Specialized Business Time Series

Introduction

Many business forecasting systems are designed around relatively regular observations: sales occur every period, variability is moderate, and historical patterns resemble future conditions. Important commercial series frequently violate these assumptions. Spare-parts demand may be zero for weeks and then suddenly large. Promotional sales may be structurally different from ordinary sales. A new product has little or no sales history. Daily electricity demand can contain intraday, weekly, and annual cycles. Financial returns may exhibit volatility clustering, while economic and operational data may shift between distinct regimes.

These situations require models that match the data-generating process rather than merely applying a standard ARIMA or exponential-smoothing model. This chapter examines specialized approaches for:

• intermittent and lumpy demand;
• sparse count data;
• promotional and event-driven demand;
• new-product cold starts;
• multiple seasonalities;
• volatility clustering; and
• structural breaks and regime changes.

The central principle is that forecast design must distinguish between the business process generating observations and the statistical appearance of the series. A zero in a demand series may mean no customer need, no stock availability, or no observation. A sudden increase may represent a promotion, a permanent level shift, or random variation. Correct diagnosis is therefore as important as model selection.

Learning Objectives

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

1. distinguish intermittent, lumpy, sparse-count, promotional, volatile, and regime-switching series;
2. select and implement Croston-family methods appropriately;
3. explain Poisson, negative-binomial, zero-inflated, and hurdle count models;
4. construct forecasting strategies for promotions and new products;
5. model multiple seasonalities using decomposition, dynamic regression, and state-space methods;
6. explain ARCH and GARCH volatility models and their limitations;
7. identify structural breaks and estimate regime-switching models;
8. evaluate forecasts using metrics appropriate to the business decision; and
9. translate specialized forecasts into inventory, pricing, risk, and managerial decisions.

1. Diagnosing Specialized Business Series

A useful first step is to classify the observation process. Let y_t denote demand or another business outcome at time t. A demand series is often called intermittent when many periods contain zero demand. It is lumpy when demand is intermittent and the non-zero quantities vary substantially. A series is sparse when events are rare relative to the observation frequency, including low-frequency claims, failures, or transactions.

Two simple descriptive measures are especially useful. If N is the number of observations and N_+ is the number of positive observations, the proportion of zero observations is

The mean inter-demand interval is commonly estimated as

where ADI is measured in periods per positive demand occurrence. Let the positive demand sizes be z₁, z₂, …, z_{N_+}. Their coefficient of variation is

where \bar{z} is the mean positive demand and s_z is its sample standard deviation. High ADI indicates infrequency; high CV indicates irregular size. These measures are diagnostic summaries, not universal classification laws. Thresholds should be validated against the operational consequences of forecast error.

The analyst should also ask:

• Are zeros genuine, or do they reflect stockouts and missing data?
• Is the time unit appropriate, or are events being aggregated too coarsely?
• Are calendar, price, promotion, weather, or supply variables available?
• Does the series represent a count, a continuous quantity, a rate, or a financial return?
• Did the measurement process change?

Data cleaning is particularly important. Treating stockout periods as zero demand systematically underestimates underlying demand. Conversely, imputing demand during every zero period can create artificial sales.

2. Intermittent and Lumpy Demand

2.1 Why ordinary models struggle

A conventional model fitted directly to an intermittent series may produce forecasts close to zero most of the time and respond too slowly after a positive observation. Standard error metrics can also be misleading: a forecast of zero may achieve a good average accuracy when demand is sparse but provide poor replenishment decisions.

A useful conceptual representation separates demand occurrence from demand size. Define an occurrence indicator I_t, where I_t = 1 if demand is positive and I_t = 0 otherwise. Then demand can be expressed as

where z_t > 0 is the positive demand size. This decomposition motivates Croston-family methods and modern two-part models.

2.2 Croston’s method

Croston’s method separately smooths positive demand size and the interval between positive demands. At a positive-demand time, let q_t be the observed positive quantity and τ_t the interval since the previous positive demand. With smoothing parameter α, where 0 < α < 1, the updates are

The forecast for demand per period is

for a one-period rate forecast, often held constant over the next h periods. Between positive-demand observations, the estimates are not updated in the original method. The method assumes that positive sizes and inter-arrival intervals can be represented by separately smoothed quantities and that recent history is informative about future demand.

