7. Specialized Forecasting Problems

Introduction and Context

Most forecasting methods perform best when demand is reasonably frequent, patterns are stable, and historical observations describe the future adequately. Supply chains, however, contain many items for which these assumptions do not hold. Spare parts may be demanded only occasionally; emergency equipment may show long periods of zero demand followed by an unusually large order; a newly launched product has little or no history; and promotional activity can create temporary surges that do not represent normal market demand. In addition, demand for one product may reduce demand for another, while stock-outs may make observed sales lower than actual customer requirements.

These situations are not merely technical exceptions. They have direct consequences for inventory investment, production planning, procurement, transport capacity, service levels, and commercial decision-making. A forecast that is statistically acceptable but operationally inappropriate can create excess inventory, missed sales, obsolete stock, or unreliable customer service. Specialized forecasting therefore requires the planner to understand the structure and origin of demand before selecting a model.

This chapter examines intermittent and lumpy demand, forecasting for new products and short life cycles, promotional demand, cannibalization and substitution, and hierarchical demand planning. It also considers Croston-type methods, analogue and attribute-based forecasting, causal models, and forecast reconciliation. The central theme is that specialized forecasting is not a search for one universally superior algorithm. It is the disciplined matching of demand characteristics, business knowledge, data quality, and supply-chain decisions.

Learning Objectives

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

1. Distinguish intermittent, lumpy, slow-moving, promotional, substituted, and hierarchical demand.

2. Explain why conventional time-series methods can perform poorly in specialized forecasting situations.

3. Apply the logic of Croston-type methods and evaluate their limitations.

4. Develop forecasting approaches for new products and short life-cycle items using analogues and attributes.

5. Separate baseline demand from promotional uplift and assess cannibalization effects.

6. Explain how substitution, stock-outs, and assortment changes distort observed demand.

7. Understand bottom-up, top-down, middle-out, and reconciled hierarchical forecasting.

8. Select appropriate accuracy measures and make managerial decisions under uncertainty.

9. Recognize governance, ethical, technological, and implementation issues in advanced demand planning.

7.1 When Ordinary Forecasting Assumptions Break Down

A conventional forecasting model generally assumes that the observed series contains enough regular information to estimate level, trend, and possibly seasonality. In specialized problems, one or more of these assumptions fails.

Intermittent demand contains many periods with zero demand. Demand occurs irregularly, although the quantity demanded on a non-zero occasion may be relatively stable. A replacement filter for industrial machinery is a typical example.

Lumpy demand is both intermittent and highly variable in size. It contains many zero observations and substantial variation in the quantities demanded when demand occurs. Service parts, emergency components, and project-based materials frequently exhibit this pattern.

Slow-moving demand refers to low-volume demand that may or may not be intermittent. The term is operational rather than strictly statistical and often describes items that require special inventory policies.

Promotional demand arises partly from a planned marketing intervention, such as a discount, display, coupon, bundle, or advertising campaign. The observed sales increase may include genuine incremental demand, timing shifts, forward buying, and purchases that would otherwise have occurred later.

New-product demand has limited or no direct historical observations. Its forecast must rely on market information, comparable products, product attributes, customer research, and managerial assumptions.

Short life-cycle demand occurs when a product is sold for a limited period, as with fashion items, seasonal models, event merchandise, or technology products. The forecast must address both demand uncertainty and the cost of being left with obsolete inventory.

Substitution and cannibalization occur when demand moves between related products. Substitution is often driven by availability, price, or customer preference; cannibalization describes the reduction in an existing product’s demand caused by introducing or promoting another product from the same portfolio.

Hierarchical demand is structured across levels such as total market, region, channel, product family, stock-keeping unit, and customer. Forecasts at separate levels may not add up, creating operational and financial inconsistency.

These categories can overlap. A newly launched spare part may be intermittent, a promoted product may cannibalize another item, and a lumpy SKU may be part of a stable product-family total. Classification is therefore a starting point, not a permanent label.

7.2 Intermittent and Lumpy Demand

Why zero observations matter

For intermittent demand, the average of all observations can be misleading. Suppose monthly demand for a spare part is:

0, 0, 0, 10, 0, 0, 0, 12, 0, 0, 8, 0.

