Specialized Forecasting Contexts

Introduction

Conventional time-series forecasting assumes that a product has a reasonably long sales history, that market conditions are not changing radically, and that demand is observed at regular intervals. Marketing decisions often violate all three assumptions. A firm may be launching a product with no historical observations, estimating demand for a technology whose adoption is accelerating, planning inventory for sporadic spare-parts demand, forecasting a fashion item with a six-week life, or estimating usage in a two-sided digital platform. These contexts require models that combine quantitative evidence with market knowledge, analogies, experimentation, and explicit judgment.

Specialized forecasting is therefore not a single technique. It is a process of matching the structure of the demand problem to an appropriate forecasting architecture. The central managerial question is not “Which model is most sophisticated?” but “What data-generating process is plausible, what decision is being supported, and how should uncertainty be represented?”

This chapter examines diffusion models, analogy-based forecasting, cohort methods, intermittent-demand methods, short-life-cycle forecasting, service and platform-market forecasting, hierarchical reconciliation, and governance of judgmental forecasts. It also considers the role of artificial intelligence, ethical risks, and organizational implementation.

Learning objectives

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

1. Diagnose why ordinary historical forecasting is inadequate in specialized contexts.
2. Select and interpret diffusion, analogy, cohort, intermittent-demand, and life-cycle approaches.
3. Construct forecasts for products, services, and platforms while distinguishing adoption, usage, and revenue.
4. Reconcile forecasts across product, channel, geographic, and organizational hierarchies.
5. Evaluate assumptions, uncertainty, bias, ethics, and governance in specialized forecasting systems.
6. Translate forecasts into inventory, capacity, pricing, launch, and investment decisions.

1. Diagnosing the forecasting context

A useful diagnosis begins with five questions:

• What is being forecast? Units sold, adopters, active users, transactions, revenue, capacity utilization, or profit?
• What is the observation process? Continuous sales, periodic purchases, censored demand, stockout-constrained sales, or unobserved potential demand?
• How much history exists? No history, sparse history, stable history, or history that is structurally obsolete?
• What drives demand? Diffusion through a population, replenishment, weather, promotions, fashion, service capacity, or network effects?
• What is the decision horizon? A launch order, weekly staffing, a five-year investment, or a real-time pricing decision?

The forecast target must be defined carefully. For a digital service, “demand” may mean registrations, active users, sessions, transactions, or paid subscriptions. These quantities are related but not interchangeable. A forecast of registrations can be impressive while paid conversion remains weak; a forecast of transactions can grow while average transaction value falls.

The forecaster should also distinguish demand from sales. If inventory is unavailable, observed sales understate demand. If a platform rejects users because of capacity constraints, observed transactions understate desired usage. Conversely, pre-orders, speculative registrations, or promotional purchases may temporarily exceed sustainable demand.

A specialized forecast should normally include a baseline, explicit drivers, scenarios, prediction intervals, and a record of assumptions. A single-point estimate hides both model risk and market risk.

2. New-product forecasting and diffusion models

2.1 The adoption problem

A new product has little or no direct sales history. Its forecast must be built from market potential, adoption timing, customer heterogeneity, competitive response, distribution, price, and communication. New-product forecasts are especially vulnerable to optimism because managers often confuse enthusiasm among early adopters with mass-market demand.

Diffusion models describe how adoption accumulates through a population. The classic Bass model separates adoption stimulated by external influence—such as advertising or publicity—from adoption stimulated by internal influence—such as word of mouth and imitation.

Let N(t) be cumulative adopters by time t, m the eventual market potential, p the coefficient of innovation, and q the coefficient of imitation. The Bass adoption-rate equation is:

Here, p and q are rates per unit time, while m and N(t) are numbers of adopters. The term m − N(t) is the remaining population that has not adopted. The model assumes a fixed potential market, homogeneous exposure structure, and adoption driven primarily by external and internal influence. It is most useful for durable or sufficiently discrete adoption decisions, not for frequent replenishment purchases.

