Demand Forecasting and Merchandise Planning

Introduction

Retail logistics begins with a deceptively difficult question: what will customers buy, where, when, and at what price? Forecasting converts evidence about future demand into an estimate that supports purchasing, allocation, replenishment, capacity, transport, labour, and cash decisions. Merchandise planning then translates that estimate into an economically viable assortment, inventory budget, sales plan, and open-to-buy commitment.

Forecasts are not promises. They are structured estimates subject to uncertainty, bias, market change, stockouts, promotions, weather, competitor actions, and data-quality problems. The professional task is therefore not to eliminate uncertainty, but to quantify it, learn from error, and design decisions that remain resilient when the forecast is wrong.

This chapter examines qualitative, time-series, and causal forecasting; forecast evaluation; seasonality, promotions, and product life cycles; and the connection between assortment, category management, open-to-buy, and supply decisions.

Learning objectives

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

1. Distinguish demand, sales, lost sales, and inventory availability.
2. Select and apply qualitative, time-series, and causal forecasting methods.
3. Calculate and interpret common forecast-error measures.
4. Adjust forecasts for seasonality, promotions, holidays, stockouts, and life-cycle effects.
5. Connect category roles, assortment decisions, merchandise plans, and open-to-buy budgets with supply-chain execution.
6. Critically assess forecasting technology, judgment, governance, and ethical risks.

1. The retail demand problem

Observed sales are not always the same as true demand. If a store sells 80 units because only 80 were available, demand may have been 80, 100, or 150. A useful conceptual relationship is:

Here, t denotes a time period. The relationship assumes that each customer who cannot obtain the product generates a lost sale and that availability is measured at the relevant selling point. In practice, substitution, delayed purchase, and customer abandonment complicate estimation.

Retailers should distinguish:

• Demand: quantity customers would purchase under specified price, promotion, assortment, and availability conditions.
• Sales: transactions actually completed.
• Lost sales: demand not fulfilled because of stockout, late delivery, store closure, or another constraint.
• Sell-through: the proportion of received or available inventory sold during a period.
• Forecast: an estimate of future demand or sales, produced at a stated level of aggregation and horizon.
• Forecast horizon: the future interval covered, such as one day, week, season, or financial year.
• Granularity: the level at which the forecast is made—for example, total chain, region, store, category, SKU, or size-colour combination.

A forecast must be aligned with the decision. A distribution centre may need a weekly category forecast, while a store requires a daily SKU forecast. Excessive granularity can create sparse, noisy data; excessive aggregation can conceal local demand differences.

Forecast hierarchy and reconciliation

Retail demand is hierarchical: total company demand is divided into channels, regions, stores, categories, products, and variants. Forecasts made independently at each level may not add up. Forecast reconciliation adjusts forecasts so that lower-level quantities sum consistently to higher-level totals. A practical approach is to forecast at several levels, compare accuracy and managerial purpose, and reconcile using a transparent rule rather than allowing contradictory numbers into the merchandise plan.

2. Forecasting approaches

2.1 Qualitative forecasting

Qualitative methods are valuable when historical data are absent, unstable, or structurally irrelevant. They are common for new products, major repositioning, fashion collections, disruptive technologies, and unusual events.

Important methods include:

• Expert-panel judgment: specialists independently assess likely demand and then discuss differences.
• Delphi method: experts provide anonymous, iterative estimates, reducing the influence of hierarchy and dominant personalities.
• Sales-force composite: sales personnel estimate customer or territory demand; useful where account knowledge is strong, but vulnerable to sandbagging or optimism.
• Executive opinion: senior managers combine market, financial, and strategic information; fast, but potentially biased.
• Consumer research: surveys, concept tests, conjoint studies, and pilot launches estimate preferences and willingness to buy.
• Analogy method: demand for a new product is inferred from comparable products or markets.

