Segmentation, Hierarchies, and Forecast Reconciliation

Introduction

Demand forecasting is rarely a single-number exercise. A supply chain may need a forecast for every stock-keeping unit (SKU), customer, sales channel, warehouse, country, product family, and planning period. These forecasts are related: SKU demand should aggregate to product-family demand; regional demand should aggregate to global demand; weekly demand should aggregate to monthly demand. If independently generated forecasts do not agree, the organization faces a governance problem as well as a statistical one.

Segmentation and hierarchy design address two connected questions:

1. Which items or demand streams should be forecast using similar methods and controls?
2. How should forecasts at different organizational and time levels be produced and made mutually consistent?

Segmentation improves proportionality: high-value or highly volatile items receive more analytical attention. Hierarchical forecasting improves coherence: forecasts at one level agree with forecasts at other levels. Reconciliation provides a systematic way to combine information from aggregate and disaggregate forecasts.

This chapter develops these ideas for graduate students and practitioners in supply chain management. It emphasizes the managerial consequences of statistical choices, the assumptions behind aggregation methods, and the need to evaluate forecasts at the level where decisions are made.

Learning Objectives

After studying this chapter, you should be able to:

• Segment demand streams using value, volume, variability, and forecastability.
• Distinguish SKU, customer, channel, geographic, product, and time hierarchies.
• Explain bottom-up, top-down, middle-out, and forecast-reconciliation approaches.
• Calculate and interpret weighted aggregation and basic reconciliation results.
• Select a forecasting architecture that reflects decision purpose, data quality, and organizational accountability.
• Identify limitations, governance risks, and implementation requirements.

1. Why Segmentation and Hierarchies Matter

A supply chain rarely has enough people, data, or computational capacity to treat every item identically. A pharmaceutical product with high margin, strict service requirements, and intermittent demand should not receive the same process as a low-value commodity with stable demand. Similarly, a national sales plan cannot be managed only at SKU level if executives make capacity and financial decisions at product-family or regional level.

Two principles are important.

First, segmentation is a resource-allocation mechanism. It determines the forecasting method, review frequency, safety-stock policy, exception thresholds, and degree of managerial intervention.

Second, aggregation changes predictability. Demand aggregated across products, customers, or locations is often less variable relative to its mean because positive and negative deviations partially offset. This is not guaranteed: common promotions, weather events, and macroeconomic shocks can produce correlated movements across all components.

A forecast architecture should therefore be designed around decisions. Replenishment may require SKU-location forecasts, while production planning may require family-level forecasts and finance may require a monthly market forecast. A single hierarchy cannot serve every decision equally well.

2. Demand Segmentation Frameworks

2.1 Value and volume segmentation

The familiar ABC analysis ranks items by annual consumption value:

The result is usually ordered from largest to smallest and divided into classes, such as A, B, and C. Typical boundaries might be approximately 70–80% of cumulative value for A items, the next 15–25% for B items, and the remainder for C items. These percentages are conventions, not universal laws.

ABC analysis is useful because it links forecasting effort to financial exposure. However, it can overlook operational importance. A low-value spare part may stop a production line, and a high-volume item may have little margin. Therefore, value should be considered alongside volume, criticality, lead time, substitutability, and service consequences.

Volume segmentation ranks items by units sold or shipped. It is especially relevant to warehouse capacity, transport planning, labor requirements, and production changeovers. A volume-based class and a value-based class may conflict. Such conflicts are informative rather than problematic.

A practical segmentation matrix combines the two dimensions:

2.2 Variability and intermittency

Variability concerns the dispersion of demand around its central tendency. A basic relative measure is the coefficient of variation (CV):

Here, σ is the standard deviation of demand and μ is its mean. Both must be measured in the same units, so CV is dimensionless. The measure is meaningful when the mean is positive and sufficiently stable; it becomes unstable when the mean is near zero.

Intermittent demand has many zero-demand periods and occasional nonzero transactions. Two useful descriptive measures are:

• Demand interval: average number of periods between nonzero demands.
• Nonzero-demand variability: variation in the size of nonzero demands.

