Hierarchical, Grouped and Multivariate Forecasting

Introduction

Most organizations do not make decisions from a single time series. A retailer forecasts sales by product, store, region, channel, and enterprise; a manufacturer forecasts demand for components, finished goods, and product families; a service firm forecasts customers by segment and location. These series are related by aggregation relationships. If independent forecasts are produced for every series, they will usually contradict one another. Regional forecasts may not sum to the national forecast, channel forecasts may not equal the product total, and short-term forecasts may be inconsistent with annual plans.

Hierarchical, grouped, and multivariate forecasting addresses this problem. It combines the statistical forecasting of individual series with structural information about how series aggregate and how they influence one another. The central managerial requirement is coherence: forecasts must satisfy the organization’s known aggregation constraints while retaining as much predictive accuracy as possible.

The topic is especially important in operations management because coherent forecasts support inventory positioning, capacity planning, procurement, workforce scheduling, budgeting, and sales and operations planning (S&OP). Coherence, however, does not guarantee accuracy. A mathematically consistent forecast can still be biased, poorly calibrated, or operationally unusable. Advanced practice therefore treats reconciliation as one component of a broader forecasting system.

Learning objectives

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

1. Distinguish hierarchical, grouped, cross-sectional, and temporal forecasting structures.
2. Represent aggregation constraints using a summing matrix.
3. Explain top-down, bottom-up, middle-out, and forecast-reconciliation methods.
4. Derive and interpret the MinT reconciliation framework.
5. Assess cross-series dependence using multivariate models and forecast-error covariance.
6. Evaluate aggregation trade-offs, forecast accuracy, coherence, and implementation risks.
7. Design a practical forecasting process for a multi-dimensional organization.

1. Structural foundations

1.1 Hierarchies and grouped structures

A hierarchy has strictly nested levels. For example:

• Total company demand
• Region A and Region B
• Stores within each region
• Products within each store

Each lower-level series belongs to one and only one parent. A grouped time series permits several valid classifications that are not strictly nested. Sales may be grouped simultaneously by product category, geography, and channel. A product can belong to one category and be sold through several channels, while a region contains several stores. These dimensions form a cross-classification rather than a single tree.

A cross-sectional hierarchy aggregates units at one time granularity. A temporal hierarchy contains the same phenomenon at different frequencies, such as daily, weekly, monthly, and quarterly demand. A hierarchical forecast system may contain both: for example, product-by-store demand reported daily and reconciled to weekly and monthly plans.

The distinction matters because the aggregation constraints differ. In a simple hierarchy, parent demand equals the sum of its children. In grouped structures, several sets of equations must hold simultaneously, and naïve sequential aggregation can double-count or create contradictions.

1.2 Coherence

Let the vector of all series at a particular time be yₜ, including bottom-level and aggregate series. Forecasts are coherent if they can be generated from a vector of bottom-level values bₜ through a summing matrix S:

Here, S is a known, non-random matrix containing zeros and ones (or, in some applications, other fixed aggregation weights), and bₜ contains the bottom-level series. A forecast vector ŷₜ|T is coherent when:

where b̂ₜ|T is a forecast of the bottom-level series. Equivalently, coherent forecasts lie in the linear subspace spanned by S. The matrix representation is useful because it converts a verbal organizational structure into explicit constraints that can be audited and implemented computationally.

For example, if total demand T equals product A plus product B, then the bottom-level vector is bₜ = (Aₜ, Bₜ)′, and the full vector can be written as yₜ = (Tₜ, Aₜ, Bₜ)′ with:

Thus, a bottom-level forecast of 60 and 40 implies a coherent total forecast of 100. A separately generated total forecast of 115 is not coherent with these product forecasts.

Figure: A hierarchical structure in which parent forecasts must equal the sum of child forecasts.

