Specialized Business Forecasting Problems

Introduction

Many forecasting methods are developed for relatively stable, regularly observed demand. In practice, however, organizations forecast products that sell only occasionally, new products with no historical observations, items whose lives last only a few weeks, demand distorted by promotions, and quantities that must agree across products, locations, and time horizons. These situations require more than selecting a standard exponential-smoothing or regression model.

Specialized forecasting problems are difficult because the data-generating process is unusual. Demand may contain many zeros, the product may have no history, observations may be counts rather than continuous quantities, or forecasts may be constrained by inventory and supply decisions. Good practice therefore begins with problem diagnosis: define the forecasting object, identify the decision it supports, understand the information available at the forecast origin, and select methods consistent with the demand process.

Learning objectives

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

1. distinguish intermittent, count, new-product, short-life-cycle, promotional, and inventory-linked forecasting problems;
2. select appropriate models and evaluation measures for specialized demand;
3. explain temporal and hierarchical forecasting and the purpose of reconciliation;
4. apply forecast combinations and interpret their advantages and limitations;
5. design a practical forecasting process that respects information timing, uncertainty, and governance.

1. Diagnosing the forecasting problem

A forecast is a conditional estimate of an uncertain future quantity. The quantity may be unit sales, orders, service arrivals, demand during lead time, or revenue. The forecast horizon, aggregation level, and decision context are as important as the statistical model.

A useful diagnostic profile includes:

• Demand frequency: How often is demand positive?
• Demand size: What is the distribution of positive quantities?
• Lead time: How long must the firm wait for replenishment?
• Product age: Is there sufficient history after launch?
• Intervention exposure: Were prices, advertising, displays, or stockouts present?
• Organizational structure: Must forecasts add across products, stores, regions, or time buckets?
• Decision loss: Is overforecasting more costly than underforecasting?

Forecast errors should be evaluated against the decision. For example, a warehouse may need the probability that demand exceeds available stock, whereas a finance team may need an unbiased monthly revenue forecast. No single accuracy measure is universally best.

2. Intermittent and count demand

2.1 Intermittent demand

Intermittent demand has many periods with zero demand and occasional positive observations. Examples include aircraft components, maintenance parts, specialist medical supplies, and slow-moving spare parts. Two features must be distinguished:

• Interarrival interval: the number of periods between positive demands;
• Demand size: the quantity demanded when demand occurs.

A series may be intermittent because demand is genuinely rare, or because the time bucket is too short. Aggregating daily demand into weekly demand can reduce zeros, although it may reduce responsiveness.

A simple measure of zero prevalence is the proportion of periods with zero demand:

Here, z is dimensionless. A high value indicates intermittency but does not by itself determine the best method. The variability of positive demand and the variability of interarrival intervals are also important.

2.2 Croston-type forecasting

A conventional moving average treats zeros and positive observations in the same sequence. Croston-type methods instead estimate the average positive demand and the average interval between positive demands.

Let y_t be demand in period t. At a positive-demand time, let q_t be the positive quantity and p_t be the interval since the previous positive demand. Exponentially smoothed estimates can be written as:

The forecast for an ordinary future period is:

where α is a smoothing parameter between 0 and 1. Updates occur when positive demand is observed; zero-demand periods generally do not update the positive-size estimate. The method assumes that positive demand sizes and intervals can be represented by separately evolving averages. It is useful as a baseline but has known bias and timing limitations, especially when demand changes rapidly.

Variants such as the Syntetos–Boylan approximation and Teunter–Syntetos–Babai methods adjust the treatment of occurrence probabilities and can perform better in some settings. These methods should be tested by rolling-origin evaluation rather than selected by reputation alone.

2.3 Count demand

Count demand records non-negative integers: numbers of calls, arrivals, claims, defects, or orders. A Poisson model is a starting point when events occur independently at a stable average rate:

Here, Y is the count in a specified interval, k is a possible count, and λ is the expected count in that interval. The Poisson model has mean and variance both equal to λ. Its rate must match the time unit: a daily rate cannot be used directly as a monthly rate without adjustment.