A stream with frequent but highly variable demand requires a different approach from one with rare, lumpy demand. Croston-type methods and their variants may be useful for intermittent demand, although their assumptions and bias properties must be examined. For items affected by known events, causal or event-based methods may be more appropriate than purely time-series methods.

2.3 Forecastability

Forecastability is not identical to low variability. A demand stream can be volatile but predictable if its volatility is explained by promotions, seasonality, or known customer schedules. Conversely, a relatively stable series may be difficult to forecast when it contains structural breaks.

Forecastability can be assessed using:

• historical forecast error, such as MAE, RMSE, or WAPE;
• forecast bias;
• performance relative to a simple benchmark;
• stability across rolling-origin evaluation periods;
• ability to explain variation using known causal variables;
• susceptibility to promotions, substitution, stockouts, and lifecycle changes.

Mean absolute error is calculated as:

where y_t is actual demand, ŷ_t is the forecast, and n is the number of evaluated periods. MAE is expressed in demand units. It is appropriate when errors in units have direct operational meaning.

Weighted absolute percentage error is often written as:

WAPE is dimensionless and is useful for aggregate portfolios, but it can hide poor performance on small items and becomes uninformative when total actual demand is zero. Percentage measures should not be used mechanically for intermittent or near-zero series.

Forecastability should be measured out of sample. In-sample fit can reward overfitting and does not establish useful predictive performance.

2.4 A multidimensional segmentation process

A robust process uses the following sequence:

1. Define the unit of analysis: SKU-location, customer-product, channel-product, or another combination.
2. Establish the decision horizon and forecast frequency.
3. Calculate value, volume, variability, intermittency, and historical forecast performance.
4. Add contextual variables such as lifecycle stage, promotion intensity, lead time, criticality, and substitution.
5. Assign segments using transparent rules or a clustering method.
6. Validate segments with planners and operational stakeholders.
7. Specify the method, review cadence, exception rule, and ownership for each segment.
8. Reassess classifications because demand patterns and business priorities change.

Clustering can reveal natural groups, but statistical clusters are not automatically managerial segments. A segment is useful only when it leads to differentiated action.

3. Hierarchies in Demand Forecasting

A hierarchy is a structured set of aggregation relationships. Common dimensions include:

• Product: SKU → brand → product family → total portfolio.
• Customer: account → customer segment → market.
• Channel: store, distributor, e-commerce, and direct sales → total channel demand.
• Geography: site → city → country → region → global.
• Time: day → week → month → quarter → year.

A hierarchy may be strict, where each child has one parent, or grouped, where multiple dimensions intersect. For example, demand may be analyzed by product and geography simultaneously. Such structures are often called grouped time series rather than a simple hierarchy.

A coherent forecast satisfies all aggregation constraints. If product-family demand is the sum of its SKUs, then the forecast of the family must equal the sum of SKU forecasts. If weekly forecasts are aggregated to months, the monthly forecast must equal the sum of the weeks within that month.

Let y_t be a vector of bottom-level demands and b_t be a vector containing all series. The aggregation structure can be represented by a summing matrix S:

The matrix contains zeros and ones in a simple additive hierarchy. Each row identifies which bottom-level series contribute to an aggregate. This formulation is useful because it makes coherence an explicit mathematical constraint.

The value of aggregation

Aggregation can reduce random noise and improve forecast accuracy. If component errors are independent, the variance of their sum is the sum of their variances. If errors are correlated, covariance matters:

If covariances are negative or small, aggregation may substantially smooth demand. If components move together because of a shared promotion or economic shock, the benefit is smaller. Aggregation can also hide important local patterns, so an aggregate forecast must not replace forecasts needed for execution.

4. Forecasting Strategies Across Hierarchies

4.1 Bottom-up forecasting

Bottom-up forecasting begins at the most disaggregated level. Each SKU-location or customer-SKU series is forecast separately, and aggregate forecasts are obtained by summation.

Advantages:

• preserves local detail;
• supports inventory and replenishment decisions;
• naturally produces coherent forecasts;
• allows different methods for different segments.

Limitations:

• bottom-level data may be sparse or noisy;
• many series increase maintenance and exception workload;
• local forecasts may miss aggregate information;
• new products and intermittent items create instability.

Bottom-up is attractive when operational decisions are made at the bottom level and data quality is adequate.

