policy change; however, review should be evidence-based rather than an invitation to override every statistical result. Privacy, access control, cybersecurity, model drift, and accountability are governance concerns. Forecasting systems should not use confidential customer information beyond legitimate and authorised purposes.

Contemporary best practice includes probabilistic forecasts, scenario forecasts, hierarchical reconciliation, demand sensing, and collaborative planning with suppliers and customers. Probabilistic forecasts represent a distribution or prediction interval rather than a single number, enabling inventory decisions to reflect risk explicitly.

11. Illustrative Case: Forecasting at Meridian Medical Supplies

Background

Meridian Medical Supplies distributes diagnostic consumables to hospitals. One test-strip SKU has a 14-day supplier lead time, weekly ordering, and a high shortage cost because hospitals may postpone procedures when the item is unavailable. Demand is affected by routine usage, public-health campaigns, and occasional hospital tenders.

Problem

The company used a six-week moving average for all items. Planners observed repeated shortages during campaigns and excess stock afterward. Reported sales also included several stockout weeks, so the historical series understated true demand.

Analysis

An analyst first separated unconstrained demand estimates from recorded sales by identifying stockout days and reviewing back orders. The item showed weekly seasonality and campaign-related spikes. A seasonal naïve benchmark, a seasonal exponential-smoothing model, and a causal model containing campaign indicators were tested using rolling-origin validation. Performance was assessed using MAE, WAPE, bias, stockout frequency, and inventory cost rather than MAPE alone, because some weeks had very low demand.

The causal model performed best during campaign periods but was not consistently superior in ordinary weeks. The company therefore adopted a segmented approach: seasonal smoothing for normal weeks and a documented campaign uplift based on the confirmed campaign schedule. Forecast intervals were passed to the inventory policy, and lead-time uncertainty was included in the safety-stock calculation.

Outcome and managerial interpretation

The process reduced systematic underforecasting during known campaigns and prevented the moving average from carrying temporary peaks too far into the future. Planners retained authority to adjust for hospital tenders, but each adjustment was recorded. Performance was reviewed monthly, and the model was recalibrated when supplier lead time changed.

Learning points

1. Sales history is not always demand history.
2. A single model for every SKU is rarely appropriate.
3. Forecast evaluation must include operational outcomes.
4. Known causal events should be represented explicitly.
5. Forecast governance and override records are essential for learning.

Challenges and Limitations

Forecasts are limited by data quality, structural breaks, changing customer behaviour, new products, cannibalisation, and unforecastable shocks. A model may appear accurate because the test period is unusually stable. Back-testing cannot fully reproduce unprecedented events. Forecast accuracy may also be improved by aggregation while the individual SKU forecasts remain unsuitable for replenishment.

Managers should avoid false precision, indiscriminate use of complex algorithms, and judging forecasts only after observing actual demand without considering information available at the time. Forecasts should be compared on consistent horizons and data vintages. Ethical practice requires transparent assumptions, responsible use of customer and employee data, and accountability for decisions influenced by automated systems.

Best-Practice Checklist

• Define demand, sales, units, horizon, and decision level clearly.
• Correct or flag stockout-censored observations.
• Use a naïve or seasonal-naïve benchmark.
• Match model structure to level, trend, seasonality, and intermittency.
• Validate chronologically rather than randomly shuffling time series.
• Report both accuracy and bias.
• Use measures appropriate to scale and zero demand.
• Separate baseline forecasts from documented overrides.
• Link forecast uncertainty to safety stock and service policy.
• Evaluate financial and service outcomes, not only forecast error.
• Monitor drift, structural change, and model governance.

Chapter Summary

Demand forecasting supports purchasing, replenishment, warehousing, production, capacity, and service decisions, but it does not remove uncertainty. Demand must be distinguished from observed sales, especially when stockouts censor the data. Qualitative methods contribute expert and market knowledge, while time-series methods exploit historical patterns and causal methods relate demand to explanatory drivers.

Moving averages smooth observations; exponential smoothing updates forecasts in response to recent errors; trend and seasonal models represent systematic patterns; causal models support scenario and driver-based analysis. Intermittent demand requires special treatment because zeros may reflect either genuine absence of demand or lack of availability.

Forecast accuracy should be evaluated with measures such as ME, MAE, RMSE, MAPE, and WAPE, selected with awareness of their assumptions and weaknesses. Bias is distinct from random error and should be monitored. Ultimately, forecast value is determined by better inventory and service outcomes, not by statistical performance alone. Effective practice combines sound data, appropriate models, documented judgement, uncertainty-aware inventory policies, and strong governance.

