Demand Analysis and Forecasting

Introduction

Inventory decisions are responses to anticipated demand. A firm must decide how much to order, when to replenish, where to position stock, and how much uncertainty to protect against. These decisions become difficult because demand is not directly observable in advance and because historical sales may reflect stockouts, promotions, price changes, assortment changes, competitor actions, and market disruptions rather than underlying customer need.

Demand analysis is therefore more than extrapolating a sales graph. It involves identifying the structure of demand, selecting an appropriate forecasting method, measuring forecast quality, incorporating managerial and market intelligence, and translating the resulting forecast into inventory policies. A statistically accurate forecast can still produce poor inventory outcomes if it is biased, poorly timed, aggregated at the wrong level, or disconnected from replenishment decisions.

Forecasting is particularly important in supply chains because lead times create a separation between the time at which a decision is made and the period in which demand occurs. The longer the replenishment lead time, the more future demand must be estimated. Forecasting also affects capacity, procurement, transportation, labour scheduling, working capital, and customer service.

Learning Objectives

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

1. Explain the distinction between demand, sales, and forecast requirements.
2. Diagnose trend, seasonality, cycles, irregular variation, intermittency, and demand distortion.
3. Select and apply qualitative, time-series, causal, and collaborative forecasting methods.
4. Calculate and interpret forecast-error measures, including bias, MAD, MSE, RMSE, and MAPE.
5. Construct simple moving-average, weighted-moving-average, and exponential-smoothing forecasts.
6. Evaluate forecast methods using both statistical and inventory-related criteria.
7. Explain how collaborative planning, forecasting, and replenishment improve supply-chain decisions.
8. Recognise the limitations, governance risks, and ethical implications of forecasting systems.

1. Demand, Sales and the Forecasting Task

Demand is the quantity customers are willing and able to purchase under specified conditions. Sales are the transactions actually recorded. Sales may be lower than demand when inventory is unavailable, orders are rationed, delivery capacity is constrained, or customers substitute another product. Consequently, a sales history is not automatically a demand history.

A useful conceptual relationship is:

This expression applies when unmet demand is lost or not recorded. If backorders are captured, observed orders may provide a better measure of demand, although cancellations and delayed fulfilment still require analysis.

Forecasting has several levels:

Forecasts should also be distinguished by forecast horizon. The lead-time horizon covers the period during which a replenishment decision cannot be changed. The planning horizon extends further and supports capacity, supplier, and budgeting decisions. Forecasts generally become less certain as the horizon increases.

A forecast is not a prediction of what must happen. It is an informed estimate conditional on available information, assumptions, and the chosen method. Forecasts should therefore be accompanied by an uncertainty estimate, review date, data definition, and owner.

2. Patterns in Demand Data

Before selecting a method, the analyst should plot the data, examine its granularity, identify unusual observations, and understand the business process that generated it.

2.1 Level and random variation

The level is the typical magnitude around which demand fluctuates. Random variation is short-term movement that cannot be systematically explained using currently available information. A stable item may have a constant level with small random deviations; a volatile item may have the same average but much greater dispersion.

2.2 Trend

A trend is a persistent upward or downward movement. It may result from market growth, product adoption, price changes, distribution expansion, or product decline. A trend can be approximately linear over a limited period, but assuming indefinite linear growth is usually unsafe.

2.3 Seasonality

Seasonality is a repeating pattern associated with a calendar or operating cycle, such as weekly restaurant demand, monthly salary-related purchases, or annual holiday sales. The seasonal period must be defined correctly: weekly observations may have a period of seven for daily data, while monthly data may have a period of twelve for annual seasonality.

2.4 Cycles

Cycles are longer and less regular movements, often related to economic conditions, industry investment, or product life cycles. Unlike seasonality, cycles do not necessarily repeat at a fixed interval.

2.5 Intermittent and lumpy demand

Intermittent demand contains many zero-demand periods and occasional non-zero events. Spare parts, emergency equipment, and low-volume components often exhibit this pattern. Lumpy demand combines intermittency with highly variable quantities. Ordinary moving averages may generate misleading forecasts because they smooth zeros and spikes without modelling the occurrence and size of demand separately.