2. Forecasting strategies

2.1 Bottom-up forecasting

Bottom-up forecasting models each bottom-level series and aggregates the resulting forecasts upward. If b̂ₜ+h|T is the vector of h-step-ahead bottom-level forecasts, the complete forecast is:

Its main advantage is detail: local patterns, promotions, store closures, and assortment effects can be represented directly. Its disadvantages are data sparsity, high computational burden, and noisy forecasts at granular levels. It is most attractive when bottom-level data are plentiful and local decisions dominate.

2.2 Top-down forecasting

Top-down methods forecast an aggregate and allocate it to lower levels. A proportional allocation approach uses historical shares. If total forecast is T̂ₜ+h|T and the estimated share of unit i is pᵢ, then:

The shares may be based on historical averages, recent proportions, or proportions adjusted for known structural changes. The method is simple and often more stable, but it can suppress genuine local dynamics and assumes that future shares are sufficiently predictable.

A forecast-proportion method first forecasts the components independently and uses their relative forecasts as allocation proportions. This retains more information than fixed historical shares but is not identical to bottom-up forecasting when the aggregate forecast is produced independently.

2.3 Middle-out forecasting

Middle-out forecasting selects an intermediate level as the forecasting and allocation anchor. For example, regional forecasts may be modeled directly, then allocated to stores using store shares. This can balance the stability of aggregation with the operational relevance of local detail. The selected level should have adequate data, meaningful managerial ownership, and a plausible relationship to the demand drivers.

2.4 Why simple strategies can fail

Top-down and bottom-up methods impose coherence, but they do so by restricting where information enters the system. Bottom-up forecasts may overreact to intermittent demand; top-down methods may hide a new store’s growth; middle-out methods may inherit errors from both directions. No strategy is universally superior. Empirical evaluation should compare methods by horizon, series, and decision use rather than relying on a general rule.

3. Forecast reconciliation

3.1 The reconciliation problem

Suppose independent base forecasts are produced for all series, giving ŷ̂ₜ+h|T. The hat distinguishes these forecasts from the reconciled forecasts. They may be statistically strong individually but incoherent collectively. Reconciliation transforms them into forecasts that satisfy the aggregation constraints:

Here, P is a reconciliation matrix. Different choices of P produce different methods.

3.2 Bottom-up as a special case

If only bottom-level forecasts are retained, P selects the bottom-level components and discards independently forecast aggregates. The resulting method is bottom-up. Top-down and middle-out methods can also be expressed as particular constrained allocation rules, although their allocation assumptions are not generally equivalent to least-squares reconciliation.

3.3 OLS and weighted least squares

A simple reconciliation approach finds the coherent vector closest to the base forecasts under squared error loss. Ordinary least squares uses equal weighting:

where ŷ̂ is the base forecast vector and y is the reconciled vector. Equal weighting is often unrealistic because series have different scales and noise levels. Weighted least squares therefore uses a positive-definite weight matrix W:

The inverse appears because a high-variance series should generally receive less weight than a precise series. Units must be consistent: if series are measured in different units, suitable scaling or a unit-consistent covariance estimate is required.

3.4 MinT reconciliation

The minimum trace, or MinT, method chooses the coherent forecast with the smallest estimated total forecast-error variance. Let Wₕ be the covariance matrix of h-step-ahead base forecast errors. Under linear unbiased reconciliation, the MinT solution is:

The matrix Wₕ captures both the variance of each base forecast error and covariance between errors. If errors are correlated, using only diagonal variances can be inefficient. In practice, estimating a large covariance matrix is difficult, so analysts use alternatives such as a diagonal variance matrix, a shrinkage estimator, or a scaled covariance estimator.

The derivation assumes a linear aggregation structure, unbiased or approximately unbiased base forecasts, and an adequately estimated error covariance matrix. MinT minimizes the trace of the reconciled forecast-error covariance matrix, not necessarily the error of every individual series. Consequently, a method can improve aggregate accuracy while worsening some low-volume series.