The corresponding cumulative adoption solution, with N(0)=0, is:

The formula requires p>0, q≥0, and a consistent time unit. The adoption rate initially rises when innovation dominates, reaches a peak, and then declines as the remaining population becomes smaller. The peak occurs at:

when q>p. The peak adoption rate is:

These expressions are theoretical benchmarks. In practice, parameters must be estimated from comparable launches, market research, experiments, or early sales observations. Market potential is rarely known with precision, and parameter estimates can be highly correlated. A forecast should therefore show sensitivity to alternative values of m, p, and q.

2.2 Interpreting diffusion forecasts

A high p suggests a market influenced strongly by advertising, publicity, regulation, or visible technological need. A high q suggests strong social contagion, referral, demonstration, or network communication. Products with high social visibility may diffuse quickly but can also decline rapidly if novelty fades.

Diffusion models should be extended or supplemented when:

• adoption is constrained by supply rather than preference;
• multiple competing technologies divide the potential market;
• price changes materially affect adoption;
• replacement cycles matter;
• adoption is recurrent rather than one-time;
• the population is segmented into groups with different propensities;
• the product is subject to regulation or interoperability requirements.

A practical extension is a segmented diffusion model. Instead of one market potential, the firm may forecast early adopters, mainstream households, and institutional buyers separately. Each segment can have distinct prices, communication channels, adoption rates, and replacement behavior. Segment forecasts are then summed, subject to avoiding double counting.

2.3 Worked example: an illustrative innovation launch

Suppose an illustrative firm estimates a potential market of 100,000 adopters, with p=0.03 per year and q=0.38 per year. At the end of year 2:

e^−(p+q)t=e^−0.82≈0.440. Therefore,

N(2)=100,000(1−0.440)/[1+(0.38/0.03)(0.440)].

The numerator is approximately 56,000 and the denominator is approximately 6.573, giving about 8,520 cumulative adopters. This result is not a statement that exactly 8,520 units will be sold. It is a model-based adoption estimate under the stated assumptions. Supply constraints, returns, replacement purchases, and channel inventory must be modeled separately.

The peak occurs at approximately ln(0.38/0.03)/0.41≈6.22 years. A manager who plans production only around the initial launch may therefore underinvest in capacity, while a manager who treats the peak as certain may overinvest if a competitor or price change alters the diffusion path.

3. Forecasting without direct history: analogies and structured judgment

3.1 Analogy-based forecasting

An analogy uses demand history from a product, market, geography, customer segment, or technology believed to share relevant causal features with the focal offering. The analogy may be:

• within-category, such as a previous model of the same product;
• cross-category, such as using adoption of a related technology;
• geographic, such as transferring a launch pattern from one country to another;
• customer-based, such as using behavior of a similar institutional segment;
• price-based, such as comparing products with similar price and perceived value.

A valid analogy is not merely a product that appears similar. It should share mechanisms affecting adoption: need intensity, switching cost, distribution access, perceived risk, regulation, replacement cycle, and competitive alternatives.

A disciplined analogy process is:

1. Define the target variable and forecast horizon.
2. Identify several candidate analogies rather than one favored example.
3. Score similarity on causal dimensions.
4. Adjust for known differences in price, reach, market size, and timing.
5. Produce a range and document judgmental adjustments.
6. Validate the method through pseudo-launches or holdout tests where possible.

A simple weighted analogy forecast is:

Here, A_j,t is the forecast from analogy j at time t, w_j is its weight, and F_t is the combined forecast. The weights should reflect validated similarity, not organizational seniority or political preference. If analogies differ in market size, a scaling factor may be required, but that factor must be justified rather than chosen to produce a desired answer.

3.2 Judgmental forecasting and bias control

Expert judgment is essential when data are absent, but it is not automatically unbiased. Common biases include anchoring on an initial target, availability bias from memorable successes, optimism bias in innovation teams, confirmation bias, and strategic bias when forecasts affect budgets or performance evaluation.

Useful controls include independent estimates, anonymous initial forecasts, explicit outside-view comparisons, pre-mortems, scenario ranges, and forecast tracking. A “reference class” asks what happened across a broad set of comparable launches, not merely what happened in the most successful example. Judgment should be recorded as assumptions that can later be audited.