Qualitative judgment should be documented: who contributed, what assumptions were made, what evidence was used, and how uncertainty was represented. A numerical forecast should not acquire false authority merely because it is expressed to several decimal places.

2.2 Time-series forecasting

Time-series models use the historical pattern of a variable. They are appropriate when past behaviour contains information about the future and when the operating environment is sufficiently stable.

A series can contain:

• Level: the underlying average magnitude.
• Trend: sustained upward or downward movement.
• Seasonality: a repeating pattern tied to calendar or trading cycles.
• Cycle: longer, less regular fluctuations associated with economic or industry conditions.
• Irregular variation: random or exceptional movement.

#### Naïve and moving-average methods

The naïve forecast uses the most recent observation:

where F_t+1 is the forecast for the next period and A_t is actual demand in period t. It is a useful benchmark, not necessarily a sophisticated final model.

A simple moving average of length k is:

It assumes that the recent k observations are equally relevant and that no strong trend or seasonal adjustment is required. Increasing k smooths noise but makes the forecast less responsive to change.

#### Weighted moving average

A weighted moving average gives different importance to observations:

The weights w_i are non-negative and sum to one. The method is suitable when recent observations should matter more, provided the weights are selected using evidence rather than convenience.

#### Exponential smoothing

Single exponential smoothing is:

α is the smoothing constant. A high value reacts quickly to recent changes; a low value produces greater smoothing. This method is appropriate for a series with a relatively stable level and no systematic trend or seasonality. Trend and seasonal variants, such as Holt and Holt–Winters methods, are required when those features are present.

#### Trend and seasonal indices

A simple trend regression may be written as:

where Y_t is demand, a is the intercept, b is average change per time unit, t is time, and ε_t is random error. A linear trend should not be extrapolated indefinitely; retail markets often saturate or change direction.

A multiplicative seasonal model expresses demand as:

where T_t is the trend-cycle component, S_t is the seasonal index, and ε_t is the irregular component. Seasonal indices are commonly normalized so their average over a complete seasonal cycle is 1.0. Multiplicative seasonality is suitable when seasonal amplitude grows with the level; an additive model is preferable when seasonal effects have approximately constant unit magnitude.

2.3 Causal forecasting

Causal models relate demand to explanatory variables such as price, display, advertising, income, temperature, holidays, competitor activity, and store traffic. A multiple regression form is:

Y_t is demand; X_jt are explanatory variables; β_j are estimated effects; β₀ is the intercept; and ε_t captures unexplained variation. The model assumes a suitably specified relationship, reliable measurement, and errors that do not invalidate estimation. Correlation is not proof of causation: advertising may rise precisely when managers expect weak demand, creating confounding.

For price decisions, a common measure is price elasticity of demand:

The estimate depends on category, time horizon, substitutes, promotion mechanics, and whether the calculation uses regular or promoted prices. Elasticity should therefore be interpreted as a decision-specific estimate, not a universal product property.

3. Forecasting promotions, events, and new products

A historical forecast is often contaminated by promotions. A promotion may increase baseline demand, shift purchases forward, generate post-promotion dips, or cause substitution from another SKU. The analyst should separate:

1. Baseline demand: expected demand without the event.
2. Incremental lift: additional demand caused by the event.
3. Cannibalization: lost demand from related products.
4. Pull-forward: purchases that would otherwise occur later.
5. Post-event effect: demand changes after the promotion ends.

A simple decomposition is:

The terms must be defined consistently; cannibalization may be measured in units or contribution margin, and timing effects may extend across several periods.

For new products, planners may combine analogous-product data, consumer research, pilot-store results, market size, distribution assumptions, and scenario analysis. Rather than one precise number, a launch plan should include conservative, expected, and aggressive cases, with explicit triggers for reorder, markdown, or discontinuation.

4. Forecast accuracy, error, and bias

For period t, forecast error is commonly defined as:

Positive error means actual demand exceeded the forecast; negative error means the forecast was higher than actual demand. Some organizations reverse the sign, so the definition must always be stated.