3.5 Numerical illustration

Consider two bottom-level products with base forecasts of 60 and 40 units. Their independently forecast total is 115 units. The bottom-up coherent total is 100 units. If management trusts the product forecasts more than the aggregate forecast, reconciliation should move the total toward 100. If the aggregate forecast is based on a strong macroeconomic model while product forecasts are noisy, a weighted reconciliation may adjust product forecasts upward so that their sum lies closer to 115. The correct adjustment depends on estimated error variances and covariances, not on the numerical discrepancy alone.

This example illustrates a central principle: reconciliation is not merely arithmetic aggregation. It is an evidence-weighting problem constrained by organizational structure.

4. Multivariate forecasting and cross-series relationships

4.1 Why model series jointly?

Univariate models forecast each series from its own history. Multivariate models use information from several series, such as substitution, common seasonality, promotions, economic indicators, or supply constraints. Joint modeling is valuable when cross-series relationships are stable, economically interpretable, and strong enough to improve out-of-sample forecasts.

A vector autoregression (VAR) for a k-dimensional vector yₜ with p lags is:

Here, c is a k-dimensional intercept, Aⱼ are k × k coefficient matrices, and εₜ is a zero-mean innovation vector with covariance matrix Σ. The model assumes weak stationarity unless transformed or extended; otherwise, coefficients and forecasts may be unstable. If series are non-stationary but cointegrated, a vector error-correction model (VECM) may be more suitable because it represents both short-run dynamics and long-run equilibrium relationships.

A dynamic regression can incorporate external drivers:

where xⱼ,ₜ may represent price, advertising, temperature, competitor activity, or availability. The analyst must distinguish demand from observed sales: stockouts, rationing, and lost sales can make the dependent variable an incomplete measure of demand.

4.2 Cross-series dependence and reconciliation

Cross-series dependence appears in two places. First, the level forecasts may use other series as predictors. Second, base forecast errors may be correlated even when models are fitted separately. Reconciliation uses the second type through Wₕ. A useful system may therefore combine multivariate modeling with reconciliation: the former improves base forecasts, while the latter enforces structural consistency.

Correlation alone does not establish a causal relationship. Two product series may rise together because of common seasonality, while a promotion may cause one product to substitute for another. Analysts should use domain knowledge, experimental variation where available, lag analysis, and out-of-sample validation before interpreting coefficients causally.

4.3 Dimensionality and regularization

Grouped organizations can generate thousands of related series. A fully unrestricted VAR then estimates too many coefficients. Regularization, factor models, dynamic factor models, and hierarchical shrinkage reduce variance by constraining or pooling parameters. For example, a dynamic factor model represents many series through a smaller number of latent common factors:

where fₜ is a low-dimensional factor vector, Λ is a loading matrix, and eₜ contains idiosyncratic components. This approach is useful when many products respond to common macroeconomic or seasonal movements, but factors can be difficult to interpret and may not preserve exact aggregation without a subsequent reconciliation step.

5. Aggregation trade-offs and evaluation

Aggregation reduces noise and often improves forecastability because random fluctuations partially cancel. However, aggregation can conceal local peaks, stockout effects, and substitution. A national forecast may be accurate while individual store forecasts are operationally poor. Disaggregation also introduces the opposite problem: intermittent demand and sparse histories increase uncertainty.

Forecast evaluation should therefore be multi-level. Common measures include mean absolute error (MAE), root mean squared error (RMSE), and weighted absolute percentage error. For observations yₜ and forecasts fₜ over n periods:

MAE has the same units as demand and gives equal weight to absolute errors. RMSE penalizes large errors more heavily. Percentage measures become unstable when actual demand is zero or near zero, so scaled or volume-weighted measures are usually preferable for intermittent and mixed-scale portfolios.

Evaluation must use rolling-origin or time-series cross-validation, not random train-test splits. The comparison should report:

• accuracy at each hierarchy level;
• accuracy by forecast horizon;
• bias and service-level consequences;
• prediction-interval coverage;
• computational cost and latency;
• impacts on inventory, capacity, and revenue decisions.

A coherent forecast can be evaluated with a weighted aggregate loss:

where wᵢ reflects volume, margin, criticality, or decision importance, and ℓ is a chosen loss function. Weights should be governed transparently because they encode managerial priorities.

