Specialized Forecasting and Demand Shaping

Introduction

Standard time-series forecasting works best when demand is sufficiently frequent, stable and historically observable. Supply-chain decisions, however, regularly involve situations in which these conditions do not hold. A product may be new and have no sales history; its commercial life may be only a few weeks; a spare part may be demanded only occasionally; or a promotion may change demand for one item while reducing demand for another. In such settings, forecasting is not merely the extrapolation of past sales. It is a structured process of combining market knowledge, product analogies, causal information and managerial judgement.

This chapter examines specialized forecasting methods for new products, short-life-cycle items, spare parts and intermittent demand. It also considers demand shaping: the deliberate use of price, promotion, assortment, availability and substitution policies to influence the timing, quantity and composition of demand. The central managerial issue is to distinguish demand that is observed from demand that could have occurred under different commercial conditions.

Learning Objectives

After completing this chapter, the learner should be able to:

1. Explain why conventional time-series models are inadequate for specialized demand situations.
2. Apply diffusion, product-lifecycle and analogous-product methods to new-product forecasting.
3. Forecast intermittent demand using Croston-type methods and understand their limitations.
4. Estimate promotion uplift, cannibalization and substitution effects.
5. Evaluate pricing, assortment and availability decisions as demand-shaping instruments.
6. Design a practical forecasting process that combines data, judgement, experimentation and governance.

1. The Nature of Specialized Demand

Specialized demand has one or more of the following characteristics:

Four distinctions are particularly important. First, sales are not always equal to demand: stockouts, rationing and distribution constraints may censor observed demand. Second, baseline demand is the demand expected without a particular intervention. Third, incremental demand is the additional demand generated by that intervention. Fourth, demand may be lost, shifted in time, or transferred from another item rather than genuinely created.

A useful decomposition is:

This is a conceptual decomposition rather than a universal estimation equation. Its usefulness depends on defining the intervention, identifying comparable observations and accounting for stock availability.

2. Forecasting New Products

2.1 The information hierarchy

A new-product forecast should not begin with a single numerical estimate. It should begin with an information hierarchy:

1. Comparable internal products: products with similar customer segments, price points, channels and use occasions.
2. External analogues: products from competitors or related industries with similar adoption conditions.
3. Market and customer evidence: surveys, concept tests, pre-orders, retailer commitments and search behaviour.
4. Commercial assumptions: distribution coverage, launch timing, price, promotional support and capacity.
5. Structural models: diffusion or lifecycle models used to translate assumptions into time-phased demand.

The forecast should normally be expressed as a range or scenario set. A point estimate can conceal uncertainty about adoption speed, market size and the effect of availability.

2.2 Analogous-product forecasting

An analogous-product method transfers information from an existing product to the new product. The analogy should be based on causal similarity, not superficial similarity. Relevant dimensions include:

• customer and use case;
• price and perceived value;
• channel and geographic coverage;
• product complexity and learning requirements;
• competitive intensity;
• seasonality and promotional calendar;
• launch support and distribution availability.

A simple proportional approach is:

where F_new,t is the forecast for the new product in period t, F_analog,t is the corresponding analogue forecast or actual demand, and r is an adjustment factor. The ratio r is dimensionless and may reflect differences in market size, distribution, price or expected acceptance.

For example, if a comparable product sold 4,000 units in its third month and the new product is expected to achieve 75% of its analogue’s demand, the forecast is 3,000 units. The calculation is easy; the difficult task is justifying 75%. That assumption should be supported by evidence and tested through sensitivity analysis.

Analogy is most credible when the analogue’s launch conditions are reconstructed. If the analogue had unusually broad distribution or a major advertising campaign, copying its sales curve without adjustment will overstate the new product’s likely demand.

2.3 Diffusion models

Diffusion models describe how adoption spreads through a population. The Bass model separates innovation effects, caused by external influence such as advertising, from imitation effects, caused by existing adopters and word of mouth.

where S(t) is new adoption during period t, p is the coefficient of innovation, q is the coefficient of imitation, F(t) is cumulative adoption as a proportion of the potential market, m is total market potential and N(t) is cumulative adoption in units. The model assumes a defined potential market, adoption rather than repeat consumption, and parameters that are reasonably stable over the forecast horizon.

An equivalent continuous form is often written using cumulative adoption proportion F(t):

Here f(t) is the rate of adoption as a proportion of market potential. The term 1 − F(t) represents the remaining non-adopters. In practice, parameters are difficult to estimate for genuinely new products, so managers often calibrate them from analogous launches or market research.

Diffusion models are less suitable when purchases are primarily replacements, when the market potential is unknown, when supply restrictions determine sales, or when the product is rapidly discontinued.