Key measures

These measures assume comparable periods and a clearly defined forecast origin. MAPE can distort evaluation for low-volume items and cannot be calculated when A_t = 0. MAE and RMSE are preferable for unit-based operational decisions. WAPE is useful for portfolios but can conceal poor performance on small, strategically important SKUs.

Worked example. Actual weekly demand is 100, 120, and 80 units; forecasts are 90, 130, and 100 units. Errors are 10, −10, and −20. Thus:

• ME = (10 − 10 − 20) ÷ 3 = −6.67 units, indicating overforecasting on average.
• MAE = (10 + 10 + 20) ÷ 3 = 13.33 units.
• MSE = (100 + 100 + 400) ÷ 3 = 200 units².
• RMSE = √200 ≈ 14.14 units.
• MAPE = (100 ÷ 3) × (0.10 + 0.0833 + 0.25) ≈ 14.44%.

Accuracy must be evaluated at the horizon and aggregation level of the decision. A forecast can have acceptable chain-level error but unacceptable store-SKU error. Bias is especially important because persistent underforecasting causes stockouts, while persistent overforecasting causes excess inventory and markdowns.

5. Merchandise planning

Merchandise planning is the commercial process of deciding what to sell, how much, when, where, at what price, and with what financial exposure. It links customer demand with supplier capacity and working capital.

Assortment planning

An assortment defines the breadth and depth of merchandise:

• Breadth: number of categories or product lines.
• Depth: number of options within a line, such as brands, sizes, colours, or price points.
• SKU productivity: sales, margin, or contribution generated per SKU, space unit, or inventory unit.

A good assortment balances customer choice against complexity. Long-tail products may improve differentiation but create slow-moving inventory, forecasting sparsity, handling cost, and supplier minimum-order exposure. Assortment decisions should consider demand variability, substitution, store capacity, channel role, lead time, margin, and strategic importance—not only historical sales.

Category roles and category plans

Category management treats a category as a strategic business unit. Common roles include destination, routine or core, seasonal, and convenience or impulse. The role influences service level, price architecture, space, promotion, and replenishment policy. A destination category may justify high availability and broad depth; a seasonal category requires time-phased commitments and rapid exit planning.

A category plan typically specifies objectives, target customers, assortment, pricing, promotion, space, forecast, inventory policy, supplier actions, and performance measures. It should connect customer-facing decisions to operational constraints.

Merchandise budget plan

A merchandise plan translates sales and margin objectives into period-by-period financial and unit targets. A basic retail inventory relationship is:

EOM means end-of-month and BOM means beginning-of-month. All terms must be measured on a consistent basis—units or value, and either retail or cost valuation. Shrinkage includes loss from theft, damage, and process error.

A common sales-based planned purchase relationship is:

This identity is useful for a merchandise financial plan, but it is not a replenishment order formula. Purchase orders must additionally consider lead time, order constraints, inventory already in transit, safety stock, availability targets, and supplier reliability.

Open-to-buy

Open-to-buy (OTB) is the remaining purchasing capacity for a defined period. In a simple retail-value formulation:

The exact organizational formula may include receipts already ordered, cancellations, transfers, and inventory adjustments. The essential control principle is that an order should consume available financial capacity unless it is explicitly approved as a plan change. OTB protects cash and limits uncontrolled buying, but a rigid OTB can also cause lost sales when demand accelerates or lead times lengthen. It should therefore be reviewed with service and supply-risk information.

6. Connecting merchandise plans with supply decisions

The connection from forecast to supply is not automatic. A planner must translate a demand estimate into timing, location, quantity, and risk decisions.

