Overview
Demand segmentation is the practice of dividing demand into groups with characteristics that justify different planning or service decisions. Depending on the decision, a segment may consist of customers, products, locations, sales channels, or combinations such as a product at a particular warehouse.
In supply chain management, segmentation connects demand patterns and customer requirements to forecasting, inventory, sourcing, and fulfillment policies. It complements the broader planning and execution processes described by the ASCM Supply Chain Operations Reference model. Transaction systems such as inventory management software can supply stock and movement data, but segment definitions and the policies attached to them require separate analytical design.
Etymology and Origins
The term combines demand, the quantity of a product or service sought, with segmentation, the division of a whole into distinct groups. Its business usage draws on market segmentation: in 1956, marketing scholar Wendell R. Smith described market segmentation as recognizing differences in demand and adjusting market offerings accordingly. The article appeared in the Journal of Marketing and treated segmentation as an alternative to relying only on product differentiation.
Supply chain practitioners later extended the idea from differentiating customers and marketing offers to differentiating operational responses. Marshall L. Fisher's 1997 framework distinguished functional products, with relatively predictable demand, from innovative products, with uncertain demand, and argued that supply chains should be matched to those characteristics. No single first use of the precise phrase demand segmentation is established here; Smith's article is a documented antecedent, not proof that he coined the phrase.
Historical Development
- 1950s: Smith's 1956 article provided a foundational account of market segmentation. Inventory classification practices developed around the same era were related to, but not identical to, modern demand segmentation. ABC classification ranks items by a concentration measure such as annual usage value; it does not, by itself, describe intermittency, seasonality, or customer service requirements.
- 1970s: J. D. Croston's 1972 research addressed intermittent demand by separately estimating demand size and the interval between demands. This established an influential alternative to treating sparse demand like regularly occurring demand.
- 1970s–1990s: Commercial barcode scanning and computerized inventory systems made it increasingly practical to collect item-level transaction data. The first retail scan of a Universal Product Code (UPC)—a barcode standard now administered within the GS1 system—occurred on 26 June 1974. Calling that symbol a "GS1 barcode" at the time would be anachronistic, because GS1 was formed later from earlier numbering organizations. Enterprise systems subsequently consolidated item, location, and transaction information across functions.
- 1990s–2000s: Fisher's 1997 article linked supply chain design to differences in product demand. In 2005, Syntetos, Boylan, and Croston published a demand-categorization framework for selecting forecasting approaches for intermittent demand. Related research by Syntetos and Boylan, also published in 2005, examined the accuracy of intermittent-demand estimates and set out a bias correction to Croston's method, often called the Syntetos–Boylan approximation.
- 2010s onward: Cloud platforms, broader data integration, and machine-learning methods have increased the practical scope for frequent reclassification and more granular forecasts. These tools do not eliminate the need to test whether a segment supports a useful decision.
Core Concepts
A segmentation scheme begins with a decision: for example, which items should receive higher availability targets or which demand series need intermittent-demand forecasting. Planners then select a unit of analysis, a review period, relevant measures, and rules for assigning groups. A SKU can belong to different segments at different locations because local demand, lead time, and service requirements differ.
Common dimensions include demand volume, value, variability, seasonality, frequency, customer criticality, margin, lead time, and required delivery speed. Historical sales should be interpreted carefully: when an item is unavailable, recorded sales can understate actual demand. Forecasting principles also emphasize evaluating methods on data not used to fit them.
Segments must lead to explicit policies. A high-volume, stable item might receive frequent automated replenishment, while a slow, erratic spare part might be pooled centrally or ordered on request. A stock replenishment workflow is one operational mechanism for executing a chosen policy; segmentation itself determines which policy is appropriate.
Types and Classifications
Value and volume
ABC classification ranks items by a chosen measure, often annual demand value calculated as annual unit demand multiplied by unit cost. A-items typically account for a large share of the chosen measure and receive closer review; B- and C-items receive progressively simpler controls. Thresholds are organizational choices, not universal constants. Inventory research on ABC classes shows that value rank alone does not determine the service level or inventory cost that a class should receive. High value does not necessarily mean high variability or high customer criticality.
Variability and intermittency
An XYZ-style classification groups items by demand predictability or variability, although definitions and cutoffs differ between organizations. For nonzero mean demand, the coefficient of variation is the standard deviation divided by the mean; it should be interpreted cautiously for seasonal, trending, or very low-volume series.
