Demand Analysis and Forecasting

Introduction

Inventory is held to reconcile the timing and scale of supply with uncertain customer demand. A firm that forecasts too low may experience stockouts, lost sales, emergency procurement, expediting costs and damaged service reputation. A firm that forecasts too high incurs carrying cost, obsolescence, markdowns and unnecessary working-capital investment. Demand forecasting is therefore not an attempt to predict the future with certainty; it is a structured process for producing decision-relevant estimates and quantifying the uncertainty surrounding them.

In logistics, a forecast supports purchasing, production, transport capacity, warehouse staffing, replenishment, safety-stock decisions and financial planning. The appropriate forecasting method depends on the demand pattern, forecast horizon, data quality, supply-chain position and cost of error. A sophisticated method applied to poor data can be less useful than a simple, transparent method applied consistently.

This chapter develops a managerial and quantitative framework for analysing demand, measuring forecast performance and selecting forecasting methods. It covers time-series models, causal and collaborative forecasting, intermittent demand, bias control, forecast value and contemporary analytical tools.

Learning Objectives

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

1. Distinguish demand, sales, orders and dependent demand, and diagnose trend, seasonality, cycles, randomness and intermittency.
2. Select suitable time-series, causal, intermittent-demand and collaborative forecasting methods.
3. Calculate and interpret forecast error, MAD, MSE, RMSE, MAPE, WAPE, tracking signal and bias.
4. Apply moving averages, exponential smoothing, seasonal indices and regression models under stated assumptions.
5. Evaluate forecast performance in relation to inventory and service decisions rather than relying on a single accuracy statistic.
6. Design a governance process that combines statistical forecasts, expert knowledge, partner information and ethical use of data.

1. Demand as an Inventory Variable

Demand is the quantity customers wish to obtain during a specified period, whereas sales are the quantities actually fulfilled. When stockouts occur, observed sales may underestimate true demand. Orders may be placed in batches, distorted by promotions, minimum-order quantities or rationing, and therefore need not represent final consumption.

Several distinctions are operationally important:

• Independent demand arises from external customers and must generally be forecast, such as demand for finished products and spare parts.
• Dependent demand is derived from the requirement for another item. If a product requires three components and the production plan is known, component requirements can often be calculated through material requirements planning rather than forecast independently.
• Primary demand is market demand before the firm’s availability constraints.
• Derived demand results from another activity, such as transport demand created by production or component demand created by a bill of materials.
• Demand during lead time is the demand that occurs between placing and receiving a replenishment order. Its distribution is central to reorder-point and safety-stock decisions.

A forecast must specify the item, location, time bucket, unit of measure, customer population and information cut-off date. Forecasting “monthly demand” is incomplete unless the month, product-location combination and treatment of returns, substitutions and stockouts are defined.

2. Recognising Demand Patterns

A demand series can be represented conceptually as a combination of systematic and irregular components. A commonly used additive representation is:

where Y_t is observed demand in period t, T_t is trend, S_t is seasonal effect, C_t is cyclical effect and I_t is irregular variation. The additive form is appropriate when seasonal fluctuations have approximately constant absolute size. A multiplicative representation is:

The multiplicative form is more appropriate when seasonal variation grows with the level of demand. In practice, decomposition is an analytical approximation, not proof that the underlying business process literally contains four independent components.

Level

Level is the underlying average magnitude of demand. A stable series fluctuates around a relatively constant level. A naive forecast, in which the next period is forecast as the latest observation, can be a strong benchmark for a stable, rapidly changing series.

Trend

Trend is a persistent upward or downward movement. It may result from market growth, product adoption, network expansion or structural decline. A trend should not be extrapolated indefinitely: saturation, competition and capacity constraints can invalidate a linear trajectory.

Seasonality

Seasonality is a recurring pattern with a known calendar or operational period, such as weekly grocery demand, annual holiday demand or monthly heating demand. Seasonality must be distinguished from a one-time event. A promotion repeated every December may appear seasonal, but a non-repeated campaign is an intervention requiring separate treatment.