1. Forecast: estimate demand by channel, store or fulfilment node, SKU, and time bucket.
2. Assortment decision: determine which items are active and the intended depth.
3. Financial plan: set sales, margin, markdown, inventory, and OTB limits.
4. Supply translation: account for lead time, minimum order quantity, case pack, supplier capacity, order cycle, and transport.
5. Allocation and replenishment: position inventory by expected demand and service objective.
6. Exception management: respond to forecast bias, stockouts, late receipts, promotion changes, and supplier failures.
7. Learning loop: compare actual demand and profitability with plan, then revise parameters.

For a seasonal product, ordering too early increases holding and markdown risk; ordering too late increases stockout risk. The correct decision depends on the cost of each outcome, lead-time uncertainty, residual value after the season, and the possibility of transshipment or reallocation.

Safety stock and uncertainty

When demand and lead time are uncertain, inventory buffers protect service. Under a simplified independent-normal assumption, safety stock can be represented as:

where SS is safety stock in units, z is the standard-normal factor for the desired cycle-service probability, and σ_LT is the standard deviation of demand during lead time. If demand per period has standard deviation σ_d over L independent periods and lead time is fixed, σ_LT = σ_d√L. If both demand and lead time vary, the variance must include lead-time uncertainty and any covariance; the simple equation should not be used uncritically.

This illustrates an important principle: a forecast is a mean estimate, while inventory policy must also address dispersion, bias, lead time, and target service.

7. Product life cycles and retail decisions

A product life cycle commonly includes introduction, growth, maturity, and decline. The stages are analytical guides, not universal laws. A new product has little direct history, growth may attract competitors, maturity may produce stable but promotion-sensitive demand, and decline may require assortment rationalization.

Forecasting and supply implications differ by stage:

A life-cycle forecast should be revised when evidence changes. Treating a temporary promotion spike as permanent growth is a common and costly error.

8. Technology, AI, and governance

Modern retail forecasting uses point-of-sale data, loyalty data, weather, web searches, clickstream events, inventory visibility, machine learning, and cloud-based planning platforms. Machine-learning models can capture nonlinear relationships and interactions, while probabilistic forecasting can produce a distribution of possible demand rather than a single point.

Technology does not remove managerial responsibility. Key controls include:

• versioned data and model pipelines;
• clear separation of training, validation, and test periods;
• prevention of data leakage, such as using future promotional information unavailable at forecast time;
• monitoring for drift, bias, and deterioration;
• human review of exceptional events;
• explainable overrides with reason codes;
• access controls and audit trails.

Ethically, retailers must govern the use of customer data, comply with privacy requirements, minimize unnecessary personal data, and test for discriminatory effects. Dynamic pricing and targeted promotions may disadvantage vulnerable customers or create perceived unfairness. Forecasting systems can also reproduce historical inequity—for example, systematically under-serving neighbourhoods with incomplete data. Governance should therefore include privacy, fairness, accountability, security, and a process for challenging automated recommendations.

9. Challenges and limitations

Forecasting and merchandise planning face recurring limitations:

• Intermittent demand: many periods contain zero sales, making ordinary percentage measures unstable.
• Censored demand: stockouts make sales appear lower than demand.
• Structural breaks: competitors, regulation, pandemics, formats, or technology can invalidate history.
• Data hierarchy problems: SKU forecasts may not reconcile with category plans.
• Promotion endogeneity: promotions are not randomly assigned; observed lift may reflect managerial expectations.
• Organizational bias: buyers may inflate forecasts to secure inventory, while finance may suppress them to protect budgets.
• Metric conflict: high availability, margin, inventory turns, and sustainability may not be maximized simultaneously.
• Sustainability: overproduction and markdown disposal create environmental costs not captured in a narrow sales forecast.

Best practice is to combine statistical baselines with structured judgment, scenario planning, causal evidence, exception-based review, and post-season learning. Overrides should be fewer, material, documented, and evaluated after the event.

Applied case: NorthStar Homeware's seasonal launch

This is an illustrative case.