4. Cohort methods and customer-level dynamics

A cohort is a group sharing a meaningful start condition, such as acquisition month, signup quarter, first purchase, installation year, or contract start. Cohort analysis separates age effects from calendar-time effects and is especially valuable for subscriptions, apps, education services, and customer retention.

For cohort c and age a, let N_c,a be the number of customers active at that age and N_c,0 the number initially acquired. The retention rate is:

If revenue per active customer at age a is r_c,a, cohort revenue is:

These formulas assume that active customers and revenue are consistently defined. They do not automatically account for refunds, upgrades, downgrades, reactivation, or users who are active but unpaid.

A forecast for a future period sums contributions from existing cohorts and expected new cohorts:

The age t−c must be mapped to the same time unit as retention and revenue. Cohort analysis reveals whether growth arises from acquiring more customers, retaining them longer, increasing usage, or raising monetization. It prevents a common error: applying the retention rate of mature cohorts to newly acquired customers without considering changes in acquisition channel or customer mix.

For a subscription service, a cohort forecast may use survival analysis. If [[MATH;h_a]] is the hazard—the probability of churn at age a conditional on survival to that age—the survival probability through age a is:

This is useful when churn varies with tenure. It assumes hazards are measured consistently and that the future hazard structure is sufficiently stable or has been adjusted for known changes.

5. Intermittent demand and spare-parts forecasting

Intermittent demand contains many zero-demand periods and occasional positive demands. Examples include aircraft components, industrial maintenance parts, emergency medicines, and replacement items. Ordinary averages can respond too slowly to a recent demand event, while conventional exponential smoothing can produce unstable timing forecasts.

Two separate processes should be forecast:

1. Occurrence: when a demand event happens.
2. Size: how much is demanded when it happens.

A widely used method is Croston’s method. It maintains an estimate of nonzero demand size, z_t, and an estimate of the interval between nonzero demands, p_t. When a nonzero demand of size y_t occurs after interval [[MATH,q_t]], updates are:

The forecast per period is:

Here, 0<α<1 is the smoothing constant; z_t is in units per demand event; p_t is periods per event; and F_t+1 is units per period. The method assumes that demand size and timing can be represented separately and that the process is reasonably stable. It is not ideal when demand is lumpy, highly seasonal, trend-driven, or affected by planned maintenance schedules.

Variants such as the Syntetos–Boylan approximation and Teunter–Syntetos–Babai method address some biases and timing behavior. Method selection should be based on out-of-sample performance and the inventory decision, not solely on average forecast error. For intermittent demand, service-level measures, stockout frequency, and inventory cost can be more meaningful than mean absolute percentage error, which is undefined or misleading when actual demand is zero.

6. Short-life-cycle and fashion products

Short-life-cycle products have limited selling windows and rapid obsolescence. Examples include seasonal apparel, event merchandise, promotional bundles, and some consumer electronics. Their main forecasting challenge is that the firm must commit inventory before uncertainty resolves. Forecast accuracy and decision quality are therefore related but not identical.

The single-period inventory model, often called the newsvendor model, provides a benchmark. Let c be unit cost, p selling price, v salvage value, and D uncertain demand. The underage cost is:

and the overage cost is:

When p>c>v, the cost-minimizing order quantity Q^* satisfies:

Thus, Q^* is the demand quantile at the critical fractile. The formula assumes one primary ordering opportunity, known unit economics, and a specified demand distribution. Markdown timing, replenishment, lost goodwill, capacity constraints, and returns may require extensions.

A numerical illustration: if p=50, c=30, and v=10 monetary units, the critical fractile is (50−30)/(50−10)=0.50. The optimal order is therefore the median of the demand distribution, not necessarily the mean. If demand is normally distributed with mean 10,000 and standard deviation 2,000, the median equals the mean, so the benchmark order is 10,000 units. If the critical fractile were 0.75, the normal-distribution order would be approximately 10,000+0.674(2,000)=11,348 units. This result depends on the normal approximation and should not be used when demand is strongly skewed without checking the distribution.

Short-life-cycle forecasting should integrate early signals: pre-orders, search activity, test-market conversion, social engagement, retailer commitments, and comparable launch curves. These signals should be calibrated because attention is not equivalent to purchase intent.