6. Implementation architecture and governance

A robust implementation normally follows these steps:

1. Define the structure. Identify products, locations, channels, time frequencies, valid parents, discontinued entities, and exceptional aggregation rules.
2. Construct the summing matrix. Validate that each series maps to the correct parents and that units and calendars are aligned.
3. Prepare data. Correct for stockouts, returns, promotions, product introductions, closures, price changes, and calendar effects.
4. Generate base forecasts. Use suitable univariate, causal, machine-learning, or multivariate models by series and horizon.
5. Estimate forecast-error covariance. Use rolling forecast residuals; apply shrinkage or diagonal approximations when sample size is limited.
6. Reconcile. Compare bottom-up, top-down, middle-out, OLS, WLS, and MinT variants.
7. Evaluate prospectively. Use rolling-origin tests and assess both statistical and operational outcomes.
8. Apply judgment with controls. Record overrides, reasons, owner, date, and expected effect; do not permit undocumented manual changes.
9. Monitor drift. Track bias, interval coverage, structural changes, and coherence violations.

Modern forecasting platforms can automate matrix construction, model selection, covariance estimation, and reconciliation. Machine learning can be useful for nonlinear effects and high-dimensional drivers, but it does not remove the need for structural knowledge. An opaque model that forecasts each series independently may produce excellent local predictions yet create impossible totals. Explainability, reproducibility, version control, and audit trails are essential in regulated or financially material planning processes.

7. Illustrative case: a multi-channel retailer

Background

A fictional retailer sells household products through stores, a website, and a wholesale channel. It plans demand monthly by product category, region, channel, and total company. The existing process produces separate spreadsheets. The sum of regional forecasts differs from the channel plan, and store managers frequently override forecasts after observing promotions and stockouts.

Problem

The retailer experiences three consequences: procurement uses an inflated total, distribution centers receive inconsistent regional requirements, and performance reviews cannot distinguish forecast error from manual override effects. The organization needs coherent forecasts without eliminating local knowledge.

Analysis

The analytics team maps each sale to product, region, channel, and month. It identifies genuine demand where possible by flagging stockout periods. A product-by-region-by-channel table becomes the bottom-level structure. Monthly category and company totals are represented as grouped aggregates. Residual analysis shows strong common seasonality across channels and correlated errors during national promotions. Store-level histories are sparse, whereas regional histories are stable.

Approach and solution

The team produces base forecasts using dynamic regression for promotions and prices, seasonal models for stable categories, and pooled models for sparse stores. Regional and channel forecasts are modeled directly because they have sufficient history. A shrinkage estimate of the base-error covariance matrix is used in MinT reconciliation. For newly opened stores, a middle-out allocation from regional forecasts is used until sufficient local history accumulates.

Overrides are permitted only for documented events such as confirmed campaigns, closures, or supply restrictions. The system records the unconstrained forecast, the reconciled forecast, and the final approved plan. Forecast accuracy is assessed using rolling-origin tests at store, region, channel, and company levels, with inventory cost and service level included in the decision evaluation.

Outcome

In the pilot, company and regional totals become mathematically coherent, while the reconciliation adjustment is smaller for high-confidence series and larger for noisy store series. Managers retain the ability to represent known events, but undocumented optimism is reduced. The retailer does not select one method permanently; it uses bottom-up forecasting for mature high-volume stores, middle-out allocation for new stores, and covariance-based reconciliation for the full grouped system.

Learning points

1. Organizational structure should be modeled explicitly rather than repaired manually at the end of the planning cycle.
2. Forecast accuracy and coherence are different objectives and must be monitored separately.
3. Cross-series error covariance can improve reconciliation, but estimating it requires sufficient historical forecast errors.
4. New entities need a transitional method because they lack reliable local histories.
5. Governance of overrides is as important as the forecasting algorithm.