Background

NorthStar Homeware plans to launch a premium insulated bottle in 120 stores and online before the summer travel season. A comparable bottle sold an average of 1,000 units per week across the chain last year, but it was heavily promoted and experienced three stockout weeks. The new bottle has a higher price, improved insulation, and four colours. Supplier lead time is six weeks, with a case pack of 12.

Problem

The buying team proposes 12,000 units based on last year's total sales. The supply team argues that demand is uncertain and that the supplier cannot guarantee rapid replenishment. Finance warns that unsold colours will require markdowns after the season.

Analysis

First, last year's sales are not a clean baseline because promotion and stockouts distorted observed demand. The team estimates lost sales from store-level availability records and removes the exceptional launch promotion. It develops three scenarios for the ten-week season:

The planners then allocate initial inventory by store potential, online demand, and colour preference rather than dividing units equally. They reserve a portion of the purchase commitment for the expected and aggressive scenarios, subject to the six-week lead time. Because the case pack is 12, store allocations are rounded to pack quantities, with a small central reserve for reallocation.

Approach and solution

NorthStar uses a baseline forecast adjusted for price, promotion, and the summer seasonal index. It sets an initial commitment near the expected scenario, monitors weekly sell-through and availability, and establishes reorder triggers. The merchandising plan includes a planned markdown budget and a colour-level exit date. OTB approval is required for any additional order, and online search and conversion data are used as leading indicators rather than treated as guaranteed sales.

The team also creates a post-launch measurement plan: compare forecast with unconstrained demand estimates, track MAE and bias by channel and colour, record stockout hours, and assess gross margin after markdown. If one colour sells slowly, inventory is transferred before a new order is placed.

Outcome

In the first four weeks, total demand is close to the expected scenario, but the green colour underperforms while black exceeds its forecast. NorthStar transfers units from low-volume stores and changes the next allocation. It avoids a broad emergency order, preserves OTB for the highest-converting channel, and reduces end-of-season residual stock.

Learning points

The case demonstrates that historical sales may be censored and promotion-contaminated; scenarios are more credible than false precision; assortment decisions must include colour and location; and a merchandise plan becomes operationally useful only when connected to lead time, pack size, allocation, OTB, and feedback measures.

Chapter Summary

Demand forecasting estimates future customer requirements, but observed sales may understate demand when inventory is unavailable. Qualitative methods are valuable for novel or unstable situations; time-series methods exploit level, trend, and seasonality; and causal models incorporate price, promotion, weather, and other drivers. Forecast choice must match the decision, data, horizon, and stability of the environment.

Forecast error should be defined consistently and assessed with measures such as ME, MAE, RMSE, MAPE, and WAPE, recognizing their different limitations. Promotions require separation of baseline, incremental lift, cannibalization, and timing effects. Product life-cycle stage affects the appropriate evidence, inventory commitment, and exit policy.

Merchandise planning connects assortment breadth and depth, category roles, sales and inventory plans, markdowns, and open-to-buy. Supply decisions must additionally reflect lead times, pack sizes, capacity, service targets, uncertainty, and inventory already in the pipeline. AI can improve scale and responsiveness, but it requires data governance, drift monitoring, explainability, privacy protection, and human accountability. Effective retail planning is a learning system: forecast, execute, measure, diagnose, and revise.

Key Terms

Demand: Quantity customers would purchase under specified conditions.

Forecast: An estimate of future demand or sales for a stated horizon.

Lost sales: Demand not fulfilled because of stockout or another constraint.

Forecast hierarchy: Levels from total business to channel, store, category, SKU, and variant.

Qualitative forecasting: Forecasting based primarily on expert, market, or consumer judgment.

Time-series forecasting: Forecasting that uses historical patterns over time.

Causal forecasting: Forecasting that relates demand to explanatory variables.

Seasonality: A repeating demand pattern linked to calendar or trading cycles.

Baseline demand: Expected demand without a specified promotion or event.

Incremental lift: Additional demand caused by an event or intervention.