7. Services and capacity-constrained demand

Services cannot generally be inventoried. An unsold airline seat, hotel room, consultation hour, or streaming-capacity slot may expire. Forecasting must therefore connect demand with capacity, timing, cancellations, no-shows, and customer waiting.

Service demand often has multiple layers:

• arrivals or booking requests;
• accepted bookings;
• completed service encounters;
• resource minutes consumed;
• revenue and contribution margin.

A hospital may forecast arrivals but need operating-room hours; a call center may forecast contacts but need workload by skill group; a platform may forecast sessions but need server capacity. Aggregating these measures without conversion assumptions creates capacity errors.

For a service period, let [[MATH;λ]] be expected arrivals per unit time and [[MATH;E[S]]] average service time. Offered workload is:

where [[MATH;s]] is the number of parallel servers or service resources. The ratio ρ is utilization and must generally be below 1 for a stable queue, although practical service systems require a buffer below 100 percent because variability causes waiting. This expression is a workload identity, not a complete waiting-time model; arrival and service distributions, queue discipline, abandonment, and priorities are also relevant.

Forecasts should distinguish unconstrained demand from capacity-limited transactions. Revenue-management systems then use prices, availability controls, overbooking, and segmentation to allocate scarce capacity. Ethical practice requires that differential pricing be transparent, legally compliant, and not discriminatory on protected characteristics.

8. Platform markets and network effects

A platform connects two or more participant groups, such as buyers and sellers, riders and drivers, advertisers and audiences, or developers and users. Demand forecasting must model both sides and their interactions. A forecast of buyer activity that ignores seller availability may be infeasible; a forecast of drivers that ignores rider demand may produce excess supply and poor economics.

Important platform measures include registrations, active participants, match rate, transactions, repeat rate, take rate, cancellations, liquidity, and contribution margin. Let B_t be buyers, S_t sellers, and [[MATH;M_t]] completed matches. A simple operational constraint is:

where d_t and s_t represent potential matching capacity per participant. This is a simplified bound, not a behavioral law. Geography, timing, quality, search frictions, and compatibility can reduce matches further.

Network effects can create nonlinear adoption. Positive cross-side effects occur when more participants on one side increase value to the other side. However, congestion, fraud, poor quality, privacy concerns, and multi-homing can weaken or reverse the effect. Forecasts should therefore use scenarios rather than extrapolating historical growth mechanically.

A platform forecast commonly combines:

1. acquisition and activation cohorts;
2. retention and frequency by participant segment;
3. supply-demand matching;
4. pricing and incentive response;
5. fraud, quality, and capacity constraints;
6. unit economics and contribution margin.

The relevant revenue identity may be:

where GMV_t is gross merchandise value and τ_t is the take rate. The terms must be measured consistently; gross value is not platform revenue.

9. Hierarchical forecasting and reconciliation

Organizations forecast at multiple levels: total company, region, channel, product family, SKU, customer segment, and time bucket. Independently produced forecasts rarely add up. A regional forecast may exceed the corporate target; SKU forecasts may not equal the category forecast; sales and finance may use different definitions of the same measure.

A hierarchy can be represented by a summing matrix S. If b_t is a vector of bottom-level forecasts, aggregate forecasts are y_t=Sb_t. In practice, forecasts may be generated at several levels and reconciled to produce a coherent vector ŷ_t satisfying:

where ĝ_t is a reconciled bottom-level or equivalent forecast representation. The requirement is coherence: every parent forecast equals the sum of its relevant children.

9.1 Reconciliation approaches

• Bottom-up: forecast the most detailed series and sum upward. It preserves local information but can be noisy when lower-level series are sparse.
• Top-down: forecast the aggregate and allocate it using historical proportions or other weights. It is stable but can miss structural changes at lower levels.
• Middle-out: forecast an intermediate level and distribute upward and downward.
• Optimal reconciliation: combine forecasts at multiple levels using estimated error covariance. In a generalized least-squares form, reconciliation seeks a coherent forecast close to the base forecasts, with less weight given to highly uncertain series.

A simple proportional allocation of a parent forecast P_t to child i is:

The weights may be historical shares, recent shares, margin-weighted shares, or model-based estimates. Historical shares should not be treated as permanent when prices, distribution, or product mix are changing.