The mean monthly demand is 2.5 units, but the supply decision cannot be based on an assumption that 2.5 units are demanded every month. The planner must estimate at least two different features: the size of a demand occurrence and the interval between demand occurrences.

Let the non-zero demand sizes be represented by \(z_t\), and let the intervals between non-zero demands be represented by \(p_t\). A simple conceptual estimate of expected demand per period is:

\[

\hat{D} = \frac{\text{estimated demand size}}{\text{estimated interval between demands}}.

\]

This decomposition is valuable because the two components may behave differently. Demand size may be stable while occurrence intervals vary, or the reverse may be true.

Croston’s method

Croston’s method was designed for intermittent demand. Instead of updating a demand-size estimate in every period, it updates the size estimate and the interval estimate only when a non-zero demand occurs. If the latest non-zero demand is \(z_t\), the previous estimated size is \(\hat{z}\), and the interval since the previous demand is \(p_t\), the smoothed estimates can be represented as:

\[

\hat{z}_t = \hat{z}_{t-1} + \alpha(z_t - \hat{z}_{t-1}),

\]

\[

\hat{p}_t = \hat{p}_{t-1} + \alpha(p_t - \hat{p}_{t-1}),

\]

with the forecast approximately equal to:

\[

\hat{D}_t = \frac{\hat{z}_t}{\hat{p}_t}.

\]

Here, \(\alpha\) is a smoothing parameter. A low value gives a more stable estimate but reacts slowly to structural change; a high value reacts quickly but can overreact to unusual demand events.

Croston’s central insight is that a zero-demand period should not be treated as evidence that the expected size of demand has fallen. It is evidence about the timing of the next demand occurrence. This makes the method more appropriate than ordinary exponential smoothing for many intermittent series.

Croston variants

Several modifications have been developed because the original method can be biased and may respond slowly to changes. Two commonly discussed variants are:

• Syntetos–Boylan adjustment: modifies the forecast to reduce systematic bias associated with the original ratio of smoothed size to smoothed interval.
• Teunter–Syntetos–Babai method: models the probability of a non-zero demand and the demand size separately, updating occurrence probability even during periods without demand.

The choice among variants should be based on empirical validation and the operational consequences of errors. No variant eliminates uncertainty, especially where demand is generated by failures, projects, or irregular customer decisions.

Classification using demand characteristics

A useful diagnostic separates demand according to the average interval between non-zero demands and the variability of demand size. One common approach uses the average demand interval and the squared coefficient of variation of non-zero demand. Low interval and low variability indicate relatively smooth demand; high interval and low variability suggest intermittent demand; low interval and high variability indicate erratic demand; and high interval and high variability indicate lumpy demand. The precise thresholds should be calibrated to the organization’s data and decision context rather than adopted mechanically.

Inventory implications

A point forecast alone is insufficient for intermittent items. The planner must also consider lead time, required service level, demand occurrence probability, criticality, repair policy, and the possibility of emergency procurement. For a critical aircraft component, holding stock may be justified even when statistical demand is extremely low. For a low-value office item, periodic review and pooled inventory may be more economical.

Demand aggregation is often effective. Several related locations may be pooled into a central inventory position, or multiple interchangeable parts may be managed as a family. Aggregation reduces variability relative to separate forecasts, although it may increase response time or transportation requirements.

7.3 Forecasting New Products and Short Life Cycles

The information problem

A new product has no direct demand history, but it is not necessarily without information. The firm may know the product’s price, features, target segment, distribution plan, launch timing, marketing expenditure, channel coverage, and similarity to existing products. The forecasting problem is to convert these sources into a transparent and revisable demand estimate.

A new-product forecast should normally be expressed as a range or probability distribution rather than a single unquestioned number. The range should reflect uncertainty in market size, adoption rate, conversion, repeat purchase, channel acceptance, and supply availability.

Analogue forecasting

An analogue is an earlier product or market situation considered sufficiently similar to provide evidence. A useful analogue is not simply a product with a similar name. It should resemble the new product in dimensions such as customer segment, price, function, channel, geographical market, launch support, competitive intensity, and life-cycle position.