2.6 Promotions, substitutions, and demand distortion

Promotional sales may represent forward buying rather than permanent demand growth. A stockout may depress observed sales and create a false impression of declining demand. Conversely, panic buying, order batching, minimum-order quantities, and allocation rules can amplify demand signals upstream. This phenomenon is associated with the bullwhip effect: order variability becomes greater than end-customer demand variability as information is delayed, aggregated, or distorted.

3. Forecasting Process and Data Preparation

A professional forecasting process normally follows these stages:

1. Define the decision, item, location, time bucket, horizon, and service objective.
2. Collect sales, orders, inventory availability, prices, promotions, returns, and relevant external variables.
3. Clean and reconcile the data while preserving an audit trail.
4. Diagnose demand patterns and exceptional events.
5. Establish a simple baseline forecast.
6. Select candidate methods and estimate parameters using historical data.
7. Evaluate forecasts using a holdout period or rolling-origin testing.
8. Incorporate documented business knowledge and collaborative inputs.
9. Publish the forecast with uncertainty, assumptions, and version control.
10. Monitor errors, bias, overrides, and inventory outcomes.

Data preparation requires care. Missing observations should not automatically be treated as zero demand. A zero during a verified stockout is a censored observation, not evidence that customers wanted nothing. Calendar changes, product substitutions, pack-size changes, returns, and one-time contracts should be recorded explicitly. Outliers should be investigated before removal; an extreme value may be an error, but it may also reveal a structural market change.

Aggregation can reduce noise but may hide local patterns. A forecast for a product family may be accurate while individual stock-keeping units remain poorly forecast. Forecasting should therefore be performed at the level at which decisions are made, while exploiting hierarchical relationships where appropriate.

4. Qualitative Forecasting Methods

Qualitative methods are valuable when historical data are limited, structural change is expected, or expert knowledge contains information not represented in the time series.

4.1 Executive opinion

Senior managers provide a consolidated estimate based on market knowledge, strategy, and commercial expectations. This method is fast but can be vulnerable to hierarchy, optimism, anchoring, and political pressure. It should be documented and compared with an objective baseline.

4.2 Sales-force composite

Sales representatives estimate demand in their territories or accounts. They may know customer projects and competitive developments, but they may also understate demand to obtain easier targets or overstate it to secure inventory.

4.3 Customer and market research

Surveys, pilot launches, conjoint studies, test markets, and customer interviews can provide insight into preferences and potential adoption. Stated intentions are not always reliable indicators of actual purchases, so research should be calibrated against observed behaviour.

4.4 Delphi method

The Delphi method obtains independent estimates from a panel of experts through multiple anonymous rounds. A coordinator summarises responses without identifying individuals, allowing participants to revise their estimates. It is useful for long-term or novel situations but can be time-consuming and dependent on panel composition.

4.5 Scenario planning

Scenario planning develops coherent alternative futures rather than a single point estimate. For example, a firm may construct base, high-growth, and disruption scenarios, each with assumptions about prices, competitors, regulation, and supply availability. Scenario planning is particularly useful for capacity and resilience decisions, although scenarios are not probability forecasts unless probabilities are explicitly assigned.

Qualitative input should not be treated as an unrestricted override. A sound governance practice records who changed the statistical forecast, why, by how much, for which period, and whether the change was later validated.

5. Time-Series Forecasting Methods

Time-series methods use the historical sequence of demand observations. Let Y_t denote actual demand in period t, and let F_t denote the forecast for that period or the forecast made at the relevant decision time.

5.1 Naïve forecast

For a stable series, the simplest baseline is:

For seasonal data with period m, a seasonal naïve forecast is:

These methods require little data and are difficult to improve upon for some volatile series. They are essential benchmarks.

5.2 Moving average

A k-period moving average is:

The method assumes that recent observations provide a reasonable estimate of the near-future level and that each of the last k observations deserves equal weight. A larger k produces more smoothing but responds more slowly to genuine change.

5.3 Weighted moving average

A weighted moving average assigns different weights to recent observations:

The weights are dimensionless and non-negative in the usual formulation. Greater weight on recent observations increases responsiveness but may increase sensitivity to noise.