2.3 Variants and bias considerations

Several variants address known properties of the original method:

• Syntetos–Boylan Approximation (SBA): applies a bias correction to the Croston forecast, commonly written as

under the standard SBA convention.

• Teunter–Syntetos–Babai (TSB): updates the probability of demand occurrence each period, including zero-demand periods. If p_t is the smoothed occurrence probability and q_t is the current positive size when demand occurs, then conceptually

TSB is often more responsive when demand may become obsolete because the occurrence probability decays during consecutive zeros.

• Modified Croston methods: alter the interval definition or correction factor. Their performance depends on the data-generating process and implementation details.

No Croston variant is universally superior. Method choice should be based on rolling-origin validation and the cost of over- and under-forecasting.

2.4 Worked example: Croston calculation

Suppose positive demands occur at periods 2, 5, and 9, with sizes 4, 6, and 5. The intervals are 2, 3, and 4 periods. Use α = 0.2, initial smoothed size 4, and initial smoothed interval 2.

At the second positive observation, the updates are:

\hat{z} = 0.2(6) + 0.8(4) = 4.4

\hat{τ} = 0.2(3) + 0.8(2) = 2.2

At the third positive observation:

\hat{z} = 0.2(5) + 0.8(4.4) = 4.52

\hat{τ} = 0.2(4) + 0.8(2.2) = 2.56

The Croston rate forecast is therefore 4.52 ÷ 2.56 ≈ 1.77 units per period. The SBA adjustment gives (1 − 0.2/2)(1.77) ≈ 1.59 units per period. These are average-rate forecasts, not predictions that 1.59 or 1.77 units will necessarily be sold in every period.

2.5 Evaluation and inventory implications

Intermittent forecasts should be evaluated at the decision level. Useful measures include mean absolute scaled error, cumulative bias, service level, fill rate, stockout frequency, and inventory holding cost. Mean absolute percentage error is unsuitable when actual demand is frequently zero. Quantile forecasts may be more useful than point forecasts because replenishment decisions depend on uncertainty.

Croston-family methods forecast demand rates but do not automatically model lead-time demand distribution. Inventory systems may require a separate model for demand during lead time, safety stock, and order quantities. A simple rate forecast should not be interpreted as a complete inventory policy.

3. Sparse Counts and Event Demand

3.1 Poisson and negative-binomial models

For count data, a natural starting point is the Poisson distribution. If Y_t is a non-negative integer count and its conditional mean is λ_t, then

The Poisson assumption implies E(Y_t) = Var(Y_t) = λ_t. This equidispersion condition is often unrealistic. When the variance exceeds the mean, a negative-binomial model introduces an additional dispersion parameter. One common parameterization has

where μ_t is the mean and κ > 0 controls overdispersion. As κ becomes large, the variance approaches the Poisson variance.

A regression model commonly uses a log link:

where x_t contains predictors such as price, exposure, day-of-week indicators, or marketing activity, and β is a coefficient vector. The log link ensures a positive conditional mean.

3.2 Zero-inflated and hurdle models

Excess zeros may arise from two processes: a structural state in which an event cannot occur, and a sampling state in which an event could occur but does not. A zero-inflated model combines a structural-zero probability with a count distribution. If π_t is the structural-zero probability and f(k; μ_t) is a Poisson or negative-binomial probability, then

A hurdle model instead uses one model to distinguish zero from positive demand and a zero-truncated count distribution for positive observations. Zero-inflated models are appropriate when structural and sampling zeros are substantively distinct; hurdle models are often easier to interpret when every positive observation is generated by a separate process.

3.3 Dependence and exposure

Counts often depend on exposure. For example, the number of claims depends partly on the number of insured policies. An offset can be used:

where e_t is known exposure. The term log(e_t) has coefficient one and converts the model into a rate model. Ignoring exposure can create spurious trends.