The procedure generally involves:

1. Defining the demand event to be forecast, such as first-month sales or cumulative sales during the first quarter.

2. Identifying several candidate analogues rather than relying on one remembered success.

3. Measuring similarities and differences explicitly.

4. Adjusting the analogue trajectory for price, distribution, marketing, capacity, and market conditions.

5. Combining statistical evidence with expert judgement.

6. Monitoring early sales and updating the forecast.

Analogues are particularly useful when historical product launches are available but the sample is small. Their weakness is that apparent similarity can conceal important differences, and successful products are often more visible than unsuccessful ones, creating selection bias.

Attribute-based models

Attribute-based forecasting estimates demand from product or market attributes. Explanatory variables may include price, size, technical performance, brand strength, delivery promise, channel availability, customer rating, or sustainability characteristics. The model can be estimated from existing products and applied to the new product’s attributes.

For example, a retailer may estimate the relationship between weekly sales and price, display position, product rating, and delivery availability across comparable products. A new item’s forecast is then generated from its planned attributes. Such models are more systematic than informal analogy, but they depend on data quality and on the assumption that historical relationships remain relevant.

Discrete-choice models can be useful where customers choose among alternatives. They estimate the probability that a customer selects a product as a function of attributes and competing options. These models are valuable for assortment design, pricing, and substitution analysis, although they require suitable customer or market data.

Diffusion and adoption models

Some innovations show an adoption pattern in which initial users influence later buyers. Diffusion models distinguish between innovation effects, generated by external communication or market exposure, and imitation effects, generated by social influence or existing adoption. Such models can support medium-term launch planning, but early observations are often too limited to estimate parameters reliably. Scenario analysis should therefore accompany the model.

Short life-cycle decisions

For a short life-cycle product, the forecast must be linked to an order decision. Ordering too little causes lost sales during the selling window; ordering too much creates markdowns, returns, disposal, or obsolete stock. The newsvendor logic is useful for a single selling season. If underage cost is the cost of ordering too little and overage cost is the cost of ordering too much, the target service fractile is:

\[

CR = \frac{C_u}{C_u + C_o}.

\]

The inventory quantity is selected at the corresponding percentile of the demand distribution. The formula is not a substitute for judgement: salvage value, replenishment options, capacity, strategic brand effects, and fairness among channels may alter the decision.

A life-cycle forecast should distinguish launch, growth, maturity, decline, and end-of-life phases. The forecast may need to be revised frequently as new information arrives. When lead times are long, postponement, modular design, late differentiation, flexible contracts, and rapid replenishment can reduce forecast dependence.

7.4 Promotions, Cannibalization, and the Difference between Sales and Demand

Baseline and incremental demand

Observed sales during a promotion combine several effects:

\[

\text{Promotional sales} = \text{baseline sales} + \text{incremental uplift} + \text{timing shift} - \text{cannibalization} + \text{availability effects}.

\]

Baseline demand is the level expected without the promotion. Incremental uplift is the additional demand created by the intervention. Timing shift occurs when customers buy earlier than they otherwise would have done. Cannibalization occurs when sales of one item reduce sales of another item in the portfolio. Availability effects arise when inventory or distribution changes constrain or enable sales.

Treating all promotional sales as permanent demand produces systematic overforecasting after the promotion ends. Conversely, removing the entire promotion effect may understate future demand if the campaign attracts new customers or increases repeat purchase.

Promotional forecasting methods

A basic approach estimates baseline demand from non-promotional periods and applies an uplift factor by promotion type. More advanced models use price, discount depth, display, advertising, seasonality, competitor activity, and inventory availability. Regression, causal inference, and machine-learning methods can be used, but the model must distinguish correlation from causal impact.

Promotions should be coded in the data at a sufficiently detailed level. A single binary variable indicating whether a promotion occurred may be inadequate if discounts, displays, media support, and duration differ substantially. The model should also account for lagged and post-promotional effects, including pantry loading and post-event dips.

Cannibalization and substitution

Cannibalization is not always undesirable. A new product may reduce sales of an older product but improve total category profit, defend market share, or prevent competitors from entering. The relevant question is therefore not simply whether an individual SKU loses sales, but whether the portfolio achieves its commercial and strategic objective.

Substitution may be:

• Customer-driven, when buyers voluntarily select an alternative based on preference or price.
• Availability-driven, when the preferred product is out of stock.
• Policy-driven, when the firm recommends or allocates an alternative.
• Functional, when products provide similar technical utility.