Reconciliation is not merely a mathematical exercise. It clarifies accountability. Finance may own the enterprise target, marketing may own category assumptions, sales may own account intelligence, and operations may own capacity constraints. A governance process must specify which adjustments are evidence-based, which are managerial overrides, and how conflicts are resolved.

10. Integrated application: illustrative launch of a connected home service

Background

The following is a fictional illustrative case. NovaHome plans to launch a connected energy-management service combining a device, mobile application, and monthly subscription. It has no direct history in the target country. Management must choose an initial device order, staffing level, subscription target, and three-year investment plan.

Problem

The product has three related but distinct forecasts: device adoption, subscription retention, and support workload. Executives initially propose a single sales target based on a successful launch in another country. Operations argues that installation capacity will constrain early sales, while finance requires a coherent forecast by region and channel.

Analysis

The team constructs three analogies: a previous smart-home device, a subscription energy application, and a regulated utility program. It scores similarity on price, channel reach, customer need, installation complexity, and competitive intensity. The reference class produces low, central, and high market-potential scenarios rather than one estimate.

A Bass model is fitted to the most comparable adoption curves, with parameters varied by scenario. Cohort assumptions then convert adopters into active subscribers. For each monthly acquisition cohort, retention is estimated from comparable subscription services, with a conservative adjustment for the higher installation burden. Support contacts are forecast from active subscribers and expected contacts per subscriber.

The model also separates unconstrained demand from installed demand. If demand requests exceed installer capacity, observed installations are capped and the model records the lost or deferred demand. Regional and channel forecasts are reconciled to the national target using a middle-out process: regional marketing teams forecast channel detail, the national team sets an independently reviewed total, and a reconciliation algorithm adjusts the detailed forecasts while preserving major evidence-based differences.

Approach and solution

The launch plan uses three stages:

1. Pilot: a controlled regional launch tests price, activation, installation time, and retention.
2. Learning update: early cohort and funnel data update adoption and churn distributions; no single early month is extrapolated without adjustment.
3. Scaled launch: inventory and staffing are committed using scenario-weighted demand and explicit service-level targets.

The executive dashboard reports adopters, installations, active subscribers, churn, support workload, contribution margin, and forecast intervals. It does not report registrations as if they were revenue.

Outcome

The pilot reveals that the most attractive customer segment has high conversion but also unusually high installation complexity. The central sales forecast is reduced, but the service forecast is increased because these customers generate more support contacts. The firm avoids a stockout of devices and prevents overstaffing in low-demand regions. After reconciliation, product, regional, channel, and finance forecasts sum to the same national totals.

Learning points

The case demonstrates that specialized forecasting is an integrated system. Diffusion describes adoption; cohort analysis describes persistence; capacity analysis describes what can actually be delivered; reconciliation creates organizational coherence. No individual model is sufficient.

11. Validation, uncertainty, and governance

Forecast evaluation should match the decision. For continuous demand, common measures include mean absolute error:

and root mean squared error:

Here, y_t is actual demand, ŷ_t is forecast demand, and n is the number of observations. MAE is in the original demand unit; RMSE gives greater influence to large errors. Percentage measures should be used cautiously with zeros or values close to zero.

For intermittent demand, evaluate inventory consequences: fill rate, stockout frequency, average inventory, obsolescence, and total cost. For new products, evaluate forecast intervals, scenario calibration, adoption-curve timing, and decision outcomes rather than expecting precise point accuracy before launch.

Forecast uncertainty has at least four sources: parameter uncertainty, process uncertainty, market-structure uncertainty, and implementation uncertainty. Scenario analysis should vary causal assumptions, not merely add arbitrary percentages. For example, scenarios may vary market potential, price elasticity, competitive entry, adoption speed, retention, and capacity.

AI and machine learning can assist with feature extraction, analogous-product search, demand sensing, causal experiments, and probabilistic forecasting. They do not eliminate the need for target definition, leakage control, regime-change detection, and governance. A model trained on historical sales may reproduce stockout constraints or discriminatory allocation patterns. Human overrides should be logged, and models should be monitored for drift, calibration, and subgroup performance.