Cycles and events

Cycles are longer, often irregular fluctuations associated with macroeconomic or industry conditions. Promotions, price changes, strikes, weather shocks, product launches and supply disruptions are interventions or external events. They should be recorded as explanatory variables rather than silently treated as ordinary random variation.

Intermittency

Intermittent demand contains many zero periods and occasional positive demands. Spare parts, maintenance items and slow-moving industrial components commonly display this pattern. Standard percentage errors and ordinary seasonal methods can perform poorly because zero observations make some metrics undefined and because the timing and size of non-zero demand are separate forecasting problems.

Stationarity and structural change

A stationary series has stable statistical properties, such as mean and variance, over the modelling period. Many operational series are only approximately stationary. A sudden change in customers, channel policy, product design or market conditions creates a structural break. Historical data before the break may be less relevant than recent observations, even if the older sample is larger.

3. The Forecasting Process

A professional process normally follows these stages:

1. Define the decision. Identify whether the forecast supports purchasing, production, deployment, staffing or financial planning.
2. Specify the forecast object. Define item, location, horizon, aggregation level, unit and cut-off date.
3. Clean and reconcile data. Identify stockouts, returns, cancellations, duplicate orders, calendar changes, substitutions and abnormal transactions.
4. Explore the series. Plot observations, inspect seasonal profiles, calculate zero proportions and compare demand with inventory availability and external drivers.
5. Establish benchmarks. Use naive, seasonal-naive or simple moving-average forecasts before evaluating complex models.
6. Select candidate methods. Match method assumptions to pattern, horizon, data frequency and managerial use.
7. Validate out of sample. Use rolling-origin or time-ordered validation; random shuffling can leak future information into training data.
8. Generate and review the forecast. Separate statistical output from documented overrides.
9. Measure error and bias. Assess performance by item, segment, location and horizon.
10. Translate uncertainty into decisions. Connect the forecast distribution to inventory, capacity and service policies.
11. Monitor and improve. Re-estimate parameters, investigate bias and retire methods that no longer perform adequately.

Forecasting should be performed at the level at which decisions are made, but aggregation can improve statistical stability. A useful hierarchy may include product, product family, region and total network. Forecasts at different levels should be reconciled so that detailed forecasts add consistently to aggregate forecasts. Bottom-up, top-down and middle-out reconciliation each have advantages; the choice depends on data quality and decision ownership.

4. Core Time-Series Methods

4.1 Naive and seasonal-naive benchmarks

The naive forecast is:

It assumes the latest observed demand is the best estimate of the next period. For seasonal demand with period m, the seasonal-naive forecast is:

where h is the forecast horizon. These methods require minimal data and are difficult to beat consistently when demand changes rapidly. They are essential benchmarks, not unsophisticated methods to be dismissed.

4.2 Moving average

A simple moving average of the most recent k observations is:

All observations receive equal weight. A larger k smooths random variation but responds more slowly to a level change. A smaller k responds faster but is more volatile. The method assumes that the recent window is representative and does not explicitly model trend or seasonality.

Example. Demand over the last three weeks is 96, 108 and 102 units. The three-week moving-average forecast is:

4.3 Weighted moving average

A weighted moving average is:

Weights reflect recency or managerial judgement. The unit of the forecast remains units because the weights are dimensionless. The weights should be chosen through validation rather than merely assumed to improve accuracy.

4.4 Simple exponential smoothing

For a series with no systematic trend or seasonality, simple exponential smoothing is:

Here α is the smoothing constant, Y_t is actual demand and F_t is the forecast for period t. Both Y_t and F_t have the same unit as demand. A high α gives greater weight to recent demand and reacts quickly; a low value provides stronger smoothing. The method assumes a locally constant level and is not designed to represent persistent trend or seasonality.

The one-step forecast error can be defined as:

Positive error means actual demand exceeded the forecast. This sign convention must be stated because some organisations define error as forecast minus actual.

4.5 Trend-adjusted smoothing

When trend is present, Holt’s method estimates a level and trend. One common form is:

Here L_t is the estimated level, B_t the estimated trend, α and β are smoothing constants between 0 and 1, and h is the number of periods ahead. The model assumes that a local trend is meaningful over the forecast horizon. Damped-trend versions reduce long-range extrapolation by gradually weakening B_t, which is often more realistic in supply-chain planning.