A forecasting system that ignores substitution may interpret the decline of one item as falling category demand, or interpret the rise of another as genuine market growth. Availability data, lost-sales indicators, product similarity, and customer choice information are therefore important.

Market response and stock-outs

Sales are censored observations of demand. If ten customers want an item but only six units are available, recorded sales may be six, not ten. Promotional periods frequently create stock-outs, making the apparent promotional uplift unreliable. Forecasting teams should flag constrained periods and estimate lost sales where possible through order records, substitution, waiting lists, web searches, or comparable locations.

7.5 Forecasting across a Demand Hierarchy

Supply chains commonly maintain forecasts at multiple levels. A manufacturer may forecast total demand, region, channel, product family, brand, and SKU. A retailer may use corporate, country, store cluster, store, category, and item levels. Forecasts independently generated at these levels may not sum consistently. For example, regional forecasts may total 9,800 units while the national forecast is 10,500 units.

Main approaches

Bottom-up forecasting begins at the lowest level and aggregates upward. It preserves local information and product detail but can be noisy when individual SKU or location histories are sparse.

Top-down forecasting begins with an aggregate forecast and allocates it to lower levels using historical proportions, recent shares, or other allocation rules. It is stable at high levels but may overlook local changes and new product effects.

Middle-out forecasting selects an intermediate level, such as product family or regional category, and forecasts both upward and downward from that level. It can be practical where middle-level data are more reliable than either extreme.

Forecast reconciliation produces forecasts that are coherent across the hierarchy. Let \(y\) represent the vector of all hierarchical series and let \(b\) represent a vector of bottom-level series. A summing matrix \(S\) maps bottom-level values to all levels:

\[

y = Sb.

\]

Reconciliation adjusts initial forecasts so that the final forecasts satisfy these aggregation relationships. Methods range from simple proportional allocation to weighted least squares, in which more reliable series receive greater weight. The goal is not merely mathematical consistency; it is improved decision quality across procurement, finance, operations, and sales.

Hierarchical and grouped structures

A product may belong simultaneously to a region, channel, brand, and category. Such structures are sometimes called grouped time series rather than a single hierarchy. Reconciliation becomes more complex because several aggregation paths must agree. The organization must define the dimensions that matter operationally and ensure that master data are stable.

New products and sparse lower levels

A common mistake is to force a new product into the same bottom-up process used for mature items. With no local history, a top-down allocation based on category potential, analogue shares, distribution plans, and launch objectives may be more defensible. As observations accumulate, the product can progressively move toward a more data-driven lower-level forecast.

7.6 A Practical Workflow for Specialized Forecasting

Specialized forecasting benefits from a structured workflow rather than ad hoc model selection.

First, define the decision. Forecasting for replenishment, capacity, financial planning, promotion evaluation, and product launch may require different horizons, aggregation levels, and loss functions.

Second, diagnose the demand. Examine zero frequency, demand-size variability, lead time, seasonality, promotions, stock-outs, returns, substitutions, and product relationships. Visual inspection and data profiling are essential.

Third, classify the item provisionally. Use categories such as intermittent, lumpy, new, promotional, short life cycle, or substitution-sensitive. Classification should be revisited when the demand process changes.

Fourth, construct a suitable benchmark. Even advanced models should be compared with simple alternatives such as moving averages, seasonal naïve forecasts, Croston-type methods, analogue forecasts, or current-plan assumptions.

Fifth, select the model and incorporate business information. The model should be explainable enough for planners and commercial managers to challenge assumptions.

Sixth, evaluate using a rolling-origin design. A random train-test split can be inappropriate for time series. Back-testing should reproduce the information available at the time the forecast would have been made.

Seventh, assess both statistical and operational performance. Metrics may include mean absolute scaled error, weighted absolute percentage error, bias, service level, stock-outs, inventory, obsolescence, and profit. For intermittent demand, ordinary percentage errors can be undefined or disproportionately influenced by small denominators.

Eighth, create scenarios and prediction intervals. A single point estimate conceals uncertainty, especially for launches and lumpy demand. Planners should consider high, central, and low cases linked to explicit assumptions.

Ninth, reconcile and communicate. Forecasts must be coherent across products, locations, channels, and financial plans. Assumptions, overrides, and reasons for changes should be recorded.