When the variance exceeds the mean, overdispersion is present, and a negative binomial model may be more suitable. When there are more zeros than the selected count model predicts, zero-inflated or hurdle models may be considered. These models require a substantive explanation for the extra zeros; they should not be used merely because they fit a sample better.

For operational planning, a probabilistic forecast is often more useful than a point forecast. If demand during lead time is D_L and the replenishment point is R, a stockout occurs when D_L > R. Thus the service probability is:

The required value of R depends on the desired service level and the estimated lead-time demand distribution.

Worked example: intermittent spare-part demand

Suppose the positive demand quantities observed over several months average 4 units, while positive demands occur every 3 months on average. A Croston-type point forecast is:

This is an average demand rate, not a claim that exactly 1.33 units will be ordered every month. The inventory decision should also consider the probability and size of the next positive demand, supplier lead time, and the cost of holding stock.

3. New-product forecasting

A new product has little or no direct sales history. The central difficulty is not simply parameter estimation; it is identifying the product’s likely market, adoption path, substitution effects, and operational constraints.

3.1 Sources of information

Useful evidence includes:

• analogous products, adjusted for price, market size, and channel;
• customer surveys, concept tests, and preorder data;
• expert judgment from sales, marketing, and operations;
• market-potential and adoption models;
• small-scale pilots and controlled regional launches;
• search, web, and engagement indicators, used cautiously;
• capacity, distribution, and availability assumptions.

Analogs should be selected by mechanism, not superficial similarity. A product can resemble an earlier product in design but differ greatly in price, regulation, competition, or channel.

3.2 Scenario and diffusion approaches

Scenario forecasting is often more defensible than a single precise number. A simple expected forecast is:

where s indexes scenarios, P(s) is the probability assigned to scenario s, and D_s is demand under that scenario. Probabilities must sum to 1 and should be documented as judgments or estimated quantities.

Diffusion models describe adoption over time. The Bass model is commonly expressed as:

Here, N(t) is cumulative adopters at time t, m is market potential, p is the coefficient of innovation, and q is the coefficient of imitation. The model assumes that adoption is influenced by external communication and by existing adopters. It is not a substitute for market research, and estimates can be unstable when early observations are scarce.

3.3 Launch learning and updating

A launch forecast should be treated as a living forecast. Prelaunch assumptions are replaced progressively by observed conversion, repeat purchase, returns, channel availability, and competitor responses. Forecasts should be revised using only information available at each forecast origin. Using later information in an earlier forecast evaluation creates look-ahead bias.

4. Short life cycles

Fashion items, seasonal gifts, event merchandise, and technology products may have a short selling window. The principal decision is often a one-time or limited replenishment decision. Historical data from mature products may be less valuable than information about current trend, reservations, early sales, and the remaining selling window.

Forecasts must distinguish:

• total demand over the life cycle;
• demand by remaining period;
• demand that can still be fulfilled before the product becomes obsolete.

A useful operational framework is to update the forecast of total remaining demand after each sales observation. If A_t is cumulative actual sales by time t, and F_t is the estimated total life-cycle demand, then estimated remaining demand is:

This calculation assumes that F_t is a total-demand estimate on the same product and market basis as A_t. It does not account automatically for lost sales caused by stockouts; observed sales may need correction using availability data.

Short-life-cycle forecasts should be linked to markdown and salvage decisions. If an extra unit has purchase cost c, selling price p, and salvage value v, then the critical fractile for a single-period newsvendor decision is:

where Q* is the target order quantity and the condition v < c < p is normally assumed. The expression is applicable when underage and overage costs can be represented by these prices and costs, and when demand is uncertain but the selling period is defined. If markdowns, capacity, or multiple replenishments exist, the basic model is incomplete.

5. Promotions and interventions

Promotions create a forecasting problem because demand is partly generated by the intervention. A forecast for a promotional week should not simply extrapolate regular demand. It should estimate baseline demand and incremental demand.