5.4 Simple exponential smoothing

Simple exponential smoothing is appropriate for data with a relatively stable level and no systematic trend or seasonality:

Here α is the smoothing constant, Y_t is actual demand, and F_t is the previous forecast. All quantities have the same unit as demand. A high α responds rapidly to recent changes; a low value creates greater smoothing. The initial forecast must be specified, for example by using the first observation or an initial average.

Worked example. Suppose demand for three periods is 100, 120, and 110 units. Let the initial forecast for period 1 be 100 and α = 0.3. The forecast for period 2 is 100. For period 3:

F₃ = 0.3(120) + 0.7(100) = 106 units.

The forecast for period 4 is:

F₄ = 0.3(110) + 0.7(106) = 107.2 units.

The result is measured in units because it is a weighted sum of quantities measured in units.

5.5 Trend and seasonality

When a series contains trend, a level-only method tends to lag. Holt’s method estimates both level and trend. One common additive form is:

L_t is the estimated level, T_t the estimated trend, h the forecast horizon, and α and β smoothing constants between 0 and 1. This additive model is suitable when trend changes are approximately constant in units. Multiplicative trend or seasonal models may be preferable when changes are proportional to the level.

Seasonal methods estimate recurring seasonal effects. An additive seasonal model assumes that the seasonal effect is approximately constant in units; a multiplicative model assumes that seasonal amplitude grows with the level. The choice must be based on plots, business knowledge, and validation rather than convenience.

5.6 Intermittent demand methods

For intermittent demand, forecasting can separate the interval between non-zero demands from the demand size. Croston-type methods follow this logic, although different variants have different statistical properties and bias corrections. Such methods should be tested against alternatives, including a suitably designed probabilistic model, and evaluated using inventory outcomes rather than only conventional percentage errors.

6. Causal and Advanced Forecasting

Causal forecasting explains demand through predictor variables such as price, advertising, income, weather, competitor activity, population, or economic indicators. A simple linear regression model is:

Here Y_t is demand, X_jt are explanatory variables, β_j are coefficients, β₀ is the intercept, and ε_t is the unexplained error. The coefficient units depend on the variables. For example, if demand is units per week and price is currency per unit, the price coefficient has units of units per week per currency unit.

Regression can support what-if analysis, but association does not prove causation. Price may be reduced precisely when demand is weak, creating a misleading positive or negative relationship. Important considerations include omitted variables, multicollinearity, endogeneity, autocorrelated errors, changing relationships, and data leakage. Predictors used in historical testing must have been available at the time the forecast would actually have been made.

Price elasticity is often represented as:

The usual demand relationship produces a negative elasticity, although exceptional market conditions can produce different results. Elasticity estimates should be interpreted within the observed price range and market context.

Machine-learning methods, including gradient boosting, random forests, and neural networks, can model nonlinear relationships and large numbers of predictors. Their value depends on data quality, forecast horizon, explainability, stability, and operational integration. More complex models do not automatically produce better inventory decisions. A transparent benchmark, robust validation, and monitoring for drift remain necessary.

7. Forecast Accuracy and Error Measurement

Let the error in period t be:

A positive error indicates underforecasting; a negative error indicates overforecasting.

Here n is the number of evaluated periods. These formulas assume actual observations are available and, for MAPE, that Y_t ≠ 0. MAPE is undefined when actual demand is zero and unstable when actual demand is close to zero. Weighted absolute percentage error or scaled errors may be more appropriate for intermittent demand.

Worked error example. Actual demand is 100, 120, and 110 units; forecasts are 105, 115, and 108 units. Errors are −5, 5, and 2 units. Therefore:

ME = (−5 + 5 + 2) ÷ 3 = 0.67 units.

MAD = (5 + 5 + 2) ÷ 3 = 4 units.

MSE = (25 + 25 + 4) ÷ 3 = 18 units², and RMSE = √18 ≈ 4.24 units.

No single measure is universally best. MAD and RMSE are expressed in units and support operational interpretation. Bias identifies systematic over- or underforecasting. Percentage measures facilitate comparison across items but can be distorted by small or zero demand. Forecast evaluation should also consider service level, inventory, obsolescence, stockouts, and the cost asymmetry between overforecasting and underforecasting.