4. Promotions, Events, and Causal Forecasting

Promotional demand is not simply an outlier to be smoothed away. A promotion may cause a temporary lift, a timing shift, cannibalization of related products, or post-promotion demand depression. A dynamic regression can represent these effects:

where n_t is a residual time series, potentially modeled by ARIMA or another error process. For a promotion with delayed and decaying effects, distributed lags may be used:

The coefficients measure associations or causal effects only when the design supports causal interpretation. Promotions are often targeted to expected high-demand periods, so naive comparisons can be confounded.

Important practical features include:

• regular price and discount depth;
• display and advertising support;
• promotion type and duration;
• competitor actions;
• inventory availability;
• product substitution and cannibalization; and
• post-promotion effects.

Forecasting should distinguish a baseline forecast without promotion from an incremental lift forecast. The total forecast is conceptually baseline plus incremental effect, but uncertainty in both components must be propagated. Leakage must be avoided: only promotion information known at forecast origin may be used.

5. New-Product Cold Starts

A new product lacks its own history, so conventional univariate forecasting is impossible or unreliable. Cold-start forecasting combines information from analogous products, product attributes, market assumptions, and early observations.

A practical hierarchy is:

1. identify comparable products by category, price, channel, launch type, and customer segment;
2. normalize historical sales for distribution, availability, and market size;
3. estimate a launch profile, such as trial, growth, peak, and decline;
4. update the profile as early sales arrive; and
5. reconcile product forecasts with category and channel totals.

Hierarchical or global machine-learning models can pool information across products. A global model estimates common relationships while allowing product-specific effects. Bayesian approaches are especially useful because prior information from analogous products can be updated with early evidence. However, similarity must be measured using commercially meaningful attributes rather than solely historical correlation.

Cold-start forecasts should be scenario-based. A base, high, and low scenario can reflect uncertain distribution, conversion, repeat purchase, and competitor response. Decision makers should see the assumptions and prediction intervals rather than a single apparently precise number.

6. Multiple Seasonalities

Business data may contain several cycles simultaneously. Electricity demand can have hourly, daily, weekly, and annual patterns. Online traffic may show intraday and weekly cycles. A single seasonal period is inadequate when seasonal frequencies are incommensurate or long.

A general decomposition is

for an additive structure, where T_t is trend, S_j,t are seasonal components, and ε_t is irregular error. A multiplicative form may be more appropriate when seasonal amplitude grows with the level, often after a logarithmic transformation.

Approaches include:

• Fourier terms: represent seasonality with sine and cosine pairs. For seasonal period m, harmonic k contributes

Fourier terms are compact for long or multiple seasonal cycles.

• TBATS and related state-space methods: accommodate multiple seasonal periods, Box–Cox transformation, trend, and damped components.

• Dynamic harmonic regression: combines Fourier terms with regression effects and autocorrelated errors.

• Intraday profiles and hierarchical models: useful when seasonal patterns are stable and operationally interpretable.

The number of harmonics controls flexibility. Too few terms underfit sharp peaks; too many overfit noise. Selection should use rolling validation and consider computation, interpretability, and forecast horizon.

7. Volatility Clustering and GARCH Models

In financial and operational risk data, the level may be difficult to predict while the conditional variance is predictable. Large changes tend to be followed by large changes, and small changes by small changes, regardless of direction. This is volatility clustering.

Let r_t be a return or mean-adjusted series and ε_t = σ_t z_t, where z_t has mean zero and variance one. A GARCH(1,1) model is

where ω > 0, α ≥ 0, and β ≥ 0. The conditional variance σ_t² is measured in squared units. The usual covariance-stationarity condition is α + β < 1. Under this condition, the unconditional variance is

provided the model assumptions hold. A high value of α + β indicates persistent volatility, but it is not itself a complete measure of risk.

GARCH assumes that positive and negative shocks have symmetric effects on variance. When bad news increases volatility more than good news, an asymmetric model such as GJR-GARCH or EGARCH may be more appropriate. Heavy-tailed innovations, such as Student’s t errors, may improve risk forecasts, but distributional assumptions must be checked.