2.4 Product-life-cycle forecasting

The product life cycle provides a managerial framework rather than a deterministic law. The usual stages are introduction, growth, maturity and decline. Each stage implies different forecasting evidence and risks:

Lifecycle labels should not be used mechanically. A product may experience a second growth phase after redesign, a temporary decline caused by stockout, or maturity in one channel while remaining in growth in another.

3. Short-Life-Cycle Items

Fashion products, seasonal merchandise, event-related goods and many technology products have a narrow selling window. Their forecast error cannot be corrected through unlimited replenishment. The principal decision is often the initial commitment, followed by allocation, replenishment and markdown decisions.

3.1 Forecasting structure

A short-life-cycle forecast should be time-phased at a sufficiently fine interval. Weekly data may be inadequate for a product whose demand changes sharply over a few days. The forecast should incorporate:

• launch and end dates;
• selling-calendar effects;
• channel and location differences;
• early sales signals;
• inventory availability;
• returns and cancellations;
• markdown and promotion plans.

An initial forecast may be updated using a weighted combination of prior expectation and early evidence:

where F_prior is the pre-launch forecast, F_signal is the forecast implied by early sales or market signals, and w lies between 0 and 1. The weight should decrease as signal quality and sample size improve. This is a managerial updating rule, not a substitute for statistical estimation.

3.2 The newsvendor logic

When only one replenishment opportunity exists, the classical single-period inventory model is useful. Let c_u be the underage cost per unit, c_o the overage cost per unit, and F(Q) the cumulative distribution of demand. The cost-minimizing service fractile is:

where Q* is the order quantity and D is random demand. Equivalently, the target quantity is the quantile of demand corresponding to this probability. The formula assumes a single ordering decision, known unit costs, a defined salvage or shortage valuation, and a meaningful demand distribution.

If the underage cost is 6 monetary units and the overage cost is 4, the critical fractile is 6 ÷ (6 + 4) = 0.60. The firm should select the 60th percentile of the demand distribution, not necessarily the mean. A high underage cost justifies a higher inventory position; a high overage cost justifies a lower one.

4. Spare Parts and Intermittent Demand

4.1 Characteristics

Intermittent demand contains many zero-demand periods separated by non-zero transactions. Spare-parts demand is often intermittent, lumpy and influenced by installed-base failures, maintenance policy and substitution. Conventional moving averages can respond poorly because zeros depress the average while a large demand occurrence creates a temporary spike.

Important measures include:

• Demand frequency: proportion of periods with non-zero demand.
• Average demand size: average quantity conditional on demand occurring.
• Average inter-demand interval: average number of periods between non-zero demands.
• Lumpiness: variability of non-zero demand sizes and intervals.

The forecast horizon and review period must be stated explicitly. A monthly spare-parts forecast is not directly comparable with a daily service-parts forecast.

4.2 Croston’s method

Croston’s method separately estimates demand size and the interval between demands. Let z_t denote non-zero demand size and p_t the interval since the previous non-zero demand. When demand occurs, the estimates are updated:

The forecast per period is:

Here α is a smoothing parameter between 0 and 1, Z_t has units of units per demand occurrence, P_t has units of periods per demand occurrence, and F_t has units of units per period. When demand is zero, the standard method leaves Z_t and P_t unchanged. The method assumes that demand size and demand interval can be estimated separately and that the process is sufficiently stable for exponential smoothing to be informative.

Croston’s original method has known statistical bias under some demand processes, particularly because of the timing convention and because it does not naturally represent parameter uncertainty. Variants such as the SBA and TSB methods modify the update logic. TSB, for example, updates the probability of demand occurrence and the demand size, making it useful when demand may disappear permanently. The chosen method should be validated through rolling-origin tests rather than selected solely by tradition.

4.3 Worked example

Suppose a spare part has an initial estimated demand size of 10 units and an initial inter-demand interval of 4 months. Use α = 0.2. After five months, a demand of 14 units occurs. The observed interval is 5 months.

Updated demand size:

Updated interval:

Forecast per month:

This forecast is not a prediction that exactly 2.57 units will be demanded in the next month. It is an average-rate estimate for inventory and capacity planning. Safety stock must additionally reflect lead time, criticality, demand variability and the consequences of shortage.

5. Promotion Uplift, Cannibalization and Substitution

5.1 Baseline and uplift

Promotion analysis begins with a credible baseline: the demand expected without the promotion. If baseline demand is 1,000 units and promoted demand is 1,300 units, incremental uplift is 300 units and percentage uplift is:

In this example, uplift is (1,300 − 1,000) ÷ 1,000 × 100 = 30%. The baseline must account for seasonality, trend, distribution and competitor activity. Comparing a promotion week with the immediately preceding week can be misleading when demand is seasonal.