4.2 Top-down forecasting

Top-down forecasting first predicts an aggregate and then allocates it to lower levels. If F_T is the top-level forecast and p_i is the allocation proportion for child i, then:

The proportions may be based on historical average shares, recent shares, proportions of fitted values, or externally informed assumptions. The constraint that proportions sum to one ensures that child forecasts add to the top forecast.

Advantages: useful when lower-level series are noisy, intermittent, or newly created; efficient for strategic planning; captures aggregate trends.

Limitations: may suppress genuine changes in product mix; historical proportions can become obsolete; low-level forecasts inherit errors from the aggregate and allocation rule; a single top-level model may overlook local causal drivers.

Top-down methods are particularly defensible for long-range planning when aggregate demand is more stable than individual demand.

4.3 Middle-out forecasting

Middle-out forecasting begins at a strategically selected intermediate level, such as product family or region. Forecasts are then disaggregated downward and aggregated upward.

This approach balances detail and stability. The middle level should be selected based on data quality, managerial responsibility, and forecastability—not merely organizational habit. For example, product-family demand may be forecast reliably while individual SKU demand is intermittent, but regional demand may be too broad to support useful product decisions.

4.4 Independent forecasts and the coherence problem

An organization may forecast every level independently because different teams own different decisions. This can produce a realistic local forecast at each level, but the results may not add up. For example, independent regional forecasts may imply a total market demand of 980 units while the corporate forecast is 1,050 units.

Incoherence creates ambiguity: Which number should production, finance, and sales use? It may also generate hidden double counting or arbitrary manual overrides. Forecast reconciliation addresses the issue by adjusting a set of forecasts so that the final forecasts satisfy aggregation constraints.

5. Forecast Reconciliation

Let ŷ_h be the vector of base forecasts for all levels and ŷ_h^rec be the reconciled forecasts. A general linear reconciliation form is:

Here, S is the summing matrix, G_h is a reconciliation or mapping matrix that uses the base forecasts, and h denotes the forecast horizon. Because the result is generated through S, it is coherent by construction.

The central design question is how to choose G_h. Different choices produce different reconciliation methods.

5.1 Bottom-up as a special reconciliation strategy

Bottom-up uses only bottom-level forecasts. If b̂_h is the vector of bottom-level forecasts, then:

It is simple and transparent, but it discards potentially useful information in aggregate forecasts.

5.2 Top-down and middle-out allocation

Top-down uses the aggregate forecast and allocation proportions. Middle-out uses an intermediate forecast and allocation or aggregation rules above and below that level. These methods are coherent if the allocation rules are applied consistently.

5.3 Bottom-up versus reconciled forecasts

Reconciliation methods use all available base forecasts. A common least-squares form is:

where S' is the transpose of S, and W_h is a positive-definite error-weight matrix. The matrix assigns relative importance to forecast-error variances and covariances. In practice, estimation of W_h can be difficult, especially with many series, short histories, structural breaks, or changing variance.

If W_h = I, where I is the identity matrix, the method is ordinary least-squares reconciliation. If the matrix is based on historical residual variances, it gives less influence to noisier series. If it estimates the full covariance structure, correlated errors can be handled, although estimation becomes demanding.

The formula minimizes a weighted distance between reconciled and base forecasts subject to coherence. It does not guarantee that reconciled forecasts outperform every individual base forecast. Performance must be evaluated empirically using rolling-origin tests and decision-relevant metrics.

5.4 Illustrative reconciliation calculation

Suppose a company has two regions, North and South, and a corporate total. The independently generated base forecasts are:

The regional forecasts sum to 110, not 120. A simple proportional adjustment preserves the regional relationship while matching the corporate forecast. The scaling factor is:

Therefore:

• North reconciled forecast = 60 × 1.0909 ≈ 65.45;
• South reconciled forecast = 50 × 1.0909 ≈ 54.55;
• total reconciled forecast = 65.45 + 54.55 = 120.

This is an allocation-based reconciliation, not a universal optimal solution. A weighted least-squares method could produce different values if historical errors show that one region is more reliable than the other.

6. Forecast Evaluation in Hierarchies

Evaluation should occur at multiple levels and horizons. A model that performs well at total-company level may perform poorly for critical SKUs. Conversely, a bottom-level model may improve item-level accuracy but create an unstable aggregate plan.