GARCH forecasts conditional variance, not necessarily the direction of returns. Applications include value-at-risk, margin setting, capacity buffers, option pricing inputs, and financial stress monitoring. Backtesting is essential: a model that produces attractive in-sample likelihood may fail during crises.

8. Change-Points and Regime Changes

8.1 Structural breaks

A structural break occurs when parameters or the data-generating mechanism changes. Examples include a competitor entry, a new regulation, a pandemic, a pricing-policy change, or a technology migration. A model estimated across incompatible regimes can produce biased forecasts.

For a mean shift, a simple model is

where τ is the unknown change point. Multiple-break methods, cumulative-sum tests, binary segmentation, and penalized likelihood procedures can identify candidate changes. Detection is not proof of a business cause; operational investigation is required.

A rolling model may adapt after a break, but it discards useful older information and can be unstable. Intervention variables are appropriate when the timing and cause are known. Forecast combinations across stable and adaptive models can reduce model-risk exposure.

8.2 Markov-switching models

A regime-switching model assumes an unobserved state S_t that follows a Markov chain. For two regimes,

where p_ij is the transition probability from regime i to regime j, and each row of the transition matrix sums to one. The observed process may be

so that regimes can differ in level, persistence, variance, or regression coefficients. The expected duration of regime i is

when p_ii < 1. This formula follows from the geometric duration assumption of a first-order Markov chain.

Regime models are useful when states are persistent and substantively meaningful, such as expansion/recession or normal/stress demand. They can be difficult to estimate, sensitive to initialization, and prone to label switching. A statistically detected regime should not be named “high demand” or “crisis” without examining its economic meaning.

9. Integrated Modeling and Forecast Evaluation

Specialized problems frequently overlap. A retailer may have intermittent demand, promotions, multiple seasonalities, and stockouts. A financial institution may need a count model for defaults and a GARCH model for market volatility. Modular architectures are often preferable: model occurrence, size, promotion effects, and residual dependence separately, then combine them coherently.

Evaluation should imitate deployment. Use rolling-origin or time-series cross-validation, preserve temporal order, and evaluate several forecast horizons. Select metrics according to the decision:

For probabilistic forecasts, calibration asks whether events occur with the probabilities stated. Sharpness asks whether forecasts are informative rather than excessively wide. A forecast can be accurate on average but poorly calibrated in the tails.

10. Illustrative Case: Regional Service-Parts Forecasting

Background

A fictional industrial-equipment company supplies replacement pumps to 40 service depots. Several pump components have zero demand in most weeks, while a failure can generate an order for multiple units. Marketing occasionally offers maintenance promotions, and a new pump model is being launched.

Problem

The existing system uses moving averages. It treats stockout weeks as zero demand, ignores promotions, and evaluates accuracy with MAPE. Depots report excess stock for some parts but repeated emergency shipments for others.

Analysis

The analytics team first reconstructs demand from shipment, inventory, and back-order records. It distinguishes genuine zero demand from periods in which demand was censored by stockouts. ADI and positive-demand CV identify intermittent and lumpy components. For rare failure events, negative-binomial models are compared with Poisson models. Promotion indicators and maintenance-contract counts are added as regressors.

Approach and solution

For stable intermittent parts, the team compares Croston, SBA, and TSB using rolling-origin validation. TSB is selected for parts showing gradual obsolescence because its occurrence probability declines during extended zero periods. For high-value parts, a two-stage model forecasts occurrence probability and positive quantity separately, then produces prediction intervals for lead-time demand.

Promotion forecasts are split into baseline and incremental lift. A dynamic regression includes discount depth, campaign duration, depot, and lagged promotion indicators. The new pump uses analogues selected by engineering family and customer segment; a Bayesian update revises the launch profile after the first six weeks.

Depot-level forecasts are reconciled to regional totals. Inventory decisions use service-level targets rather than point forecasts alone. A monitoring dashboard tracks forecast bias, stockout-censored observations, data drift, promotion effects, and changes in inter-demand intervals.