4.6 Seasonal methods

A seasonal index measures relative demand for a season compared with an average season. Under a multiplicative model, a deseasonalised value can be expressed as:

where S_t is the seasonal index and D_t is deseasonalised demand. For a complete seasonal cycle, indices are commonly normalised so their average is 1, or their sum is m seasons. A forecast is reseasonalised by multiplying the forecast level or trend by the relevant future index. Additive seasonal models instead add a seasonal effect measured in demand units. The choice should be based on whether seasonal amplitude is approximately constant or proportional to the level.

5. Causal and Regression Forecasting

Time-series methods infer future demand from its past. Causal methods use variables that plausibly influence demand, such as price, promotion, temperature, income, service availability, competitor activity or construction starts. A simple linear model is:

where y_t is demand, x_jt are explanatory variables, β_j are coefficients, β₀ is the intercept and ε_t is unexplained error. The coefficient β_j represents the expected change in demand for a one-unit increase in x_j, holding other included variables constant.

A causal model is useful only when the drivers are available or can themselves be forecast at the time the demand forecast is produced. It is also necessary to consider lagged effects, reverse causality and omitted variables. For example, observed sales may rise with promotions, but promotions may be scheduled precisely when demand was already expected to rise. A coefficient therefore should not automatically be interpreted as a causal effect.

For a log-log specification, the coefficient can be interpreted approximately as an elasticity:

A one-percent change in x_t is associated approximately with a β₁ percent change in expected y_t, subject to model assumptions. Regression diagnostics should examine residual patterns, autocorrelation, heteroscedasticity, multicollinearity, influential observations and out-of-sample performance. High in-sample R² does not guarantee good forecasts.

6. Intermittent-Demand Forecasting

Intermittent demand requires separate attention to occurrence and size. Croston’s method estimates the size of non-zero demand and the interval between non-zero demands, then forecasts their ratio. Variants such as SBA and TSB modify the original method to reduce bias or respond to changing occurrence probabilities. These methods are useful for slow-moving items but have limitations: original Croston does not naturally provide forecasts for every zero period in a probabilistically coherent way, and parameter selection remains important.

For inventory decisions, a method should be judged not only by point forecasts but also by whether it represents lead-time demand and service risk. A spare-parts planner may prefer a method that gives reliable stockout-risk estimates even if its ordinary point forecast has a slightly larger MAD. Segmentation is valuable: fast-moving seasonal items, lumpy items and highly intermittent items should not be forced into one model family.

7. Forecast Error, Accuracy and Bias

No forecast is perfect. Error statistics should be calculated on a consistent holdout or rolling-origin evaluation set. Important measures include:

For MAPE, all actual values in the denominator must be non-zero, and the result can be distorted by very small actuals. WAPE is often more useful for aggregate volume, but it gives high-volume items more influence. MAE or MAD is preferable when the business wants an interpretable unit measure. RMSE is appropriate when large errors have disproportionate operational consequences.

A tracking signal can be defined as:

where:

RSFE is running sum of forecast errors. A large positive tracking signal indicates cumulative underforecasting under the stated error convention; a large negative value indicates cumulative overforecasting. Control limits such as ±4 or ±5 are managerial conventions, not universal statistical laws. They should be calibrated to the item, horizon and review process.

Worked accuracy example. Suppose actual demand is 100, 120 and 80 units, while forecasts are 90, 130 and 100 units. Errors e_t = Y_t − F_t are 10, −10 and −20 units. Thus:

• ME = (10 − 10 − 20) ÷ 3 = −6.67 units.
• MAD = (10 + 10 + 20) ÷ 3 = 13.33 units.
• MSE = (100 + 100 + 400) ÷ 3 = 200 units².
• RMSE = √200 ≈ 14.14 units.
• WAPE = 100(40) ÷ 300 = 13.33%.

The negative ME indicates cumulative overforecasting in this sample, despite the errors not all having the same magnitude. Accuracy measures should be reported with the horizon, aggregation level, sample period and treatment of zero demand.