Useful measures include:

• MAE for unit error;
• RMSE when large errors deserve greater penalty;
• WAPE for portfolio-level volume comparison;
• mean absolute scaled error (MASE) for comparisons across series with different units or scales;
• bias, such as mean error, to detect systematic over- or under-forecasting;
• service-level or stockout consequences when forecasts drive inventory decisions.

Accuracy should be measured against actual unconstrained demand where possible. Observed sales during stockouts are censored and may understate true demand. Promotions, price changes, lost sales, substitutions, and product discontinuations should be identified before interpreting errors.

A sound evaluation design uses rolling-origin backtesting. At each origin, the model is trained using information available at that date and evaluated over the subsequent horizon. This replicates operational forecasting and avoids look-ahead bias. Results should be segmented by item class, hierarchy level, horizon, and demand regime.

Reconciled forecasts should be compared with base forecasts and simple benchmarks. Statistical significance is not the only criterion: a small accuracy improvement may not justify substantial implementation complexity, while a modest improvement for high-value or critical items may be highly valuable.

7. Applied Case: Orion Home Appliances

The following is an illustrative case, designed to demonstrate an integrated forecasting process.

Background

Orion Home Appliances sells 1,800 SKUs through retailers, distributors, and its online channel in four geographic regions. The company produces a monthly demand plan for finance, a weekly production plan for factories, and daily replenishment recommendations for distribution centers.

The existing process has three problems. Regional sales teams forecast independently, product managers forecast families independently, and supply planners create SKU forecasts using a statistical system. The resulting numbers do not reconcile. In addition, all items are reviewed monthly, although the top 12% of SKUs account for 76% of annual consumption value.

Problem diagnosis

An initial analysis finds four segments:

1. Core high-value items: regular demand, high margin, low-to-moderate CV.
2. Promotional items: demand strongly affected by retailer campaigns.
3. Long-tail items: low value, intermittent demand, many zero months.
4. New and lifecycle items: limited history, frequent substitutions and phase-outs.

The company also discovers that observed sales for several high-demand items are constrained by stockouts. Using sales without adjustment would systematically understate demand.

Approach

Orion establishes a product-by-region hierarchy and a monthly-to-weekly temporal relationship. It applies the following architecture:

• high-value core items receive SKU-region statistical forecasts with frequent monitoring;
• promotional items use event calendars, promotion depth, and account information;
• long-tail items use simple intermittent-demand rules and inventory-policy review;
• new products use analogues, launch assumptions, and explicit uncertainty ranges;
• product-family and regional forecasts are generated independently to capture aggregate information;
• all base forecasts are reconciled using a weighted method, with weights estimated from rolling residuals and regularized to avoid unstable covariance estimates.

The company preserves two versions of demand: an unconstrained demand estimate for planning and an order-fulfillment forecast for operational execution. This distinction prevents historical stockouts from being treated as evidence of weak market demand.

Outcome

After three quarterly evaluation cycles, the reconciled plan improves coherence immediately because all approved forecasts satisfy the hierarchy. Portfolio WAPE improves modestly, while the largest benefit occurs in product families affected by noisy regional forecasts. Item-level performance is mixed: some long-tail series do not improve, but high-value items receive more focused review. Planners also report fewer meetings devoted to resolving numerical inconsistencies.

The company does not claim that reconciliation alone caused every improvement. Better event data, stockout correction, segmentation, and governance were implemented at the same time. Orion therefore continues controlled backtesting to isolate the contribution of each change.

Learning points

• Segmentation must lead to differentiated action.
• Forecast coherence is a governance requirement, not merely a statistical preference.
• Aggregate information is valuable, but it should not erase local demand drivers.
• Reconciliation cannot correct poor data, an unrecorded promotion, or a structural market change.
• Evaluation must recognize both statistical accuracy and operational consequences.

8. Implementation, Governance, and Technology

Successful implementation requires a documented hierarchy, a data dictionary, stable identifiers, version control, and clear ownership. Every forecast should have a timestamp, horizon, source, method, overrides, and reason codes. Overrides should be logged rather than silently replacing statistical outputs.