A rolling-origin evaluation repeatedly trains a method on an earlier period and tests it on the next period. This better represents real forecasting than randomly mixing observations, particularly when time order matters. Forecasts should be compared against a naïve benchmark, and statistical differences should not be mistaken for economically significant improvements.

8. From Forecast Error to Inventory Decisions

A forecast point estimate is insufficient for inventory control. Uncertainty affects safety stock, reorder points, capacity, and service commitments. If demand during lead time has mean μ_LT and standard deviation σ_LT, a basic reorder point is:

Here SS is safety stock and z is the standard-normal service factor associated with a chosen cycle-service target, under the approximation that lead-time demand is normally distributed and lead time is known and constant. These assumptions may be inappropriate for intermittent, highly skewed, promotional, or short-life-cycle demand.

If independent period demand has standard deviation σ_D and constant lead time is L periods, then:

This relationship assumes independent and identically distributed demand. If lead time is also variable, a common approximation is:

where μ_D is mean demand per period, σ_D its standard deviation, L mean lead time, and σ_L lead-time standard deviation. This approximation assumes independence between demand and lead time and is most defensible when distributions are reasonably well behaved.

Forecast error should be included in the uncertainty estimate. Using only historical demand variation can understate risk when the forecasting process is inaccurate. The correct inventory response depends on the service measure: cycle service level, fill rate, stockout frequency, lost sales, and backorder performance are not interchangeable.

9. Collaborative Demand Planning

Collaborative Planning, Forecasting and Replenishment (CPFR) coordinates trading partners around shared information and agreed processes. Participants may exchange point-of-sale data, inventory positions, promotions, new-product plans, event calendars, and exception alerts.

A practical collaborative process includes:

1. Agreeing on product, location, calendar, unit, and data definitions.
2. Establishing a baseline statistical forecast.
3. Sharing commercial events and supply constraints.
4. Reconciling differences through structured exception management.
5. Creating a jointly agreed forecast and replenishment plan.
6. Assigning accountability for decisions and overrides.
7. Reviewing forecast accuracy, bias, service, and inventory outcomes.

Collaboration is not simply adding every partner’s estimate. It requires incentives, data quality, trust, confidentiality controls, and clear decision rights. The bullwhip effect can be reduced by sharing consumption data, shortening information delays, reducing artificial order batching, and aligning promotional and replenishment policies.

Consensus forecasting can nevertheless fail when participants manipulate forecasts for targets, negotiate rather than analyse, or use inconsistent definitions. Statistical forecasts should remain visible as a baseline, and overrides should be evaluated after the fact.

10. Illustrative Case: Meridian Medical Supplies

Background

Meridian Medical Supplies distributes sterile procedure kits to hospitals. Monthly sales of one kit increased sharply during a public-health emergency, then fell as demand normalised. Several months also contained stockouts. The sales team proposed using the most recent month as the forecast because demand appeared to be recovering.

Problem

The proposed method treated stockout-affected sales as true demand and ignored the difference between emergency demand, normal seasonal demand, and replenishment orders placed by hospitals to rebuild depleted inventories. It produced high purchase quantities and rising expiry risk.

Analysis

The supply-chain analyst reconstructed demand using shipment records, hospital order cancellations, inventory availability, and emergency allocation data. Stockout months were flagged rather than treated as ordinary observations. Demand was segmented into emergency and routine periods. A seasonal naïve benchmark, a three-month moving average, and exponential smoothing were tested using rolling-origin validation. Forecasts were evaluated using MAD, bias, and service-related outcomes.

The three-month average responded too slowly during the initial decline. The latest-observation method reacted quickly but overemphasised a single replenishment event. Exponential smoothing with a moderate smoothing constant produced lower absolute error and less positive bias. The commercial team added a documented hospital tender expected to begin in month six; this was treated as a causal event with a specified probability rather than embedded invisibly in the baseline.

Approach and solution

Meridian adopted a two-layer process. The baseline forecast was generated statistically from adjusted demand history. A consensus review then considered confirmed tenders, planned promotions, and known contract changes. Safety stock was recalculated using forecast-error variability and the required hospital service target. Slow-moving kits were subject to expiry review, while high-criticality kits received a different service policy from low-criticality accessories.