A basic decomposition is:

where D_t is observed demand, B_t is baseline demand without the promotion, I_t is incremental promotional demand, and ε_t is unexplained variation. The units of all four terms must be the same, such as units per store per week.

Regression with intervention variables can estimate effects while controlling for seasonality, price, distribution, and competitor activity. A simplified model is:

Here, Promo_t may be an indicator or intensity measure. Causal interpretation of β₂ requires that important confounders be controlled and that the promotion variable is not merely a proxy for an unmeasured event.

Important complications include:

• Pull-forward: customers buy early and reduce later demand;
• Cannibalization: a promoted product takes sales from another product;
• Post-promotion dip: demand falls below baseline after stockpiling;
• Execution failure: the advertised price or display is not available;
• Endogeneity: promotions are targeted to periods already expected to be weak or strong.

Promotional forecasts should therefore use event calendars, execution data, and uplift evaluation. A model trained on promotional sales without promotion indicators may overforecast ordinary demand.

6. Inventory-linked forecasts

Observed sales are not always demand. If stock is unavailable, the recorded sales quantity is censored below true customer demand. Let S_t be observed sales and D_t be demand. In a simple stockout situation:

where A_t is available inventory. The identity assumes that all demand exceeding availability is lost and that there are no substitutions, backorders, or rationing effects. In real systems, these assumptions must be checked.

Forecasting from censored sales can systematically understate demand for popular products. Analysts should incorporate inventory-on-hand, availability hours, backorders, substitutions, and lost-sales indicators. Possible approaches include:

1. remove periods with severe stockouts from model training;
2. estimate unconstrained demand using censored-data methods;
3. use substitution and traffic data as supplementary evidence;
4. forecast demand and supply separately, then simulate inventory outcomes.

A forecast is also affected by replenishment policy. If orders are generated from a forecast, and future sales depend on whether stock was available, a feedback loop exists. Model monitoring must distinguish forecast error from supply execution error.

7. Temporal and hierarchical forecasting

7.1 Temporal aggregation

A business may need forecasts hourly, daily, weekly, and monthly. Forecasts generated independently at each frequency may disagree. For example, twelve monthly forecasts may not sum to the annual forecast. Temporal coherence means that forecasts at fine time scales aggregate to forecasts at coarser time scales.

If y is a vector of bottom-level time observations and Y is a vector containing observations at several temporal levels, an aggregation matrix S gives:

The matrix S contains zeros and ones, or other aggregation weights where appropriate. It records which lower-level periods contribute to each higher-level total. Temporal reconciliation adjusts forecasts so that the same relationship holds for forecasts.

7.2 Hierarchical forecasting

A hierarchy may contain total company demand, regional demand, store demand, and product-store demand. Bottom-up forecasting models the lowest level and sums upward. Top-down forecasting begins with an aggregate forecast and allocates it using proportions. Middle-out forecasting forecasts selected intermediate levels and aggregates or disaggregates from them.

Each method has advantages:

A coherent set of forecasts can be expressed as:

where ĝ is a vector of forecasts or reconciled components and S maps components into all required levels. Reconciliation methods such as ordinary least squares, weighted least squares, and MinT seek forecasts that are coherent while considering forecast-error variances and covariances. In practice, covariance estimates can be unstable in large hierarchies, so shrinkage and robust estimation are important.

8. Forecast reconciliation in practice

Suppose a retailer requires a company forecast and forecasts for two regions. The accounting identity is:

An independent process might produce a company forecast of 1,000 units, regional forecasts of 620 and 450 units, and therefore a regional total of 1,070. These forecasts cannot all be used directly in a plan. Reconciliation modifies them according to a chosen objective, such as minimizing squared adjustment subject to the identity.

Reconciliation is not merely an arithmetic exercise. The analyst must decide:

• which forecasts are generated independently;
• which constraints are mandatory;
• whether negative forecasts are possible or must be prevented;
• how uncertainty is estimated;
• whether accuracy is measured before or after reconciliation;
• how changes are communicated to planners.