Key Terms

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

Forecast: A conditional estimate of future demand.

Independent demand: Demand generated by external customers.

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

Time series: Ordered observations recorded over time.

Level: Underlying average magnitude of a demand series.

Trend: Sustained upward or downward movement.

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

Intermittent demand: Demand containing many zero-demand periods.

Lumpy demand: Intermittent demand with substantial variation in non-zero quantities.

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

Exponential smoothing: Forecasting method that weights recent information more heavily.

Causal forecasting: Forecasting based on explanatory variables or demand drivers.

Forecast error: Difference between actual demand and forecast demand.

Bias: Persistent directional tendency to overforecast or underforecast.

MAE: Mean absolute error measured in demand units.

WAPE: Aggregate weighted absolute percentage error.

Safety stock: Inventory held to protect against uncertainty.

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

Forecast value: Business benefit produced by the forecasting process.

Review Questions

Short Answer Questions

1. Why may observed sales be lower than actual customer demand?
2. State one assumption of simple exponential smoothing.
3. What is the difference between forecast accuracy and forecast bias?
4. Why is MAPE problematic when actual demand is zero?
5. Give two examples of causal variables that may influence demand.

Long Answer Questions

1. Compare qualitative, time-series, and causal forecasting methods, including their appropriate applications and limitations.
2. Explain how seasonality can be represented using additive and multiplicative models.
3. Discuss how forecast errors should influence inventory and safety-stock decisions.
4. Explain the methodological challenges of forecasting intermittent and lumpy demand.
5. Design a governance process for statistical forecasts and managerial overrides.

Analytical / Application Questions

1. Actual weekly demand for three weeks is 200, 240, and 220 units. Calculate the three-week moving-average forecast for week four and explain how the result responds to the observations.
2. A forecast system produces errors of 20, −5, and −15 units. Calculate the mean error and interpret its direction using e_t = Y_t − F_t.
3. A warehouse experiences repeated stockouts during promotions. Propose a forecasting and evaluation approach that distinguishes censored demand, seasonal effects, and known promotional drivers.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes demand from observed sales?

A. Demand always equals the quantity shipped

B. Observed sales may be lower than demand when stockouts occur

C. Demand includes only cancelled orders

D. Observed sales are independent of product availability

Correct Answer: B. Observed sales may be lower than demand when stockouts occur

Question 2

A three-period moving-average forecast uses demands of 90, 120, and 150 units. What is the forecast for the next period?

A. 110 units

B. 120 units

C. 130 units

D. 150 units

Correct Answer: B. 120 units

Question 3

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

A. Less responsive to recent observations

B. More responsive to recent observations

C. Unaffected by actual demand

D. Automatically seasonal

Correct Answer: B. More responsive to recent observations

Question 4

Using e_t = Y_t − F_t, a positive mean error indicates:

A. Systematic overforecasting

B. Systematic underforecasting

C. Zero forecast variance

D. Perfect accuracy

Correct Answer: B. Systematic underforecasting

Question 5

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

A. MAE

B. RMSE

C. MAPE

D. ME

Correct Answer: C. MAPE

Question 6

Which method is most directly designed to use explanatory variables such as price and advertising?

A. Causal regression

B. Seasonal naïve forecasting

C. Simple moving average

D. Croston-type forecasting

Correct Answer: A. Causal regression

Question 7

Intermittent demand is characterised primarily by:

A. A constant positive demand every period

B. Many periods with zero demand

C. Only a long-term upward trend

D. A guaranteed seasonal cycle

Correct Answer: B. Many periods with zero demand

Question 8

In a multiplicative seasonal model, a seasonal index of 1.20 indicates that demand is approximately:

A. 20% below the baseline

B. Equal to the baseline

C. 20% above the baseline

D. 120 units regardless of the baseline

Correct Answer: C. 20% above the baseline

Question 9

If actual demand is 100 units and the forecast is 90 units, what is the forecast error under e_t = Y_t − F_t?

A. −10 units

B. 0 units

C. 10 units

D. 190 units

Correct Answer: C. 10 units

Question 10

Which statement best describes forecast value?

A. It is always identical to lower MAPE

B. It is the business improvement created by using the forecast

C. It is the number of models in a forecasting system

D. It is the forecast horizon measured in days

Correct Answer: B. It is the business improvement created by using the forecast