Intermittent demand contains many periods with no demand. Classification can consider both the average interval between demand occurrences and variation in nonzero demand sizes. Syntetos and colleagues' research provides a framework for distinguishing demand patterns and selecting forecasting methods. An ABC–XYZ matrix combines value and variability, but it may need additional flags for critical spare parts or regulatory requirements.
Croston's method and its variants estimate average demand per period; they are not, by themselves, a complete inventory policy. Compare candidate methods using a relevant holdout and inventory outcomes, rather than assuming that one intermittent-demand method is always best. Research comparing intermittent-demand estimators illustrates why the choice of estimator matters, including the positive bias that can arise from Croston's original estimator.
Customer and service needs
Customer or channel segments can reflect order size, promised lead time, geography, delivery windows, and willingness to accept substitutions. Fisher's product-demand framework illustrates why predictable products and products with uncertain demand can require different supply chain responses. Customer priority is not the same as product-demand pattern: a low-volume item can still warrant a responsive policy if the consequence of a stockout is severe.
Applications and Practical Example
A distributor might classify each SKU-location by annual demand value and demand pattern, then overlay a criticality flag:
- High-volume, stable consumable: Forecast regularly, replenish frequently, and place stock near demand.
- Low-volume, critical spare part: Set a service target based on the cost of equipment downtime rather than sales value alone; consider a central stock pool.
- Seasonal promotional product: Incorporate promotion dates and expected uplift, then review excess-stock exposure after the event.
- Highly intermittent, noncritical item: Consider supplier ordering or shared inventory rather than stocking every location.
For illustration, an item selling 1,000 units a year at a unit cost of $20 has annual demand value of $20,000. If another item sells 20 units at $500 each, its value is $10,000, but its operational priority could still be higher if a stockout stops production. These figures are illustrative, not industry benchmarks.
For the critical spare, the distributor could compare local stocking with a central pool against a stated delivery promise. If central delivery cannot meet that promise, the lower-inventory option may be unsuitable. This is a hypothetical decision, not a documented case study; it follows the principle of matching supply chain responsiveness to demand requirements.
Purchase order management and order management systems can record and execute differentiated purchasing and fulfillment decisions once policies have been chosen. Their suitability depends on configuration, available data, and the organization's planning process; software does not define the segments.
Implementation and Best Practices
- Define the decision and ownership. Agree on whether the objective is forecast selection, inventory placement, service differentiation, or another outcome. Assign a business owner for segment rules and exceptions.
- Prepare the data. Align units of measure, product identifiers, locations, and time buckets. Investigate returns, stockouts, promotions, substitutions, and product launches before interpreting historical sales as unconstrained demand. Reliable inventory records support this work but do not by themselves reveal lost demand. Document purchase and stocking units and their unit-of-measure conversions before comparing demand values.
- Choose interpretable rules. Start with a small number of dimensions and document thresholds. Add criticality and lifecycle overrides where a simple value ranking would create unacceptable risk.
- Set differentiated policies. Specify the forecast method, review frequency, service target, replenishment approach, and escalation process for each segment. Define whether availability means an order fill rate or another measure, and use the same definition when comparing policies.
- Test and monitor. Compare outcomes with a baseline using forecast error, bias, availability, inventory, obsolescence, and cost. Use holdout periods and monitor results by segment; forecast evaluation guidance explains why in-sample fit alone can be misleading. Test over a horizon relevant to the replenishment decision and compare like-for-like periods where possible.
- Review classifications periodically. Reclassify after meaningful changes in demand, lead time, product lifecycle, or customer commitments. Use minimum history requirements or buffers around thresholds to prevent frequent switching between segments.
Challenges and Limitations
Segmentation can create false precision when data are sparse, sales are censored by stockouts, or products are new. Aggregating across locations may conceal local volatility, while excessively granular segments can leave too little data for reliable estimates. Where forecasts are needed at both local and aggregate levels, hierarchical forecasting provides methods for reconciling forecasts across levels; it does not replace the need to choose appropriate segmentation rules.
Forecast accuracy measures also need care: percentage errors can be undefined or misleading when actual demand is zero, a common issue for intermittent-demand forecasting. Use measures suited to the series and decision, and examine bias and stock outcomes alongside forecast error.
Policies can conflict across objectives. For example, increasing availability for every segment may raise working capital and obsolescence, while concentrating stock can lengthen delivery times. Supply chain design therefore requires trade-offs among responsiveness, cost, and uncertainty. Segmentation is most useful when those trade-offs are explicit, classes are not treated as having universal cutoff percentages, and the results are reviewed against actual outcomes.