Non-negativity, capacity, and supply constraints may require optimization beyond standard linear reconciliation. A coherent forecast can still be inaccurate if all component models are biased.

9. Forecast combinations

Forecast combinations average or otherwise combine forecasts from different models, data sources, or experts. If forecasts are f₁, f₂, …, f_K and weights are w₁, w₂, …, w_K, a linear combination is:

For a convex combination, w_k ≥ 0 and ∑w_k = 1. Equal weighting is often a strong baseline because estimated weights can overfit, particularly when the historical sample is short. Weighted combinations may use validation accuracy, but weights must be estimated using information available at the forecast origin.

Combination works especially well when component forecasts have different errors. If errors are positively correlated, averaging provides less diversification. For two unbiased forecasts with error variances σ₁² and σ₂² and error covariance σ₁₂, the variance of a weighted combination is:

Here, w is the weight on the first forecast. The formula assumes finite variances and a linear combination. Lower covariance generally increases the potential benefit of combination, but combination does not guarantee improvement.

Useful component forecasts may include a seasonal naive forecast, an exponential-smoothing forecast, a causal promotion model, a machine-learning model, and a judgmental adjustment. Their data definitions must be consistent. Combining forecasts made for different horizons, product scopes, or promotion assumptions is invalid unless they are first made comparable.

10. Integrated applied case: illustrative omnichannel retailer

Background

Consider an illustrative retailer, Northstar Home, selling household appliances through stores and an online channel. It must forecast a newly launched compact air purifier, replacement filters, and a promotional bundle. Forecasts are required by product, channel, region, week, and month.

Problem

The air purifier has no direct history. Filter demand is intermittent. The launch promotion creates a temporary sales spike. Several stores stocked out, so observed sales understate demand. Independently generated weekly and monthly forecasts do not add to the annual plan.

Analysis

The analytics team first separates the demand processes. It uses analogous-product data and preorder information for the purifier, Croston-type and count-based methods for filters, and a baseline-plus-uplift model for the promotion. Inventory availability is joined to transactions so that stockout periods are flagged. The team also checks whether online purchases substituted for unavailable store purchases.

For the purifier, three scenarios are constructed: conservative, expected, and strong adoption. Probabilities are documented with marketing and sales leadership rather than hidden inside a model. Early launch observations are incorporated through rolling-origin updates.

For filters, positive demand size and interarrival intervals are modeled separately. The replenishment calculation uses a lead-time demand distribution rather than relying only on the average monthly forecast.

For the promotion, the team estimates regular demand from comparable nonpromotional periods and models incremental units using price, display, and channel-execution variables. A post-promotion dip is included in the planning calendar.

Finally, forecasts are reconciled across product, channel, region, and time. A weighted reconciliation gives greater reliability to aggregate forecasts where bottom-level series are sparse, while retaining detailed signals where sufficient data exist. A simple equal-weight combination of statistical and causal forecasts is retained as a benchmark.

Outcome

The resulting plan does not claim that one model is universally correct. It produces coherent forecasts, identifies stockout-adjusted demand, separates baseline from promotional uplift, and attaches scenarios to the new product. Performance is evaluated by decision-relevant measures: unit error, service-level attainment, waste, and forecast bias by product type.

Learning points

1. Forecast design follows the demand process and decision, not the other way around.
2. Sales data must be interpreted in light of availability.
3. New products require structured uncertainty and rapid updating.
4. Promotional spikes should not be treated as ordinary recurring demand.
5. Coherence across time and organizational levels is a practical planning requirement.
6. Combining diverse, well-defined forecasts can improve robustness.

11. Evaluation, governance, and implementation

Specialized forecasts require specialized evaluation. For intermittent data, percentage errors can be undefined or misleading when actual demand is zero. Alternatives include mean absolute error, scaled errors, service-level measures, and measures evaluated only on positive-demand events, with the choice explicitly documented. For count forecasts, assess both count accuracy and calibration of predictive probabilities.