Forecast bias is not the same as random error. Random errors fluctuate around zero; bias is persistent directional error. Bias may arise from optimistic sales targets, unrecorded stockouts, systematic promotional overestimation, delayed master-data updates or deliberate service protection. An unbiased forecast is not necessarily accurate: errors can average zero while remaining very large.

8. Forecast Selection and Inventory Decisions

Forecast selection should minimise relevant decision cost, not merely a generic error statistic. If underforecasting causes a critical production stoppage while overforecasting causes modest holding cost, asymmetric loss may justify a deliberately conservative forecast or a separate safety-stock policy. Conversely, for perishable goods, overforecasting may be much more costly.

Forecast uncertainty can be represented through prediction intervals or a distribution of future demand. Safety stock should not be confused with changing the point forecast. The point forecast estimates expected demand; safety stock protects against uncertainty in demand and, often, lead time. A forecast that is systematically biased should be corrected before adding safety stock, because safety stock is not a substitute for process correction.

Forecast value added (FVA) compares a baseline statistical forecast with successive interventions, such as sales overrides or marketing adjustments. If an intervention does not improve out-of-sample performance or decision outcomes, it should be challenged. Overrides may be valuable when they incorporate genuine new information, but they can reduce accuracy when driven by incentives, anchoring or unsupported optimism.

9. Collaborative Forecasting

Collaborative planning, forecasting and replenishment (CPFR) integrates information among suppliers, manufacturers, distributors and customers. A typical cycle includes agreement on roles and data, joint business plans, demand sensing, exception review, forecast sharing, replenishment decisions and post-period evaluation.

The value of collaboration comes from access to information unavailable to one party: promotional calendars, customer launches, inventory positions, production constraints, point-of-sale data and planned price changes. Collaboration does not mean that every participant simply averages forecasts. It requires common definitions, agreed ownership, version control, confidentiality arrangements and rules for resolving disagreement.

Point-of-sale data can reduce the distortion caused by order batching, while electronic data interchange and cloud platforms can shorten information delays. However, shared systems can also spread bad data rapidly. Forecast hierarchies, master-data governance and audit trails remain necessary.

10. Technology and Artificial Intelligence

Modern forecasting platforms combine statistical models, machine learning, optimisation and workflow controls. Gradient-boosted trees, random forests and neural networks can model nonlinear relationships and many external variables. Probabilistic forecasting methods produce quantiles or distributions rather than only a single expected value, supporting differentiated service targets and inventory-risk analysis.

Technology should be evaluated against transparent baselines and operational outcomes. Common risks include data leakage, unstable performance after a regime change, biased training data, excessive feature complexity and poor explainability. A model trained on sales during stockouts may learn availability constraints rather than customer demand. Machine learning therefore complements, rather than eliminates, domain knowledge and governance.

AI use should observe data-protection, confidentiality and cybersecurity requirements. Customer-level data should be minimised, access-controlled and retained only as necessary. Automated overrides or replenishment recommendations should have escalation rules, traceability and human accountability, especially for medical, safety-critical or strategically important items.

11. Illustrative Case: Northstar Industrial Supplies

Background

Northstar distributes replacement pumps and seals to mining customers. The company used a single twelve-month moving average for all items. Buyers reported frequent stockouts of high-criticality seals, while the warehouse held excess slow-moving pumps. Sales forecasts were also based on orders, although several customers placed quarterly orders that did not match monthly consumption.

Problem

A review found three distinct patterns: high-volume seals with weekly seasonality, pumps with a declining trend after a technology change, and emergency repair kits with intermittent demand. The existing forecast was evaluated using only aggregate MAPE, which was unstable for the many zero-demand kit months.

Analysis

The team reconstructed demand from point-of-use issues where available, flagged stockout weeks, and separated planned promotions from ordinary demand. Seasonal-naive and exponential-smoothing benchmarks were compared with Holt-Winters models for seals. A causal model included mine maintenance schedules for pumps. Croston-type and TSB forecasts were tested for repair kits. Evaluation used rolling-origin forecasts over six months, with WAPE for network volume, MAD for unit planning and service-related measures for critical items.