8. Challenges, limitations, and ethical considerations

Structural definitions can change when products are reclassified, stores close, or channels merge. Historical data must then be restated or modeled with time-varying structures. Reconciliation also distributes errors: improving a total may worsen a strategically important component. Covariance estimates are unstable when the number of series is large relative to the number of forecast-error observations. Forecast intervals are more difficult to reconcile than point forecasts, and naïve interval combination can produce misleading coverage.

Bias is a further concern. If historical sales reflect unequal service levels, discriminatory allocation, or systematic underinvestment in certain regions, forecasting those sales without adjustment can reproduce the inequality. Promotional data may encode commercial power relationships, while algorithmic overrides may favor high-revenue channels over essential low-volume customers. Governance should define fairness objectives, data provenance, override accountability, and escalation procedures.

Confidentiality also matters in grouped systems. A highly detailed forecast can reveal commercially sensitive information about individual stores, suppliers, or customer segments. Access controls and aggregation thresholds should be applied without undermining the mathematical structure. Forecasts used in labor or credit decisions require additional review for legal compliance and disparate impact.

Best-practice checklist

• Treat the hierarchy or grouping as a governed business object.
• Maintain a validated summing matrix and explicit units.
• Compare several reconciliation methods using rolling-origin evaluation.
• Use covariance information where sample size and quality justify it.
• Evaluate forecasts at the levels where decisions are made.
• Separate observed sales from censored demand caused by stockouts.
• Reconcile prediction intervals carefully, not only point forecasts.
• Record every judgmental override and its ex post performance.
• Use shrinkage, pooling, or factor models for high-dimensional systems.
• Monitor structural breaks, bias, drift, and coherence continuously.

Chapter Summary

Hierarchical and grouped forecasting addresses the need to forecast many related series while preserving aggregation relationships. A summing matrix expresses these relationships formally, and coherent forecasts are generated as S times a vector of bottom-level forecasts. Bottom-up methods preserve local information; top-down methods emphasize aggregate stability; middle-out methods use an intermediate planning level. Forecast reconciliation provides a broader framework in which independent base forecasts are projected onto the coherent subspace.

OLS, weighted least squares, and MinT differ primarily in how they measure the cost of altering base forecasts. MinT uses the covariance matrix of forecast errors and can exploit cross-series dependence, although covariance estimation is challenging in large systems. VAR, dynamic regression, factor models, and related multivariate methods can improve base forecasts when relationships are stable and validated out of sample. Aggregation may reduce noise but conceal local variation, so evaluation must consider levels, horizons, intervals, and operational consequences. Effective implementation combines statistical methods with data governance, transparent overrides, ethical safeguards, and continuous monitoring.

Key Terms

Coherence: The property that forecasts satisfy all required aggregation constraints.

Hierarchical forecasting: Forecasting series arranged in a nested parent–child structure.

Grouped forecasting: Forecasting series classified through multiple simultaneous, possibly non-nested dimensions.

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

Top-down forecasting: Forecasting an aggregate and allocating it to lower levels.

Middle-out forecasting: Forecasting an intermediate level and allocating upward or downward.

Reconciliation: Adjusting base forecasts so that they satisfy structural constraints.

Summing matrix: A matrix that maps bottom-level series to all hierarchical or grouped series.

Base forecast: An independently generated forecast before reconciliation.

MinT: Minimum-trace reconciliation that minimizes estimated total forecast-error variance.

Forecast-error covariance: The variance and cross-series covariance structure of forecast errors.

VAR: Vector autoregression using lagged values of several series.

VECM: Vector error-correction model for non-stationary but cointegrated series.

Dynamic regression: Regression with time-dependent errors and explanatory variables.

Dynamic factor model: A model representing many series through a smaller set of common factors.

Aggregation bias: Distortion caused by forecasting at an aggregate level that conceals component behavior.

Rolling-origin evaluation: Repeated time-ordered training and testing for forecast comparison.

Forecast override: A documented human adjustment to a statistical forecast.

Censored demand: Unobserved demand constrained by stockouts or limited availability.