Outcome

The company reduced systematic overforecasting, lowered expiry exposure, and preserved service for critical hospitals. The most important improvement did not come from selecting a sophisticated algorithm; it came from distinguishing true demand from constrained sales, segmenting exceptional conditions, and linking forecasts to differentiated inventory policies.

Learning points

1. Sales history requires operational interpretation.
2. A forecast method should be judged on future-like validation data.
3. Confirmed commercial information should be documented separately from assumptions.
4. Inventory decisions require uncertainty and service objectives, not only a point forecast.
5. Segmentation by criticality and demand behaviour often improves outcomes more than universal modelling.

11. Challenges, Governance and Contemporary Practice

Forecasting systems face structural breaks, missing data, new products, changing customer behaviour, supplier disruption, cyber incidents, and climate-related volatility. Historical accuracy may deteriorate when the environment changes. Forecast models should therefore be monitored for data drift, parameter instability, bias by customer or region, and deteriorating service outcomes.

AI-enabled forecasting introduces additional governance requirements. Organisations should document data provenance, feature availability, model version, assumptions, performance by segment, and override history. Models should not use protected or inappropriate personal information, expose confidential customer data, or make unexplained allocation decisions that disadvantage particular customers. Human review is particularly important when forecasts drive scarce-resource allocation, healthcare supply, food access, or emergency response.

Forecasting teams should avoid false precision. A forecast of 10,247 units may imply a level of certainty unsupported by the data. Prediction intervals, scenarios, and rounded operational quantities may communicate uncertainty more honestly. Explainability is valuable not because every model must be simple, but because managers need to challenge implausible assumptions and understand consequences.

Best practice includes:

• use a transparent baseline and test improvement against it;
• forecast at the decision-relevant level while reconciling hierarchies;
• correct or flag stockouts, promotions, substitutions, and one-off events;
• select metrics consistent with demand characteristics and cost consequences;
• evaluate methods with time-ordered validation;
• monitor bias as well as magnitude of error;
• connect forecast uncertainty to safety stock and service policy;
• record overrides and assess whether they improve outcomes;
• segment products by value, criticality, volatility, and intermittency;
• review models after structural market changes.

Chapter Summary

Demand forecasting supports inventory, procurement, capacity, and service decisions by estimating future requirements under uncertainty. Sales are not always equivalent to demand because stockouts, rationing, substitutions, and promotions can distort observed history. Effective analysis begins with clear definitions, suitable time buckets, data cleaning, pattern diagnosis, and recognition of level, trend, seasonality, cycles, intermittency, and exceptional events.

Qualitative methods contribute market knowledge when data are sparse or structural change is expected, while time-series methods provide disciplined baselines. Naïve, moving-average, weighted-average, exponential-smoothing, trend, seasonal, intermittent, causal, and machine-learning methods have different assumptions and applicability conditions. Forecasts must be evaluated through time-ordered testing and compared with simple benchmarks.

Bias, MAD, MSE, RMSE, and percentage measures describe different aspects of error. No metric alone determines inventory performance. Forecast uncertainty must be connected to safety stock, reorder points, lead-time variability, service objectives, and the cost of shortages and excess. Collaborative planning improves information quality when partners share definitions, consumption data, events, and accountability. Governance is essential to control overrides, model drift, data leakage, unfair allocation, and false precision. The central managerial principle is that forecasting is not merely a statistical exercise: it is a coordinated decision process linking demand information to inventory action.

Key Terms

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

Observed sales: Transactions recorded, which may be constrained by supply availability.

Forecast horizon: Future period for which a forecast is prepared.

Level: Typical magnitude around which demand varies.

Trend: Persistent upward or downward movement in demand.

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

Cycle: Longer, less regular movement in demand.

Intermittent demand: Demand with many zero periods and occasional non-zero observations.

Bullwhip effect: Amplification of demand variability as information moves upstream.

Naïve forecast: Forecast based on the latest observation or corresponding seasonal observation.

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

Exponential smoothing: Forecasting method that assigns geometrically declining weights to past observations.

Causal forecasting: Forecasting using explanatory variables believed to influence demand.

Forecast error: Difference between actual demand and forecast demand.

Bias: Systematic tendency to overforecast or underforecast.