Rolling-origin evaluation is preferred: repeatedly train using information available up to a historical date, forecast the next required horizon, and compare with what actually occurred. The evaluation design must reproduce the operational horizon, product availability, promotion information, and revision schedule.

Governance practices include:

• maintaining a data dictionary for demand, sales, returns, stockouts, and promotions;
• recording forecast versions and the information set used;
• distinguishing statistical forecasts from approved judgmental overrides;
• monitoring bias, drift, calibration, and subgroup performance;
• setting approval thresholds for material overrides;
• protecting commercially sensitive and customer-level information.

AI and machine learning can support feature generation, analog selection, probabilistic forecasting, and anomaly detection. They do not remove the need for causal reasoning, coherent aggregation, or inventory knowledge. Black-box systems can reproduce stockout bias or promotional confounding. Human review should focus on assumptions, data quality, uncertainty, and decision consequences rather than simply accepting a model score.

Best-practice workflow

1. Define the target, horizon, frequency, hierarchy, and decision.
2. Diagnose zeros, counts, product age, promotions, and availability constraints.
3. Establish transparent baselines, including seasonal naive and simple intermittent methods.
4. Add causal, scenario, or diffusion information only when it is available at forecast time.
5. Generate probabilistic information where service, capacity, or risk decisions require it.
6. Reconcile forecasts across time and organizational levels.
7. Combine models when their errors or information sources are meaningfully different.
8. Evaluate through rolling-origin backtesting and decision-relevant metrics.
9. Monitor forecast bias, data changes, overrides, and operational outcomes.
10. Retire methods that no longer improve decisions, even if they are technically sophisticated.

Chapter Summary

Specialized business forecasting problems arise when ordinary historical extrapolation is unreliable. Intermittent demand should be analyzed through positive-demand sizes and interarrival intervals; count demand may require Poisson, negative binomial, hurdle, or zero-inflated thinking. New products depend on analogs, scenarios, market evidence, diffusion assumptions, and rapid postlaunch updating. Short-life-cycle forecasting must estimate remaining demand and connect forecasts to markdown, salvage, and one-period inventory decisions.

Promotional forecasting separates baseline from incremental demand and considers pull-forward, cannibalization, execution, and post-promotion effects. Inventory-linked forecasting recognizes that observed sales may be censored by stockouts. Temporal and hierarchical forecasting require coherent aggregation, while reconciliation adjusts forecasts to satisfy structural relationships. Forecast combinations can improve robustness when component forecasts contain complementary information and are evaluated without leakage. Across all problems, sound data definitions, rolling-origin evaluation, uncertainty communication, and governance are as important as model selection.

Key Terms

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

Count demand: Demand recorded as non-negative integer events or units.

Interarrival interval: Time between successive positive-demand observations.

Croston-type method: A forecasting approach that separately estimates positive demand size and occurrence interval.

Overdispersion: Count-data variation greater than the mean implied by a Poisson model.

New-product forecast: A forecast made when direct product history is limited or absent.

Analog product: An earlier product used as evidence for a new product forecast.

Diffusion model: A model describing adoption or cumulative market penetration over time.

Short life cycle: A selling period in which demand and product relevance end quickly.

Baseline demand: Expected demand without a specified intervention.

Incremental demand: Additional demand attributed to an intervention such as a promotion.

Pull-forward: Purchases shifted from future periods into a promotional period.

Stockout censoring: The condition in which observed sales are limited by product availability.

Temporal coherence: Agreement between forecasts at different time aggregations.

Hierarchical forecasting: Forecasting across related organizational or product levels.

Forecast reconciliation: Adjusting forecasts so that structural aggregation constraints hold.

Forecast combination: Forming one forecast from multiple component forecasts.

Rolling-origin evaluation: Repeated historical testing that preserves the timing of information availability.

Forecast bias: Systematic tendency to overforecast or underforecast.

Service level: The probability or proportion of demand fulfilled under a specified inventory policy.