Review Questions

Short Answer Questions

1. What distinguishes a hierarchy from a grouped time-series structure?
2. Define forecast coherence and explain why it matters operationally.
3. What is the principal advantage and disadvantage of bottom-up forecasting?
4. Why can forecast-error covariance matter in reconciliation?
5. Why are random train-test splits inappropriate for most time-series evaluation?

Long Answer Questions

1. Explain top-down, bottom-up, middle-out, OLS, WLS, and MinT forecasting, comparing their assumptions and managerial uses.
2. Derive the role of the summing matrix in constructing coherent forecasts and illustrate it with a simple product hierarchy.
3. Discuss how VAR, dynamic regression, and factor models can incorporate cross-series relationships.
4. Analyze the trade-off between forecast accuracy at aggregate and granular levels.
5. Design a governance framework for judgmental overrides in a hierarchical forecasting system.

Analytical / Application Questions

1. A company’s total forecast is 10,000 units, while independently forecast regional totals sum to 10,800. Propose a reconciliation approach and explain what additional evidence is needed before selecting the adjustment.
2. A retailer has many new stores with sparse demand histories but stable regional data. Compare bottom-up, top-down, and middle-out approaches for the first six months of each store’s operation.
3. Design a rolling-origin experiment to test whether MinT improves a grouped product–region–channel forecasting system. Specify accuracy metrics, aggregation levels, and operational outcomes.

Multiple Choice Questions with Answers

Question 1

Which condition defines a coherent forecast vector?

A. Every series uses the same forecasting algorithm

B. All forecasts satisfy the required aggregation relationships

C. All forecasts have identical prediction intervals

D. The aggregate forecast is always more accurate than component forecasts

Correct Answer: B. All forecasts satisfy the required aggregation relationships

Question 2

In bottom-up forecasting, what is forecast first?

A. The total company series

B. The highest geographic level

C. The bottom-level series

D. The annual budget

Correct Answer: C. The bottom-level series

Question 3

What is the primary role of the summing matrix S?

A. It estimates promotional elasticity

B. It maps bottom-level values to all required aggregate series

C. It removes seasonality from every series

D. It calculates forecast accuracy

Correct Answer: B. It maps bottom-level values to all required aggregate series

Question 4

Top-down forecasting typically requires which operation after forecasting the aggregate?

A. Allocating the aggregate forecast using component shares

B. Deleting all component histories

C. Estimating a VAR with no lags

D. Replacing demand with inventory

Correct Answer: A. Allocating the aggregate forecast using component shares

Question 5

What distinguishes MinT reconciliation from equal-weight OLS reconciliation?

A. MinT ignores aggregation constraints

B. MinT uses estimated forecast-error covariance information

C. MinT forecasts only annual totals

D. MinT requires all bottom-level series to be zero

Correct Answer: B. MinT uses estimated forecast-error covariance information

Question 6

In a VAR model, lagged values of what may explain a series?

A. Only the same series

B. Only external economic variables

C. The series itself and other modeled series

D. Only the forecast horizon

Correct Answer: C. The series itself and other modeled series

Question 7

Why can aggregation improve forecastability?

A. It guarantees causal relationships

B. Random fluctuations can partially cancel at higher levels

C. It eliminates all structural breaks

D. It makes every series stationary

Correct Answer: B. Random fluctuations can partially cancel at higher levels

Question 8

Which metric is expressed in the same units as demand?

A. MAE

B. MAPE

C. Percentage bias only

D. Share of total demand

Correct Answer: A. MAE

Question 9

Why is a stockout important when preparing demand data?

A. Observed sales may understate actual demand

B. It proves the forecast was unbiased

C. It makes reconciliation unnecessary

D. It always increases seasonal variation

Correct Answer: A. Observed sales may understate actual demand

Question 10

What is the main purpose of rolling-origin evaluation?

A. To randomize the temporal order of observations

B. To imitate repeated future forecasting using time-ordered data

C. To force all models to have equal parameters

D. To aggregate every series into one total

Correct Answer: B. To imitate repeated future forecasting using time-ordered data