Cannibalization: Demand lost by related products because of a product or promotion.

Forecast bias: Persistent tendency to overforecast or underforecast.

MAE: Mean absolute error, the average absolute forecast error in units.

Assortment: The set of products and variants offered by a retailer.

Breadth: Number of categories or product lines in an assortment.

Depth: Number of options within a product line or category.

Category role: Strategic purpose assigned to a merchandise category.

Open-to-buy: Remaining purchasing capacity within an approved merchandise plan.

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

Forecast reconciliation: Adjustment of forecasts so hierarchical totals are consistent.

Review Questions

Short Answer Questions

1. Why can observed retail sales be lower than true demand?
2. State one suitable use of qualitative forecasting.
3. What is the difference between seasonality and trend?
4. Why can MAPE be problematic for low-volume items?
5. Define open-to-buy.

Long Answer Questions

1. Compare qualitative, time-series, and causal forecasting, including their assumptions and retail applications.
2. Explain how promotions contaminate historical data and how a planner can estimate baseline and incremental demand.
3. Discuss how assortment breadth and depth affect inventory, service, and supply-chain complexity.
4. Explain the relationship among a merchandise budget plan, planned receipts, and open-to-buy.
5. Evaluate the benefits and governance requirements of AI-based retail forecasting.

Analytical / Application Questions

1. A category has actual demand of 240, 300, and 180 units and forecasts of 200, 330, and 210 units. Calculate ME, MAE, and RMSE, and interpret the bias.
2. A retailer has a six-week supplier lead time, uncertain seasonal demand, and four colour variants. Design a launch forecasting and allocation process that limits markdown and stockout risk.
3. A promotion doubles sales of one SKU but reduces sales of a similar SKU. Explain how you would estimate total category lift, cannibalization, and the implications for the next purchase decision.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes demand from observed sales?

A. Demand is always equal to inventory available

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

C. Demand excludes purchases made during promotions

D. Observed sales include all customers who abandoned a purchase

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

Question 2

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

A. 0.10

B. 0.25

C. 0.333 approximately

D. 1.00

Correct Answer: C. 0.333 approximately

Question 3

In single exponential smoothing, a higher α generally makes the forecast:

A. More responsive to recent observations

B. Independent of recent observations

C. More affected by distant observations only

D. Valid when seasonality is automatically removed

Correct Answer: A. More responsive to recent observations

Question 4

Using the chapter’s error convention, actual demand is 100 and forecast demand is 90. What is the error?

A. −10 units

B. 0 units

C. 10 units

D. 190 units

Correct Answer: C. 10 units

Question 5

Which measure is undefined when actual demand is zero?

A. MAE

B. RMSE

C. ME

D. MAPE

Correct Answer: D. MAPE

Question 6

Which item is part of promotion analysis?

A. Incremental lift

B. Supplier legal ownership only

C. Warehouse floor colour

D. Employee tenure

Correct Answer: A. Incremental lift

Question 7

What is the primary purpose of open-to-buy?

A. To calculate customer elasticity exactly

B. To control remaining purchasing capacity

C. To replace all demand forecasts

D. To eliminate lead-time uncertainty

Correct Answer: B. To control remaining purchasing capacity

Question 8

Which category is most likely to justify broad assortment and high availability?

A. Destination category

B. Declining clearance category

C. Unproven one-off item only

D. Category with no customer relevance

Correct Answer: A. Destination category

Question 9

Under the simplified safety-stock equation, z represents:

A. The desired service-level standard-normal factor

B. The number of stores

C. The supplier lead time in weeks

D. The forecast error sign

Correct Answer: A. The desired service-level standard-normal factor

Question 10

Why should an AI forecasting system be monitored after deployment?

A. Models cannot use historical data

B. Forecast relationships and data quality may drift over time

C. Monitoring guarantees zero forecast error

D. Human review is always mathematically invalid

Correct Answer: B. Forecast relationships and data quality may drift over time