Ethically responsible forecasting requires privacy protection for cohort and platform data, explainable use of sensitive variables, non-discriminatory pricing and allocation, security against manipulation, and clear accountability for automated decisions. Forecasts should not be used to justify exploitative scarcity, deceptive urgency, or exclusion of vulnerable customers.

12. Best-practice implementation framework

A practical implementation sequence is:

1. Define the decision and target. Specify unit, population, time horizon, and whether the target is demand or constrained sales.
2. Map the demand process. Identify adoption, repeat purchase, churn, capacity, network, and competitive mechanisms.
3. Segment where mechanisms differ. Avoid averaging away meaningful heterogeneity.
4. Select complementary models. Combine diffusion, analogy, cohorts, intermittent methods, and capacity models where appropriate.
5. Create a reference class. Use broad comparable evidence and document differences.
6. Build scenarios and intervals. Make uncertainty actionable for inventory, staffing, and investment.
7. Reconcile hierarchies. Enforce coherent totals while preserving reliable local information.
8. Pilot and update. Treat launch as a learning process, not a one-time forecast event.
9. Evaluate decision outcomes. Track service, profit, inventory, retention, and forecast calibration.
10. Govern judgment and technology. Record assumptions, overrides, data provenance, and ethical controls.

Challenges and limitations

Specialized models face recurring limitations. Market potential is difficult to identify because the boundary of the market changes with price and technology. Analogies can conceal structural differences. Diffusion models may mistake supply expansion for adoption. Cohort models can confound acquisition-channel changes with customer aging. Intermittent-demand models may fail when maintenance schedules create predictable bursts. Short-life-cycle models are sensitive to salvage assumptions and markdown policy. Platform forecasts can become self-fulfilling when incentives create the very growth being forecast.

Organizational incentives are equally important. Sales teams may inflate forecasts to secure inventory; finance may suppress them to protect budgets; product teams may select favorable analogies. Reconciliation can create an appearance of precision if definitions and assumptions are not aligned. The remedy is not to remove judgment, but to make it explicit, testable, and accountable.

Chapter Summary

Specialized forecasting contexts require a model of the demand mechanism, not merely an extension of a historical time series. New products can be forecast with diffusion models, analogies, market research, and staged learning. The Bass model separates external innovation from internal imitation, but its assumptions about market potential and adoption structure must be tested through scenarios.

Analogy methods are valuable when direct history is absent, provided that analogies are selected on causal similarity and adjusted transparently. Cohort methods forecast acquisition, retention, usage, and monetization by customer age. Intermittent-demand methods separate the timing and size of nonzero demand and should be assessed using inventory consequences. Short-life-cycle forecasting links demand distributions to ordering and pricing decisions through overage and underage costs.

Service forecasts must distinguish arrivals, completed encounters, resource workload, and revenue. Platform forecasts must model both sides of the market, matching constraints, network effects, and unit economics. Hierarchical reconciliation ensures that product, channel, regional, and corporate forecasts add up coherently. Across all contexts, scenario analysis, pilots, forecast tracking, ethical governance, and explicit uncertainty are essential. The professional forecaster’s task is therefore both analytical and organizational: to produce useful evidence and to create a disciplined process for acting under uncertainty.

Key Terms

Diffusion model: A model describing how adoption spreads through a population over time.

Bass model: A diffusion model separating innovation and imitation effects.

Market potential: The maximum number of customers or adopters under defined conditions.

Analogy forecasting: Forecasting based on evidence from comparable products, markets, or customers.

Reference class: A broad set of comparable cases used to reduce inside-view bias.

Cohort: A group sharing a common start event or time period.

Retention rate: The proportion of an initial cohort remaining active at a specified age.

Hazard: The conditional probability of an event, such as churn, at a given age.

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

Croston’s method: A method forecasting nonzero demand size and demand intervals separately.

Short-life-cycle product: A product with a limited selling window or rapid obsolescence.

Underage cost: The cost of ordering too little relative to realized demand.

Overage cost: The cost of ordering too much relative to realized demand.

Critical fractile: The demand probability used to determine a single-period order quantity.

Service demand: Demand for encounters, capacity, time, or resource consumption rather than inventory units.