The sales department supplied upcoming shutdown schedules and customer project information. Each override required a reason code and was assessed later through FVA. Forecasts were reconciled from customer-site level to regional and total demand so that planning totals remained consistent.

Approach and solution

Northstar adopted segmented forecasting. Seasonal exponential smoothing was used for seals, a damped trend model for pumps, and an intermittent-demand method for kits. Forecasts were converted into replenishment parameters using lead-time demand and target service levels. Bias dashboards were reviewed monthly, and stockout periods were excluded or modelled as censored demand rather than treated as genuine zero demand.

Outcome

In the illustrative scenario, six months after implementation, critical-seal stockouts declined, excess pump inventory was reduced and planner overrides became less frequent. The improvement was not attributed to one algorithm: it resulted from better demand definitions, segmentation, collaborative information, rolling validation and governance of overrides.

Learning points

1. A sales history is not automatically a demand history.
2. One accuracy measure cannot represent all inventory objectives.
3. Intermittent demand requires specialised evaluation.
4. Collaboration is valuable only when information is timely, defined and accountable.
5. Forecasting and inventory policy must be improved together.

12. Challenges, Limitations and Best Practices

Forecasting is constrained by short histories, changing product assortments, missing data, promotions, substitution, cannibalisation, stockouts and political or economic shocks. Historical accuracy can also be misleading when the forecast process itself changes. A model may appear accurate because planners override it using information unavailable to the benchmark, or inaccurate because an unforeseeable disruption occurred.

Best practice includes:

• maintain a clear demand calendar and event database;
• distinguish lost sales from zero demand;
• use simple benchmarks and rolling-origin validation;
• segment by demand pattern, value, criticality and lifecycle;
• select metrics aligned with cost and service consequences;
• report both accuracy and bias;
• separate point forecasts from safety stock;
• document every manual override and evaluate its value;
• reconcile forecasts across product and geographic hierarchies;
• use prediction intervals where decisions depend on risk;
• monitor structural breaks and re-estimate models;
• protect data, explain automated recommendations and retain human accountability.

Chapter Summary

Demand forecasting provides structured estimates of future requirements for inventory and logistics decisions. Demand must be distinguished from sales and orders, especially when stockouts, batching or promotions distort observed history. Demand patterns include level, trend, seasonality, cycles, interventions and intermittency; additive and multiplicative decompositions help describe these patterns.

Naive forecasts, moving averages and exponential smoothing provide transparent benchmarks. Holt-type models address trend, seasonal methods address recurring variation, causal regression incorporates explanatory drivers, and Croston-type approaches address intermittent demand. Method selection must reflect data characteristics, horizon and decision consequences.

Forecast performance requires both accuracy and bias analysis. ME, MAD, MSE, RMSE, MAPE, WAPE, bias ratios and tracking signals answer different questions and have distinct limitations. A forecast should be evaluated out of sample and translated into inventory decisions through uncertainty, service and cost considerations. Collaborative forecasting can improve visibility when partners share reliable, timely information. AI and machine learning can add value, but only with sound data, transparent benchmarks, governance, security and human accountability.

Key Terms

Demand: Quantity customers wish to obtain during a specified period.

Sales: Quantity actually fulfilled or recorded as sold.

Independent demand: Demand originating from external customers.

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

Level: Underlying average magnitude of a demand series.

Trend: Persistent upward or downward movement in demand.

Seasonality: Recurring variation linked to a calendar or operating cycle.

Intermittent demand: Demand containing many zero periods and occasional positive observations.

Naive forecast: Forecast that uses the latest observation as the next forecast.

Exponential smoothing: Method that weights recent and prior forecasts through a smoothing constant.

Causal forecast: Forecast based on explanatory variables that influence demand.

Forecast error: Difference between actual demand and forecast under a stated sign convention.

MAD: Mean absolute deviation, the average absolute forecast error in demand units.

RMSE: Root mean squared error, which penalises large errors more strongly.

WAPE: Volume-weighted absolute percentage error.

Forecast bias: Persistent tendency to overforecast or underforecast.

Tracking signal: Running sum of errors divided by an error-scale measure, commonly MAD.