5.2 Promotion evaluation

A promotion may increase unit sales but reduce contribution margin. A simplified incremental contribution calculation is:

where P_p and P_b are promoted and baseline prices, C_v is variable cost per unit, Q_p and Q_b are promoted and baseline quantities, and C_promo is the promotion cost. This expression assumes the same variable cost and ignores longer-term effects such as customer acquisition, reference-price changes and post-promotion demand dips.

5.3 Cannibalization

Cannibalization occurs when an intervention increases sales of one item but reduces sales of another item in the same portfolio. If the promoted product gains 400 units while a sister product loses 150 units, portfolio incremental volume is only 250 units before considering category effects.

A simple cannibalization ratio is:

In the example, the ratio is 150 ÷ 400 = 37.5%. This measure should be interpreted cautiously because the loss may also reflect competitor actions, seasonality or stock availability.

5.4 Substitution and stockouts

Substitution is the transfer of demand from an unavailable or unattractive item to another item. It may be customer-driven, retailer-driven or designed by the firm. For planning, an unconstrained demand estimate can be written conceptually as:

Lost sales are not directly observed and must be estimated from inventory records, search activity, customer complaints, substitution patterns or comparable locations. Treating observed sales as demand during a stockout systematically understates future requirements.

6. Demand Shaping Decisions

Forecasting and demand shaping are related but distinct. Forecasting estimates what is likely to happen under specified conditions; shaping changes those conditions. Common instruments include price, promotion, assortment, lead-time promises, allocation, product configuration and communication.

6.1 Pricing

Price elasticity is commonly defined as:

A negative elasticity indicates that quantity generally decreases as price increases. Elasticity estimates are causal only when confounding factors, such as promotions and availability, are controlled. A price test should therefore define treatment and control groups where feasible and monitor competitor prices.

Pricing decisions must consider contribution, not volume alone. A lower price can raise volume while reducing total profit if the margin loss exceeds the incremental contribution from additional units.

6.2 Assortment

Assortment decisions shape demand by determining which alternatives customers can see and buy. Removing a low-volume item may transfer demand to a substitute, but it may also cause customers to leave the category. Assortment evaluation should therefore distinguish:

• direct item demand;
• transferred demand within the category;
• lost category demand;
• operational savings from simplification;
• effects on service level and customer perception.

A portfolio forecast should be reconciled so that item-level changes do not imply impossible category totals.

6.3 Allocation and availability

When supply is scarce, allocation is a demand-shaping decision because it affects observed sales and customer choice. Allocation should consider margin, strategic customers, service commitments, substitution opportunities and fairness. It should not be evaluated only through historical sales, since historical sales may reflect earlier allocation rules.

7. Integrated Forecasting Process

A robust process has seven stages:

1. Define the decision: launch quantity, service stock, promotion volume or assortment change.
2. Define the demand unit: product, location, channel and time period.
3. Diagnose the demand pattern: continuous, seasonal, intermittent, lifecycle or intervention-driven.
4. Construct a baseline: correct for stockouts, promotions and distribution changes.
5. Select methods: analogy, diffusion, lifecycle, Croston-type or causal models.
6. Create scenarios: include conservative, central and upside cases with explicit assumptions.
7. Monitor and learn: compare forecast, unconstrained demand and realized outcomes; revise assumptions.

Forecast accuracy should not be the only evaluation criterion. Bias, service level, working capital, obsolescence, margin and decision quality are also relevant. For intermittent demand, percentage-error measures can be unstable or undefined when actual demand is zero. Scaled errors, absolute errors, service outcomes and inventory costs are often more informative.

8. Technology, AI and Governance

Modern systems can combine point-of-sale data, search activity, weather, web traffic, installed-base records and promotion calendars. Machine-learning models may identify nonlinear relationships and interactions, but they require sufficient representative data. A model trained on normal availability may learn that low sales during stockouts represent low demand.

AI-supported forecasting should therefore include:

• data lineage and version control;
• human review of assumptions and overrides;
• monitoring for drift and bias;
• explainable drivers for important decisions;
• protection of customer and supplier data;
• controls against leakage of future information into training data.

Governance is especially important when forecasts affect allocation among customers, access to critical spare parts or prices offered to different market segments. A technically accurate forecast can still produce unacceptable outcomes if its data or decision rule is discriminatory, opaque or inconsistent with contractual obligations.