Forecasting systems increasingly use machine learning, probabilistic models, and automated hierarchy reconciliation. These tools can process many series and estimate nonlinear relationships, but complexity does not remove the need for segmentation or governance. A model may achieve low error by exploiting information unavailable at the forecast issue date, producing misleading results through leakage.

AI-generated forecasts require particular controls:

• verify that training data represent the relevant demand regime;
• prevent leakage from future prices, promotions, or inventory status;
• monitor performance by segment rather than only in aggregate;
• preserve human accountability for overrides and consequential decisions;
• explain material changes to planners and decision makers;
• protect commercially sensitive customer and transaction data.

Hierarchical forecasts also raise organizational questions. If a regional manager's forecast is reconciled downward, the manager may perceive the process as a performance judgment. The purpose of reconciliation should be communicated clearly: it creates a consistent planning signal; it does not by itself determine accountability or sales targets.

9. Limitations and Critical Evaluation

No hierarchy is permanently correct. Products are redesigned, channels emerge, customers merge, and geographic markets change. A rigid hierarchy can create false precision. Grouped structures may also contain overlapping totals, making careless aggregation prone to double counting.

Top-down methods can fail when mix changes rapidly. Bottom-up methods can fail when data are sparse or when common shocks dominate. Middle-out methods depend heavily on selecting a level that is both statistically stable and managerially meaningful. Reconciliation can improve coherence without improving accuracy, and weighted methods can become unstable when error covariances are poorly estimated.

Forecasts are conditional statements, not facts. They depend on assumptions about prices, promotions, availability, competitors, economic conditions, and customer behavior. Scenario planning and prediction intervals are therefore appropriate when uncertainty is material. A single reconciled point forecast should not be mistaken for a complete decision analysis.

Best-Practice Checklist

• Define the decision, horizon, and level of aggregation before choosing a model.
• Segment using value, volume, variability, forecastability, and business criticality.
• Maintain a documented summing matrix and test hierarchy integrity.
• Use out-of-sample, rolling-origin evaluation.
• Correct for stockouts, promotions, substitutions, and lifecycle events.
• Compare bottom-up, top-down, middle-out, and reconciliation approaches empirically.
• Evaluate accuracy at both aggregate and operational levels.
• Use reason-coded human overrides and retain an audit trail.
• Reassess segments, weights, and hierarchy definitions as the business changes.
• Treat coherence, forecast accuracy, service, inventory, and stakeholder trust as related but distinct outcomes.

Chapter Summary

Segmentation directs forecasting resources toward demand streams with different financial, operational, and statistical characteristics. ABC and volume analysis describe exposure, while variability, intermittency, and forecastability explain how demand should be modeled and governed. Effective segments produce differentiated methods and review practices.

Demand hierarchies connect SKU, customer, channel, geographic, product, and time levels. Aggregation can improve predictability, but common shocks and local patterns limit its benefits. Bottom-up forecasting preserves detail, top-down forecasting exploits aggregate stability, and middle-out forecasting balances the two. Reconciliation combines information from independently generated forecasts while imposing mathematical coherence.

The choice of method should be evaluated through rolling-origin tests, multiple accuracy measures, operational outcomes, and managerial practicality. Reconciliation is not a substitute for good data or sound causal reasoning. Its greatest value often lies in creating a transparent, consistent planning system in which forecasts support decisions at every relevant level.

Key Terms

Segmentation: Dividing demand streams into groups that receive differentiated forecasting and management treatment.

ABC analysis: Classification based on cumulative annual consumption value.

Coefficient of variation: Standard deviation divided by the mean, measuring relative variability.

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

Forecastability: The extent to which future demand can be predicted reliably from available information.

Hierarchy: Structured levels connected by aggregation relationships.

Grouped time series: Multiple overlapping hierarchies, such as product and geography together.

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

Top-down forecasting: Forecasting an aggregate and allocating it downward.

Middle-out forecasting: Forecasting an intermediate level and extending forecasts upward and downward.

Base forecast: An independently generated forecast before reconciliation.

Coherence: The condition that forecasts at all levels satisfy aggregation constraints.

Reconciliation: Adjusting base forecasts to produce a coherent set of forecasts.