Tenth, monitor forecast performance and model drift. A method that performs well during stable operations may fail after a product redesign, competitor entry, service-policy change, or distribution shift.

7.7 Technology, Analytics, and Governance

Enterprise planning systems can automate demand classification, model selection, hierarchical aggregation, exception alerts, and forecast version control. Machine-learning systems may identify nonlinear relationships among price, availability, marketing activity, weather, web traffic, and customer behaviour. They are most valuable when they improve decisions rather than merely increase algorithmic complexity.

Artificial intelligence can support analogue discovery, feature generation, anomaly detection, and probabilistic forecasting. However, automated systems may reproduce historical biases, confuse stock-constrained sales with demand, or treat promotional correlation as causation. Human review remains necessary for launches, exceptional events, strategic products, and changes in business policy.

Governance should address data ownership, model validation, override authority, audit trails, access control, and the treatment of sensitive customer information. Commercial teams should not inflate forecasts to obtain inventory or manipulate performance targets, while operations teams should not suppress forecasts merely because capacity is constrained. A forecast should represent expected demand; a separate plan should represent the response to that demand.

7.8 Challenges and Common Mistakes

A frequent error is applying ordinary exponential smoothing to a series dominated by zeros. Another is interpreting a single large order as a permanent level shift when it may represent a project or one-time failure.

For new products, planners may select an analogue because it is memorable rather than statistically similar. They may also create a precise forecast before distribution and marketing assumptions are settled. In short life cycles, organizations often ignore salvage values and end-of-life constraints until too late.

Promotional forecasting commonly fails because baseline demand is poorly estimated, promotions are not coded consistently, stock-outs are treated as low demand, and post-promotion dips are omitted. Portfolio forecasts fail when cannibalization is ignored or when a new item’s growth is counted as entirely incremental.

Hierarchical planning fails when product and location master data are inconsistent, forecasts are reconciled only at the end of the process, or planners override models without documenting reasons. Finally, evaluation can be misleading if the metric rewards a low forecast for an item that was frequently out of stock.

7.9 Illustrative Case: Forecasting a New Service Part with Intermittent Demand

Background

A fictional industrial-equipment manufacturer, Meridian Handling Systems, introduces a redesigned hydraulic control module for a fleet of warehouse vehicles. The module is expected to replace an older module gradually. The replacement part has a high unit value, long supplier lead time, and critical service implications. Demand is expected to be intermittent because modules are required only when failures occur.

Situation and problem

The service division initially forecasts the new module by applying the average monthly demand of the old module. This generates a forecast of 18 units per month. The approach ignores the gradual fleet conversion, differences in failure rates, customer stockholding, and the fact that some customers can use a refurbished old module temporarily.

The company must determine an initial stocking policy for regional depots and a central warehouse. It must also estimate whether demand for the old module will be cannibalized by the new module or merely transferred as installed equipment reaches the end of its life.

Analysis

Historical demand for the old module shows many zero months and occasional orders ranging from one to nine units. The service team classifies the series as lumpy. Failure records indicate that the new module has a different expected failure profile because its design is more robust. The installed base is known by customer and region, but the conversion schedule is uncertain.

The planning team develops three demand layers. First, an installed-base model estimates the number of eligible vehicles by region. Second, a failure-rate assumption converts the eligible population into expected replacement events. Third, a Croston-type method is applied to early observed orders to update the occurrence interval and demand size. The team also creates high and low scenarios for customer conversion and failure rates.

Because the regional histories are sparse, the company forecasts the total network demand first and allocates initial stock using installed-base exposure, service criticality, and regional lead times. Forecasts are reconciled with the total company plan. The old module is forecast separately, but the two items are linked through a transition matrix showing the expected movement from old-module demand to new-module demand.

Approach and solution

The company establishes a central pool for the expensive module and holds smaller emergency quantities at selected depots. The forecast is reviewed monthly, but inventory decisions use lead-time demand and service-criticality rules rather than the point forecast alone. The planning team records actual failures, customer substitutions, emergency shipments, and cases in which the old module was used instead.

The first six months are treated as a learning period. Forecast accuracy is assessed with scaled absolute errors and service outcomes, not ordinary percentage error. A review process separates genuine demand changes from one-time fleet projects and from demand suppressed by stock-outs.