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

Forecast value added: Assessment of whether a forecasting intervention improves results over a baseline.

Prediction interval: Range intended to contain a future observation with a stated probability.

Review Questions

Short Answer Questions

1. Why can recorded sales underestimate true customer demand?
2. Distinguish trend from seasonality.
3. State the main assumption of simple exponential smoothing.
4. Why is MAPE problematic when actual demand is zero?
5. What does a positive cumulative forecast error indicate under the convention e_t = Y_t − F_t?

Long Answer Questions

1. Explain the forecasting process from problem definition through monitoring and governance.
2. Compare moving averages, exponential smoothing, Holt’s method and seasonal models, including their assumptions and limitations.
3. Discuss why forecast accuracy should be measured with several metrics rather than one universal statistic.
4. Explain how causal forecasting can improve demand planning and identify the conditions under which regression results may be misleading.
5. Describe how collaborative forecasting and AI can improve supply-chain planning, and discuss their governance risks.

Analytical / Application Questions

1. A warehouse records actual demand of 240, 300 and 180 units and forecasts of 210, 330 and 240 units. Calculate ME, MAD, RMSE and WAPE using e_t = Y_t − F_t, and interpret the direction of bias.
2. A spare-parts item has 70% zero-demand periods and highly variable non-zero quantities. Recommend a forecasting and evaluation approach, explaining why an ordinary MAPE-based comparison is unsuitable.
3. A retailer’s forecast is consistently below demand during promotions but accurate in ordinary weeks. Design a revised process using event variables, collaboration, validation and inventory-policy implications.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes sales from demand?

A. Sales always equal customer demand

B. Sales are fulfilled quantities and may be below demand during stockouts

C. Demand is always measured after returns

D. Sales apply only to dependent items

Correct Answer: B. Sales are fulfilled quantities and may be below demand during stockouts

Question 2

For a stable series without trend or seasonality, which method is an appropriate benchmark?

A. Simple exponential smoothing

B. A bill-of-materials explosion

C. A randomised train-test split only

D. A seasonal index with no seasonal cycle

Correct Answer: A. Simple exponential smoothing

Question 3

If actual demand is 120 and the forecast is 100, what is the error under e_t = Y_t − F_t?

A. −20 units

B. 0 units

C. +20 units

D. +220 units

Correct Answer: C. +20 units

Question 4

Which metric is expressed in the original demand units and averages absolute errors?

A. MAD

B. MAPE

C. WAPE

D. MSE

Correct Answer: A. MAD

Question 5

What is a major limitation of MAPE?

A. It cannot be calculated when actual demand is zero

B. It always penalises large errors more than RMSE

C. It is measured in squared units

D. It measures only forecast bias

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

Question 6

A positive running sum of errors under e_t = Y_t − F_t generally indicates that the firm has been:

A. Overforecasting

B. Underforecasting

C. Using seasonal data

D. Eliminating all uncertainty

Correct Answer: B. Underforecasting

Question 7

Which demand pattern is most associated with Croston-type methods?

A. Smooth, high-volume demand with strong daily seasonality

B. Intermittent demand with many zero periods

C. Demand with a known deterministic bill of materials

D. Demand that has no observations

Correct Answer: B. Intermittent demand with many zero periods

Question 8

In a regression demand model, an explanatory variable must be available at forecast time because:

A. Historical coefficients cannot be estimated

B. Future values of unavailable drivers cannot support an operational forecast

C. Regression models cannot use numerical variables

D. Causal models require zero forecast error

Correct Answer: B. Future values of unavailable drivers cannot support an operational forecast

Question 9

What is the primary purpose of forecast value added analysis?

A. To guarantee that every manual override is retained

B. To compare interventions with a baseline forecast

C. To replace all statistical forecasts with opinions

D. To convert MAD into MAPE

Correct Answer: B. To compare interventions with a baseline forecast

Question 10

Why should stockouts be identified during demand-data preparation?

A. They may make observed sales lower than underlying demand

B. They prove that the forecast was unbiased

C. They eliminate the need for safety stock

D. They convert independent demand into dependent demand

Correct Answer: A. They may make observed sales lower than underlying demand