Review Questions

Short Answer Questions

1. Why is intermittent demand better described through demand size and interarrival interval than through a single average?
2. What is the distinction between baseline and incremental promotional demand?
3. Why can observed sales understate demand during a stockout?
4. Define temporal coherence.
5. State one advantage and one limitation of forecast combinations.

Long Answer Questions

1. Explain Croston-type forecasting, including its assumptions and limitations.
2. Discuss a rigorous forecasting process for a new product with no direct sales history.
3. Compare bottom-up, top-down, middle-out, and reconciliation approaches to hierarchical forecasting.
4. Explain how promotions can create pull-forward, cannibalization, and post-promotion effects.
5. Design an evaluation and governance framework for specialized business forecasts.

Analytical / Application Questions

1. A spare part has an average positive demand of 6 units and an average interval of 4 months. Calculate a Croston-type monthly point forecast and explain why it should not be interpreted as a fixed monthly order.
2. A retailer forecasts 800 units of baseline demand and expects a promotion to add 25% incremental demand. Calculate the total forecast and identify two assumptions that should be tested.
3. A company forecast is 1,200 units, while two regional forecasts are 700 and 650 units. Explain why reconciliation is necessary and describe two possible reconciliation strategies.

Multiple Choice Questions with Answers

Question 1

Which characteristic most directly defines intermittent demand?

A. Demand has a constant positive value every period

B. Demand contains many zero periods and occasional positive observations

C. Demand is always normally distributed

D. Demand is observed only after a promotion

Correct Answer: B. Demand contains many zero periods and occasional positive observations

Question 2

In a Croston-type method, what is estimated separately from positive demand size?

A. The tax rate

B. The market potential

C. The interval between positive demands

D. The inventory carrying cost

Correct Answer: C. The interval between positive demands

Question 3

If a count process has variance substantially greater than its mean, which issue is suggested?

A. Overdispersion

B. Perfect seasonality

C. Temporal coherence

D. Negative demand

Correct Answer: A. Overdispersion

Question 4

What is the main purpose of a baseline-plus-uplift promotional model?

A. To remove all seasonality from the data

B. To distinguish ordinary demand from intervention-generated demand

C. To force all forecasts to be equal

D. To replace inventory records

Correct Answer: B. To distinguish ordinary demand from intervention-generated demand

Question 5

When a product is out of stock, recorded sales are often best described as:

A. A guaranteed measure of unconstrained demand

B. A lower bound on potential demand

C. A measure of competitor demand

D. A forecast of future demand

Correct Answer: B. A lower bound on potential demand

Question 6

A hierarchy is temporally coherent when:

A. All forecasts use the same algorithm

B. Every forecast is identical

C. Fine-period forecasts aggregate to the corresponding coarse-period forecasts

D. Only bottom-level forecasts are produced

Correct Answer: C. Fine-period forecasts aggregate to the corresponding coarse-period forecasts

Question 7

Which approach begins by forecasting the lowest organizational or product level and then aggregates upward?

A. Bottom-up

B. Top-down

C. Judgmental override

D. Scenario diffusion

Correct Answer: A. Bottom-up

Question 8

If baseline demand is 800 units and incremental promotional demand is 25% of baseline, what is the total forecast?

A. 825 units

B. 1,000 units

C. 1,025 units

D. 1,200 units

Correct Answer: B. 1,000 units

Question 9

Why can combining forecasts improve accuracy?

A. It guarantees that every component model is unbiased

B. It can diversify errors when component forecasts contain complementary information

C. It eliminates the need for evaluation

D. It makes all demand deterministic

Correct Answer: B. It can diversify errors when component forecasts contain complementary information

Question 10

Which evaluation design best avoids look-ahead bias?

A. Using future observations to select past model parameters

B. Evaluating only the largest sales period

C. Rolling-origin evaluation using information available at each forecast date

D. Replacing missing future demand with zero

Correct Answer: C. Rolling-origin evaluation using information available at each forecast date