Summing matrix: A matrix representing how bottom-level series form aggregate series.

WAPE: Weighted absolute percentage error based on total absolute error divided by total actual demand.

Rolling-origin evaluation: Repeated out-of-sample testing using historical forecast origins.

Forecast bias: Systematic tendency to over-forecast or under-forecast.

Censored demand: Demand hidden by stockouts or other supply constraints.

Override governance: Rules and records controlling human changes to statistical forecasts.

Review Questions

Short Answer Questions

1. Why should ABC analysis be supplemented with operational criticality?
2. Define forecast coherence in a hierarchical forecasting system.
3. What is the principal difference between bottom-up and top-down forecasting?
4. Why can aggregation improve forecastability?
5. What is the purpose of a summing matrix?

Long Answer Questions

1. Explain how value, volume, variability, intermittency, and forecastability can be integrated into a segmentation policy.
2. Compare the advantages and limitations of bottom-up, top-down, and middle-out forecasting.
3. Explain the logic of weighted forecast reconciliation and the role of the error-weight matrix.
4. Design an evaluation framework for a company forecasting across SKU, region, and monthly time levels.
5. Discuss why stockouts, promotions, and product lifecycle changes complicate hierarchical forecasting.

Analytical / Application Questions

1. A company forecasts a total of 500 units, while its three independently forecast regions total 460 units. Propose and calculate a proportional allocation adjustment, and discuss when that adjustment may be inappropriate.
2. Design a segmentation scheme for a spare-parts portfolio containing high-value intermittent items, low-value stable items, and items critical to production uptime. Specify forecast methods and review policies.
3. A reconciled forecast improves corporate-level WAPE but worsens accuracy for several high-value SKU-location combinations. How should management decide whether to retain the reconciliation process? Identify the additional evidence required.

Multiple Choice Questions with Answers

Question 1

Which measure combines demand quantity with financial exposure in ABC analysis?

A. Annual consumption value

B. Coefficient of variation

C. Mean absolute error

D. Demand interval

Correct Answer: A. Annual consumption value

Question 2

A demand series has a mean of 200 units and a standard deviation of 50 units. What is its coefficient of variation?

A. 0.04

B. 0.25

C. 4.00

D. 250

Correct Answer: B. 0.25

Question 3

Which forecasting strategy begins by forecasting the most disaggregated series?

A. Top-down

B. Middle-out

C. Bottom-up

D. Proportional allocation

Correct Answer: C. Bottom-up

Question 4

What is the defining property of a coherent hierarchical forecast?

A. Every series uses the same forecasting model

B. All forecasts are unbiased

C. Forecasts at different levels satisfy aggregation constraints

D. All forecasts have identical variances

Correct Answer: C. Forecasts at different levels satisfy aggregation constraints

Question 5

In a top-down approach, allocation proportions for all children must normally:

A. Be equal to zero

B. Sum to one

C. Be greater than one

D. Be independent of the aggregate forecast

Correct Answer: B. Sum to one

Question 6

Which issue can cause observed sales to understate true demand?

A. A stockout

B. A positive covariance

C. A summing matrix

D. A high service level

Correct Answer: A. A stockout

Question 7

In the reconciliation expression ŷ_h^rec = S G_h ŷ_h, what does S represent?

A. The forecast horizon

B. The summing matrix

C. The standard deviation matrix

D. The seasonal index

Correct Answer: B. The summing matrix

Question 8

A corporate forecast is 120 units and two regional forecasts are 60 and 50 units. Under proportional scaling, what is the reconciled North forecast?

A. 60.00

B. 65.45

C. 70.00

D. 72.00

Correct Answer: B. 65.45

Question 9

Which evaluation design best reduces look-ahead bias?

A. In-sample fit only

B. One comparison using future information

C. Rolling-origin out-of-sample evaluation

D. Judgment based only on the latest forecast

Correct Answer: C. Rolling-origin out-of-sample evaluation

Question 10

Why may reconciliation fail to improve every individual series?

A. Coherence and accuracy are different objectives

B. Reconciliation always removes all local information

C. A coherent forecast cannot be evaluated

D. Aggregation prevents any forecast from being calculated

Correct Answer: A. Coherence and accuracy are different objectives