Demand Classification for Forecasting

By MLAIA Data Science Ltd. · Published 13 July 2026

Demand classification is the practice of segmenting an item portfolio by the statistical character of each item's demand, so that the forecasting method, review cadence, and inventory policy can be matched to the pattern rather than applied uniformly. In portfolios with thousands of SKUs — the normal case in spare parts and distribution — classification is what makes item-level forecasting operationally feasible.

Why classify before forecasting

Forecasting methods embed assumptions: exponential smoothing assumes demand in every period, ARIMA assumes Gaussian errors, Croston-family methods assume intermittency, machine-learning models assume enough signal to learn from. Applying one method to a whole portfolio guarantees a poor fit somewhere. A classification step assigns each item to a segment with a known-appropriate default method and a sensible level of planner attention.

ABC analysis: classifying by value

ABC analysis ranks items by annual usage value (unit cost × annual volume) and exploits the Pareto concentration typical of inventories: a small fraction of items accounts for most of the value. Conventionally, A items (roughly the top 70–80% of value) receive tight control and frequent review; B items intermediate treatment; C items (the long tail, often half the SKUs but a small share of value) simple, automated policies. ABC says where control effort pays off — but nothing about how hard an item is to forecast.

XYZ analysis: classifying by variability

XYZ analysis segments items by the stability of their demand, typically using the coefficient of variation of demand per period: X items are stable and predictable, Y items fluctuate (trend, seasonality), and Z items are irregular or sporadic. Combined as a 3×3 ABC–XYZ matrix, the two views separate, for example, high-value stable items (AX — forecast tightly, hold lean buffers) from high-value erratic items (AZ — forecast humbly, protect with stock or lead-time agreements).

ADI–CV² in practice

For forecasting-method selection specifically, the ADI–CV² scheme is the standard: compute the average inter-demand interval (ADI) and the squared coefficient of variation of non-zero demand sizes (CV²) per item, and cut at ADI = 1.32 and CV² = 0.49 into smooth, erratic, intermittent, and lumpy quadrants (Syntetos, Boylan & Croston, 2005). Practical notes:

• The classification depends on the time bucket: an item intermittent at weekly granularity may look smooth monthly. Classify at the bucket used for planning.
• Items drift between quadrants as demand evolves; reclassify on each forecast refresh and treat quadrant transitions (smooth → intermittent, or interval lengthening toward zero demand) as signals of lifecycle change and obsolescence risk.
• Very short histories classify unreliably; new items need a provisional treatment until enough periods accumulate.

Matching method to pattern

Machine-learning models (gradient boosting, neural approaches) earn their place when there is cross-item structure to pool — many related series, promotions, installed-base or causal data. On a single short, lumpy series they have little signal to learn and rarely beat SBA-class baselines; empirical forecasting competitions have repeatedly shown that strong simple baselines are hard to beat on intermittent data.

Forecastability

Classification also sets expectations. A series dominated by irregular, independent events has a floor on achievable accuracy no method can break. Chasing that noise with ever more complex models produces instability, not accuracy. The mature response is to measure forecastability (for example, by benchmarking against naive methods), accept the limit, and shift the burden to inventory buffers and service-level policy for the genuinely unforecastable segment.

Aggregate vs. item level

Aggregation trades relevance for stability. Demand summed across items, locations, or longer time buckets is smoother and easier to forecast — statistical noise cancels — but replenishment decisions are executed per item per location. Common practice is to forecast where the signal lives and reconcile: top-down (forecast the family, disaggregate by historical mix), bottom-up, or middle-out. Temporal aggregation is a further lever: an intermittent weekly series may become a well-behaved monthly one, and forecasting at the aggregate bucket then disaggregating (the ADIDA approach of Nikolopoulos et al., 2011) can improve intermittent-demand forecasts at no extra data cost.

References

• Syntetos, A. A., Boylan, J. E., & Croston, J. D. (2005). On the categorization of demand patterns. Journal of the Operational Research Society, 56(5), 495–503.
• Nikolopoulos, K., Syntetos, A. A., Boylan, J. E., Petropoulos, F., & Assimakopoulos, V. (2011). An aggregate–disaggregate intermittent demand approach (ADIDA) to forecasting. Journal of the Operational Research Society, 62(3), 544–554.
• Boylan, J. E., & Syntetos, A. A. (2021). Intermittent Demand Forecasting: Context, Methods and Applications. Wiley.

See also

• Demand Forecasting Wiki — all articles
• Intermittent and Lumpy Demand
• Croston's Method, SBA and TSB
• Safety Stock and Reorder Points
• Dead Stock and Excess Inventory

Related: NextDemand classifies each item's demand pattern automatically and selects the forecasting method per item.