Outcome

In this illustrative scenario, the company reduces emergency shipments without applying a single model to every part. The main improvement comes not from algorithmic complexity alone but from correcting the observation process, separating demand occurrence from size, and linking evaluation to inventory decisions.

Learning points

1. A zero observation is a business event requiring interpretation.
2. Forecasting and inventory policy should be evaluated together.
3. Promotions and launches require explanatory information, not only historical demand.
4. Model portfolios are often more robust than a universal forecasting method.
5. Monitoring is essential because intermittency, product life cycle, and regimes can change.

11. Limitations, Governance, and Contemporary Practice

Specialized models can create false confidence. Croston-family methods may not represent covariates or serial dependence adequately. Count models can fail under unobserved heterogeneity. GARCH forecasts are vulnerable to structural breaks and tail misspecification. Regime models may identify statistical states that lack managerial meaning. Machine-learning models can improve nonlinear prediction but require sufficient comparable data and careful leakage control.

Governance should address:

• traceable data definitions and stockout treatment;
• versioned model code and reproducible training sets;
• documented assumptions and forecast horizons;
• bias and service-level monitoring across regions or customers;
• human review for promotions, launches, and exceptional events;
• access controls for commercially sensitive and financial data; and
• escalation rules when drift or regime change is detected.

AI and machine learning are most useful when integrated with domain structure. Global models, gradient boosting, and neural forecasting can pool information across many related series, while probabilistic and hierarchical methods preserve uncertainty and aggregation consistency. They do not eliminate the need for causal reasoning, data-quality assessment, or operational validation.

Best-Practice Workflow

1. Define the decision, horizon, and cost of errors.
2. Audit observation, stockout, exposure, and calendar data.
3. Diagnose zeros, counts, seasonality, volatility, and breaks.
4. Establish simple benchmarks, including seasonal-naive and zero-demand baselines where appropriate.
5. Select candidate specialized models based on the diagnosis.
6. Use rolling-origin validation and decision-relevant metrics.
7. Produce probabilistic forecasts when uncertainty affects the decision.
8. Integrate forecasts with inventory, pricing, staffing, or risk policies.
9. Monitor drift, calibration, bias, and post-event performance.
10. Re-estimate or redesign the model when the business process changes.

Chapter Summary

Specialized business time series require models aligned with their observation and decision processes. Intermittent and lumpy demand can be decomposed into occurrence intervals and positive sizes; Croston, SBA, and TSB methods provide useful rate forecasts, with TSB particularly responsive to changing demand occurrence. Sparse counts are modeled using Poisson, negative-binomial, zero-inflated, and hurdle frameworks, with exposure and overdispersion treated explicitly. Promotions require baseline–incremental decomposition and careful causal interpretation. New products require analogues, hierarchical pooling, priors, scenarios, and rapid updating. Multiple seasonalities can be represented with Fourier terms, dynamic harmonic regression, and state-space approaches. GARCH models conditional variance and volatility persistence, while change-point and Markov-switching methods address structural instability and latent regimes. Across all applications, data diagnosis, rolling validation, uncertainty quantification, governance, and decision-linked evaluation are as important as model selection.

Key Terms

Intermittent demand: Demand with many zero observations.

Lumpy demand: Intermittent demand with highly variable positive quantities.

ADI: Average demand interval between positive-demand occurrences.

Croston’s method: A method that separately smooths positive demand size and inter-demand interval.

SBA: Syntetos–Boylan Approximation, a bias-adjusted Croston variant.

TSB: Teunter–Syntetos–Babai method that updates demand occurrence probability each period.

Sparse count: A low-frequency non-negative integer event series.

Overdispersion: Count variance exceeding its mean.

Zero-inflation: Excess zeros generated by a structural-zero process.

Hurdle model: A two-part model separating zero occurrence from positive counts.

Exposure: Population or opportunity size used to model a count rate.

Promotion lift: Incremental demand attributable to a promotional activity.

Cold start: Forecasting a product or entity with little or no own history.

Multiple seasonality: The presence of two or more recurring seasonal cycles.

Fourier terms: Sine and cosine functions used to represent periodic patterns.