9. Illustrative Case: Launching a Connected Refrigeration Sensor

Background

A manufacturer of commercial refrigeration equipment plans to launch a sensor that monitors temperature and sends alerts to facility managers. The product is new, but the firm has historical sales of a simpler monitoring device and a large installed base of refrigeration units.

Problem

Management must determine the first six months of demand, initial component purchases and the service-parts requirement. The product has uncertain adoption, a promotional launch and possible cannibalization of the simpler device.

Analysis

The planning team identifies the simpler device as an analogue but finds that the new sensor has a higher price and requires software integration. It therefore reduces the analogue launch curve by 25% for the central scenario. Market interviews suggest a potential market of 20,000 eligible sites, but adoption is uncertain. A diffusion model is used for scenario structure rather than as a precise statistical forecast.

The launch promotion produces a forecast of 1,300 units in the first month against a baseline of 1,000 units. However, the simpler device is expected to lose 150 units because some customers will choose the new sensor. The portfolio’s incremental volume is therefore 250 units before considering category expansion. The team also identifies a battery module as a spare part. Historical demand for a similar module is intermittent, so a Croston-type forecast is used, with a separate criticality-based safety-stock policy.

Approach and solution

The firm establishes three scenarios:

The launch is staged. Initial production covers the central forecast, while component contracts include an option for additional volume. Weekly data are reviewed, but sales are adjusted for stockouts and channel inventory. Promotion performance is evaluated on contribution margin, not units alone. The simpler product remains in the assortment because its lower price serves a distinct segment.

Outcome

The staged approach reduces obsolescence exposure while preserving the ability to respond to strong adoption. The organization also avoids interpreting promotional volume as permanent baseline demand. The spare-part forecast is managed separately from the finished-product launch forecast because failure demand follows a different process.

Learning points

1. A product can require several forecasts: market adoption, channel sell-in, customer sell-through and spare parts.
2. Analogies are useful only when differences in price, distribution and use case are made explicit.
3. Promotion uplift must be separated from cannibalization and stockout effects.
4. Scenario-based procurement is often more robust than committing to one precise launch number.

Challenges and Limitations

Specialized forecasting is vulnerable to structural breaks, biased market research, selective analogue choice and managerial optimism. New-product forecasts may confuse interest with purchase intent. Diffusion models may imply a stable potential market when competitors or technologies change it. Croston-type methods may perform poorly when demand is lumpy, declining or affected by planned maintenance. Promotion estimates may fail when treatment and control markets are not comparable.

The response is not to eliminate judgement but to make judgement explicit. Every major assumption should have an owner, evidence source, review date and sensitivity range. Forecasts should also distinguish uncertainty that can be reduced through information from variability that remains irreducible.

Best-Practice Checklist

• Separate observed sales from unconstrained demand.
• Classify the demand process before selecting a method.
• Use analogues based on causal similarity.
• Treat diffusion and lifecycle models as structured assumptions, not certainties.
• Use intermittent-demand methods for intermittent data, and validate variants empirically.
• Estimate promotion uplift against a credible baseline.
• Measure cannibalization at portfolio and category levels.
• Include price, availability, assortment and substitution in scenario design.
• Evaluate forecasts using financial and service outcomes as well as statistical error.
• Document overrides, data limitations and ethical implications.

Chapter Summary

Specialized forecasting addresses demand situations in which historical sales are limited, intermittent, censored or altered by managerial intervention. New-product forecasts combine analogies, market evidence, diffusion concepts and lifecycle reasoning. Short-life-cycle products require time-phased forecasts and explicit treatment of overage and underage costs. Spare-parts demand is often intermittent; Croston-type methods estimate non-zero demand size and the interval between demand occurrences separately.

Demand shaping includes promotion, pricing, assortment, allocation and availability decisions. Effective analysis distinguishes baseline demand, incremental uplift, cannibalization and substitution. Forecasting systems should use scenarios, causal evidence, continuous monitoring and governance rather than relying on a single unexamined point estimate. The strongest professional practice integrates statistical methods with commercial knowledge while clearly documenting assumptions and uncertainty.

Key Terms

Baseline demand: Demand expected without a specified intervention.

Cannibalization: Reduction in demand for one product caused by increased demand for another related product.

Croston’s method: An intermittent-demand method that separately smooths demand size and demand interval.

Demand shaping: Deliberate managerial action intended to influence the timing, quantity or composition of demand.

Diffusion model: A model describing how adoption spreads through a potential market.

Intermittent demand: Demand containing multiple zero-demand periods.

Lifecycle: The sequence of introduction, growth, maturity and decline associated with a product or market.

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

Newsvendor model: A single-period inventory model balancing underage and overage costs.