Platform market: A market in which a business facilitates interaction between participant groups.

Network effect: A change in value resulting from changes in the number or activity of other participants.

Hierarchical forecasting: Forecasting across aggregated and disaggregated organizational levels.

Reconciliation: Adjusting forecasts so that related levels sum consistently.

Forecast governance: Policies and processes controlling data, models, judgment, accountability, and ethical use.

Review Questions

Short Answer Questions

1. Why is observed sales data not always a valid measure of demand?
2. What do the parameters p and q represent in the Bass model?
3. Why are cohort methods useful for subscription services?
4. What are the two processes separated by Croston’s method?
5. What is the purpose of hierarchical forecast reconciliation?

Long Answer Questions

1. Explain the assumptions, interpretation, and limitations of the Bass diffusion model.
2. Compare analogy forecasting with cohort forecasting in a new-product launch.
3. Explain how the newsvendor critical fractile supports decisions for short-life-cycle products.
4. Discuss the forecasting challenges created by two-sided platform markets.
5. Design a governance process for combining statistical forecasts and managerial judgment.

Analytical / Application Questions

1. A firm is launching a subscription service with no direct history but has three comparable services. Design a forecasting process using analogies, cohorts, scenarios, and pilot data.
2. A spare-parts item has frequent zero-demand months and highly variable positive demand. Explain how you would forecast it and evaluate the forecast for inventory decisions.
3. A national sales forecast is 20,000 units, while regional and channel forecasts sum to 23,500. Propose a reconciliation approach and explain what evidence should determine the allocation of the adjusted total.

Multiple Choice Questions with Answers

Question 1

In the Bass diffusion model, the coefficient q primarily represents:

A. Market potential

B. Influence from imitation and word of mouth

C. Unit production cost

D. Inventory service level

Correct Answer: B. Influence from imitation and word of mouth

Question 2

If p = 0.03 and q = 0.38 per year, which force is relatively stronger in the diffusion process?

A. External innovation

B. Internal imitation

C. Salvage value

D. Capacity utilization

Correct Answer: B. Internal imitation

Question 3

A cohort forecast is especially useful because it separates:

A. Price from cost only

B. Age effects from calendar-time and acquisition effects

C. Supply from accounting profit

D. Advertising from regulation

Correct Answer: B. Age effects from calendar-time and acquisition effects

Question 4

Croston’s method forecasts intermittent demand by separately estimating:

A. Price and margin

B. Demand size and interval between nonzero demands

C. Market share and market potential

D. Capacity and utilization

Correct Answer: B. Demand size and interval between nonzero demands

Question 5

If a product sells for 50, costs 30, and has salvage value 10, its newsvendor critical fractile is:

A. 0.25

B. 0.40

C. 0.50

D. 0.75

Correct Answer: C. 0.50

Question 6

Which measure is most problematic for evaluating intermittent demand when many actual observations are zero?

A. Mean absolute error

B. Root mean squared error

C. Mean percentage error based on division by actual demand

D. Fill rate

Correct Answer: C. Mean percentage error based on division by actual demand

Question 7

In a platform market, gross merchandise value multiplied by take rate most directly estimates:

A. Platform transaction revenue

B. Total number of users

C. Supplier capacity

D. Customer churn

Correct Answer: A. Platform transaction revenue

Question 8

A coherent hierarchical forecast requires that:

A. Every lower-level forecast is identical

B. Parent forecasts equal the sums of their relevant child forecasts

C. Only the top-level forecast is used

D. Historical proportions never change

Correct Answer: B. Parent forecasts equal the sums of their relevant child forecasts

Question 9

For service capacity, ρ = λE[S]/s represents:

A. Market potential

B. Utilization or offered workload relative to parallel capacity

C. Customer retention

D. The innovation coefficient

Correct Answer: B. Utilization or offered workload relative to parallel capacity

Question 10

Which practice best reduces anchoring bias in new-product forecasting?

A. Adopting the first executive target

B. Using only the most successful analogy

C. Obtaining independent estimates and comparing with a broad reference class

D. Removing all managerial judgment

Correct Answer: C. Obtaining independent estimates and comparing with a broad reference class