MAD: Mean absolute deviation, expressed in demand units.

RMSE: Root mean squared error, which penalises large errors and is expressed in demand units.

CPFR: Collaborative Planning, Forecasting and Replenishment between supply-chain partners.

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

Reorder point: Inventory position at which replenishment is initiated.

Review Questions

Short Answer Questions

1. Distinguish between true demand and observed sales.
2. What is the difference between seasonality and a business cycle?
3. State the assumptions of simple exponential smoothing.
4. Why can MAPE be unsuitable for intermittent demand?
5. What is the purpose of maintaining a naïve forecast benchmark?

Long Answer Questions

1. Explain the complete demand-forecasting process from problem definition through monitoring.
2. Compare moving averages, exponential smoothing, and causal regression in terms of assumptions, responsiveness, data requirements, and limitations.
3. Explain how stockouts, promotions, and substitutions can distort historical demand data.
4. Discuss how forecast error should influence safety-stock and service-level decisions.
5. Evaluate the role of CPFR in reducing the bullwhip effect and identify conditions required for successful collaboration.

Analytical / Application Questions

1. A product has monthly demand of 240, 260, 250, and 290 units. Calculate the three-month moving-average forecast for month five and explain how the choice of window affects responsiveness.
2. Actual demand for three periods is 80, 100, and 120 units, while forecasts are 90, 95, and 110 units. Calculate mean error, MAD, and RMSE, and interpret the direction of bias.
3. A retailer experiences repeated stockouts during a promotion and then observes sharply lower sales after the promotion. Design a forecasting and inventory-analysis approach that distinguishes underlying demand, forward buying, lost sales, and post-promotion effects.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes demand from observed sales?

A. Demand is always equal to shipped quantity.

B. Observed sales may be lower than demand because of stockouts.

C. Demand includes only fulfilled orders.

D. Observed sales are independent of product availability.

Correct Answer: B. Observed sales may be lower than demand because of stockouts.

Question 2

A three-period moving-average forecast gives what weight to each included observation?

A. Equal weight of one-third

B. Weight of one-half for the latest observation

C. Weight proportional to the observation value

D. Zero weight for the oldest observation

Correct Answer: A. Equal weight of one-third

Question 3

If actual demand is 100 units and the forecast is 110 units, what is the forecast error using e_t = Y_t − F_t?

A. 10 units

B. −10 units

C. 100 units

D. −110 units

Correct Answer: B. −10 units

Question 4

Which measure is expressed in the same units as demand and penalises large errors more strongly?

A. MAPE

B. Mean error

C. RMSE

D. Forecast horizon

Correct Answer: C. RMSE

Question 5

Simple exponential smoothing is most appropriate when demand has:

A. A stable level without systematic trend or seasonality

B. Only zero observations

C. A known deterministic upward trend

D. No historical observations at all

Correct Answer: A. A stable level without systematic trend or seasonality

Question 6

Why can MAPE be problematic for intermittent demand?

A. It always produces negative values.

B. It cannot compare percentages.

C. It is undefined when actual demand is zero.

D. It is measured in squared units.

Correct Answer: C. It is undefined when actual demand is zero.

Question 7

In a causal forecasting model, a predictor variable is used primarily to:

A. Replace all historical observations

B. Explain or estimate changes in demand

C. Guarantee that correlation proves causation

D. Remove the need for validation

Correct Answer: B. Explain or estimate changes in demand

Question 8

Under constant lead time and independent period demand, lead-time-demand standard deviation is commonly approximated as:

A. σ_D ÷ L

B. σ_D + L

C. σ_D√L

D. L²σ_D

Correct Answer: C. σ_D√L

Question 9

Which practice helps reduce the bullwhip effect?

A. Withholding point-of-sale information

B. Increasing order batching without coordination

C. Sharing consumption data and reducing information delays

D. Changing product definitions between partners

Correct Answer: C. Sharing consumption data and reducing information delays

Question 10

Why should forecast overrides be recorded and reviewed?

A. To prevent any use of managerial knowledge

B. To identify whether judgement improves forecast and inventory outcomes

C. To make all forecasts identical

D. To eliminate the need for statistical baselines

Correct Answer: B. To identify whether judgement improves forecast and inventory outcomes