Volatility clustering: Persistence in the magnitude of shocks over time.

GARCH: A model in which conditional variance depends on past squared shocks and past variance.

Change point: A time at which a statistical parameter or process changes.

Regime switching: A model in which observations depend on an unobserved state that evolves over time.

Calibration: Agreement between forecast probabilities and observed frequencies.

Review Questions

Short Answer Questions

1. Distinguish intermittent demand from lumpy demand.
2. Why can MAPE be inappropriate for a series containing many zeros?
3. What is the main difference between Croston’s method and TSB?
4. State the equidispersion assumption of the Poisson model.
5. What does a GARCH model forecast?

Long Answer Questions

1. Explain the assumptions, calculations, and limitations of Croston-family methods.
2. Compare Poisson, negative-binomial, zero-inflated, and hurdle models for sparse business counts.
3. Discuss how promotions should be incorporated into a business forecasting system.
4. Explain how Fourier terms and dynamic harmonic regression address multiple seasonalities.
5. Compare change-point analysis with Markov-switching models for structural instability.

Analytical / Application Questions

1. A spare part has 85% zero-demand weeks and highly variable positive quantities. Design a forecasting and evaluation strategy, including inventory-relevant metrics.
2. A retailer launches a product with no historical sales but has data for five analogous products. Propose a cold-start forecasting procedure and identify key uncertainties.
3. A financial return series has persistent volatility and a major regulatory break. Explain how you would combine GARCH diagnostics with break detection and how you would validate the resulting risk forecasts.

Multiple Choice Questions with Answers

Question 1

Which characteristic most directly indicates intermittent demand?

A. A high proportion of zero observations

B. Constant positive demand every period

C. Normally distributed measurement errors

D. A deterministic upward trend

Correct Answer: A. A high proportion of zero observations

Question 2

In Croston’s method, what two quantities are smoothed separately?

A. Trend and seasonal amplitude

B. Positive demand size and inter-demand interval

C. Price and promotion depth

D. Variance and covariance

Correct Answer: B. Positive demand size and inter-demand interval

Question 3

Using the SBA convention with α = 0.2 and a Croston forecast of 1.77 units per period, the adjusted forecast is approximately:

A. 0.89 units

B. 1.59 units

C. 1.77 units

D. 2.21 units

Correct Answer: B. 1.59 units

Question 4

Which condition characterizes the standard Poisson model?

A. Variance is always less than the mean

B. Variance equals the conditional mean

C. All observations must be positive

D. The mean must be zero

Correct Answer: B. Variance equals the conditional mean

Question 5

A negative-binomial model is especially useful when count data exhibit:

A. Overdispersion

B. No observations

C. Perfect seasonality only

D. A continuous response variable

Correct Answer: A. Overdispersion

Question 6

What is the purpose of an exposure offset in a count model?

A. To remove all zeros

B. To convert a count model into a rate model relative to opportunity size

C. To force equal variance and mean

D. To estimate seasonal period automatically

Correct Answer: B. To convert a count model into a rate model relative to opportunity size

Question 7

In a GARCH(1,1) model, which terms determine current conditional variance?

A. Only the current mean

B. Past squared shocks and past conditional variance

C. Only future observations

D. Only seasonal indicators

Correct Answer: B. Past squared shocks and past conditional variance

Question 8

Under the usual covariance-stationarity condition for GARCH(1,1), which inequality is required?

A. α + β > 1

B. α + β = 2

C. α + β < 1

D. α − β < 0 only

Correct Answer: C. α + β < 1

Question 9

If a two-state Markov regime has a stay probability p_ii = 0.8, its expected duration is:

A. 0.2 periods

B. 1.25 periods

C. 4 periods

D. 8 periods

Correct Answer: C. 4 periods

Question 10

Why should promotion forecasts distinguish baseline demand from incremental lift?

A. Because promotions can alter demand through temporary, delayed, or substitution effects

B. Because promotions never affect demand levels

C. Because baseline demand is always zero

D. Because count models cannot include predictors

Correct Answer: A. Because promotions can alter demand through temporary, delayed, or substitution effects