Outcome

The revised approach produces a lower initial average forecast than the original 18-unit assumption but a wider uncertainty range. Central pooling protects service without requiring every depot to carry a full quantity. As failure observations accumulate, the occurrence estimate is updated and regional allocations are adjusted. The company avoids treating early launch orders as a permanent demand level and prevents the decline of the old module from being misinterpreted as total market loss.

Learning points

The case demonstrates that intermittent demand forecasting must be connected to the physical driver of demand. Croston-type methods can estimate demand timing and size, but installed-base information, substitution rules, pooling, and criticality are equally important. A forecast should support a stocking decision, not replace the decision. Finally, coherent forecasts across old and new products help management understand portfolio transition and avoid double counting.

Best Practices for Managers and Planners

Effective specialized forecasting generally follows five principles. First, model the demand-generating process, not just the sales history. Second, separate baseline demand, promotional effects, availability constraints, and portfolio transfers. Third, use aggregation and pooling where item-level data are too sparse. Fourth, combine statistical models with transparent business assumptions, especially for launches and life-cycle transitions. Fifth, evaluate forecasts using operational consequences such as service, inventory, obsolescence, and profit as well as statistical accuracy.

Organizations should maintain a forecast taxonomy, standardized event data, product-relationship records, and a documented exception process. Models should be benchmarked regularly, assumptions should be visible, and manual overrides should be reviewed for bias. Forecasting maturity is demonstrated not by the complexity of the algorithm but by the quality of decisions made under uncertainty.

Chapter Summary

Specialized forecasting problems arise when demand is sparse, irregular, newly created, promotion-driven, portfolio-dependent, or distributed across multiple levels. Intermittent demand requires separate consideration of demand occurrence and demand size; Croston-type methods provide a useful foundation but should be evaluated against operational needs and alternative variants. New products require analogues, attribute models, market knowledge, and frequent updating, while short life-cycle items require probabilistic forecasts linked to overage and underage costs.

Promotional forecasting must distinguish baseline demand from incremental uplift, timing shifts, stock constraints, and cannibalization. Substitution and stock-outs mean that observed sales may not equal unconstrained customer demand. Hierarchical forecasting uses bottom-up, top-down, middle-out, or reconciliation methods to create coherent plans across products, locations, and channels. Across all these problems, good practice combines data diagnosis, suitable benchmarks, scenario analysis, operational metrics, human judgement, and governance. Technology and AI can strengthen this process, but they cannot compensate for poor data, unclear decisions, or unexamined assumptions.

Key Terms

Intermittent demand: Demand with many periods of zero sales and irregular non-zero occurrences.

Lumpy demand: Intermittent demand combined with substantial variability in demand size.

Croston’s method: A forecasting method that separately smooths non-zero demand size and the interval between demand occurrences.

Demand occurrence: The event in which a non-zero quantity is demanded.

Demand interval: The number of periods between successive non-zero demand events.

Syntetos–Boylan adjustment: A modification of Croston’s method intended to reduce forecast bias.

Analogue forecasting: Forecasting a new product by adapting demand evidence from comparable earlier products or situations.

Attribute-based model: A model that estimates demand from product, market, customer, or channel characteristics.

New-product forecasting: Estimating demand for an item with limited or no direct historical sales.

Short life cycle: A selling period in which a product becomes obsolete, replaced, or commercially irrelevant relatively quickly.

Baseline demand: Expected demand in the absence of a particular promotion or special intervention.

Promotional uplift: Additional demand attributed to a promotional activity above baseline demand.

Cannibalization: Reduction in one product’s demand caused by another product in the same portfolio.

Substitution: Selection or use of an alternative product when the preferred product is unavailable, less attractive, or functionally replaceable.

Demand censoring: The failure of recorded sales to reveal full demand because inventory or capacity was insufficient.

Bottom-up forecasting: Forecasting at the lowest level and aggregating forecasts upward.

Top-down forecasting: Forecasting at an aggregate level and allocating the result to lower levels.

Forecast reconciliation: Adjusting forecasts so that values at different hierarchical levels add up consistently.

Forecast fractile: The demand-distribution percentile used to determine an order quantity under uncertain demand.