Potential market: The maximum population or volume considered capable of adopting a product under defined conditions.

Promotion uplift: Incremental demand associated with a promotion relative to a credible baseline.

Reference product: An existing product used as an analogue for forecasting a new product.

Service part: A component held to support maintenance or repair of an installed product.

Substitution: Transfer of demand from one item, channel or supplier to another.

Underage cost: Economic cost of having too little inventory.

Overage cost: Economic cost of having too much inventory.

Unconstrained demand: Demand that would have occurred if inventory and capacity constraints had not limited fulfilment.

Review Questions

Short Answer Questions

1. Why can observed sales differ from true demand?
2. What are the two quantities separately estimated in Croston’s method?
3. Define promotion uplift and explain why a baseline is necessary.
4. What is cannibalization?
5. Identify two limitations of diffusion models.

Long Answer Questions

1. Explain how analogous-product forecasting can be used for a new product. Discuss the criteria for selecting a credible analogue.
2. Describe the assumptions, calculations and limitations of Croston’s method.
3. Explain how the newsvendor logic supports inventory decisions for short-life-cycle products.
4. Discuss how price, promotion and assortment decisions shape demand and affect portfolio forecasts.
5. Design an integrated forecasting process for a new product with uncertain adoption and limited supply.

Analytical / Application Questions

1. A promoted product sells 1,500 units compared with a baseline of 1,100 units. A related product loses 120 units. Calculate the promotion uplift percentage and the cannibalization ratio, then explain what additional evidence is needed before judging the promotion successful.
2. A spare part has an estimated demand size of 8 units and an inter-demand interval of 3 months. With α = 0.25, the next non-zero demand is 12 units after 5 months. Calculate the updated demand size, updated interval and Croston forecast per month.
3. A retailer must order a seasonal item once. The underage cost is 9 monetary units per unit and the overage cost is 6 monetary units per unit. Calculate the required service fractile and explain how uncertainty in the demand distribution should influence the decision.

Multiple Choice Questions with Answers

Question 1

Which situation is most appropriate for an intermittent-demand method such as Croston’s method?

A. A product with strong daily sales and stable seasonality

B. A spare part with many zero-demand periods

C. A newly launched product with no analogue

D. A product whose demand is entirely determined by a promotion

Correct Answer: B. A spare part with many zero-demand periods

Question 2

In the Bass diffusion model, the coefficient of imitation primarily represents the influence of:

A. Existing adopters and word of mouth

B. Unit variable cost

C. Inventory holding cost

D. Supplier lead time

Correct Answer: A. Existing adopters and word of mouth

Question 3

A product sells 1,300 units during a promotion, compared with baseline demand of 1,000 units. What is the percentage uplift?

A. 13%

B. 23%

C. 30%

D. 43%

Correct Answer: C. 30%

Question 4

If a promoted item gains 400 units and a related item loses 100 units, the cannibalization ratio is:

A. 10%

B. 25%

C. 40%

D. 75%

Correct Answer: B. 25%

Question 5

In Croston’s method, when demand is zero in a period, the standard method generally:

A. Sets the forecast to zero permanently

B. Updates demand size using zero demand

C. Leaves the demand-size and interval estimates unchanged

D. Doubles the interval estimate

Correct Answer: C. Leaves the demand-size and interval estimates unchanged

Question 6

If underage cost is 8 and overage cost is 2, the newsvendor critical fractile is:

A. 0.20

B. 0.40

C. 0.67

D. 0.80

Correct Answer: D. 0.80

Question 7

Which statement best describes cannibalization?

A. Demand transferred from a competitor to the category

B. Demand lost because all inventory is unavailable

C. Demand for one portfolio item reduced by demand for another related item

D. Demand generated solely by word of mouth

Correct Answer: C. Demand for one portfolio item reduced by demand for another related item

Question 8

Why should promotion performance be evaluated using contribution as well as volume?

A. Promotions always reduce demand

B. Higher volume can occur despite lower margin and promotion costs

C. Volume cannot be measured during a promotion

D. Contribution is independent of price

Correct Answer: B. Higher volume can occur despite lower margin and promotion costs

Question 9

Which is a major risk when historical sales during a stockout are used directly as demand?

A. Demand will necessarily be overstated

B. Seasonality will disappear

C. Unconstrained demand may be understated

D. Cannibalization will be eliminated

Correct Answer: C. Unconstrained demand may be understated

Question 10

For a new product, the strongest analogue is generally one that is similar in:

A. Product colour only

B. Causal market conditions such as customer, channel and price

C. Supplier name only

D. Accounting classification only

Correct Answer: B. Causal market conditions such as customer, channel and price