Review Questions

Short Answer Questions

1. What distinguishes intermittent demand from lumpy demand?

2. Why does Croston’s method separate demand size from demand interval?

3. Define baseline demand and promotional uplift.

4. What is cannibalization in a product portfolio?

5. Why must forecasts in a hierarchy be reconciled?

Long Answer Questions

1. Explain Croston’s method, its logic, its main variants, and its limitations in inventory planning.

2. Discuss how analogues and attribute-based models can be combined to forecast a new product.

3. Explain the components of promotional sales and the consequences of confusing sales with unconstrained demand.

4. Compare bottom-up, top-down, middle-out, and reconciled hierarchical forecasting.

5. Discuss the managerial and governance requirements for using AI in specialized demand forecasting.

Analytical / Application Questions

1. A spare part has 20 months of history, with 14 zero-demand months and highly variable non-zero orders. Which diagnostic steps and forecasting method would you recommend, and why?

2. A retailer launches a premium product that reduces sales of an existing standard product while increasing total category profit. How should the forecast and performance evaluation reflect this outcome?

3. A seasonal product has a three-month selling window, uncertain demand, no replenishment opportunity, and a low salvage value. Explain how you would construct the demand distribution and determine an appropriate order quantity.

22. Multiple Choice Questions with Answers

Question 1

Which feature most directly characterizes intermittent demand?

A. A stable upward trend in every period

B. Many periods with zero demand and irregular non-zero occurrences

C. High demand during every promotional period

D. Demand that is always equal across locations

Correct Answer: B. Many periods with zero demand and irregular non-zero occurrences

Question 2

What is the principal logic behind Croston-type forecasting?

A. Treat every zero observation as a permanent decline in demand

B. Forecast only the maximum historical demand

C. Estimate non-zero demand size separately from the interval between occurrences

D. Replace all historical observations with market-share data

Correct Answer: C. Estimate non-zero demand size separately from the interval between occurrences

Question 3

A product has many zero-demand periods and highly variable quantities when demand occurs. How should it initially be classified?

A. Smooth demand

B. Seasonal demand only

C. Lumpy demand

D. Purely promotional demand

Correct Answer: C. Lumpy demand

Question 4

Which information is most useful when selecting an analogue for a new product?

A. Similarity of product name alone

B. Similarity in customer segment, price, channel, and launch conditions

C. The analogue’s highest-ever sales regardless of context

D. Only the analogue’s manufacturing cost

Correct Answer: B. Similarity in customer segment, price, channel, and launch conditions

Question 5

Which statement best describes promotional uplift?

A. All sales recorded during a promotion, including transferred demand

B. The additional demand attributed to a promotion above the estimated baseline

C. The reduction in sales after a promotion ends

D. The quantity available in inventory during a promotion

Correct Answer: B. The additional demand attributed to a promotion above the estimated baseline

Question 6

A promotion causes customers to purchase earlier than they otherwise would have, followed by unusually low sales. This is primarily an example of:

A. Timing shift or forward buying

B. Permanent market growth

C. Forecast reconciliation

D. Demand aggregation

Correct Answer: A. Timing shift or forward buying

Question 7

Why can recorded sales underestimate actual demand during a stock-out?

A. Customers always increase their orders during stock-outs

B. Observed sales are limited by available inventory

C. Stock-outs eliminate all substitution effects

D. The baseline forecast automatically increases

Correct Answer: B. Observed sales are limited by available inventory

Question 8

Which forecasting approach begins at the lowest product or location level and aggregates upward?

A. Top-down

B. Middle-out

C. Bottom-up

D. Causal allocation only

Correct Answer: C. Bottom-up

Question 9

What is the main purpose of forecast reconciliation?

A. To remove all managerial judgement

B. To ensure forecasts at different hierarchical levels are mathematically coherent

C. To make every SKU forecast identical

D. To increase promotional uplift estimates

Correct Answer: B. To ensure forecasts at different hierarchical levels are mathematically coherent

Question 10

For a short life-cycle product with no replenishment opportunity, the order quantity should primarily reflect:

A. Only the average of last year’s sales

B. The demand distribution and the relative costs of shortage and excess inventory

C. The highest possible forecast without considering salvage value

D. The lowest forecast to eliminate all obsolescence

Correct Answer: B. The demand distribution and the relative costs of shortage and excess inventory