XYZ analysis: classify demand without hiding zero periods

Keep the observation window and missing-data rules beside each class because the same average demand can conceal steady use, seasonality or long gaps.

Kartik Deshpande··6 min read

Two parts both average ten units of demand a week, but one moves every week while the other moves in occasional batches. XYZ analysis separates those patterns by comparing the standard deviation of demand with its mean over a declared window. Calculate that coefficient of variation for each item and location, apply documented class boundaries, and review zeros, seasonality and stockouts before using the class to change replenishment.

Use XYZ for variation in consumption, alongside the value priorities in ABC inventory analysis. An expensive item with erratic demand needs a different planning conversation from an expensive item used at a steady rate. Neither label determines a purchase quantity on its own, and a class computed from incomplete issues can misrepresent demand that the warehouse could not fulfill.

Define what one observation means

Choose the item, location, base unit, demand measure and time bucket before exporting data. Weekly customer orders, daily production consumption and monthly shipments describe different processes. Do not combine them into one series without explaining the transformation. Keep the extraction time and source filters so another planner can reproduce the result.

SAP’s variability classification documentation describes consumption from goods issues by calendar day. That is one specified implementation, not a requirement to use calendar days for every analysis. If your plant reviews weekly buckets, document that choice and keep shutdown weeks visible under the same rule each time.

Build a complete calendar for the selected window and join each item to it. A confirmed week with no consumption becomes zero; a missing source extract stays missing. Preserve an availability flag where stockouts may have prevented issues. Otherwise the calculation can interpret a supply interruption as low demand and recommend a stock policy based on the wrong cause.

Check pack and unit changes inside the history before aggregating periods. A case-to-each conversion needs the factor that applied to the transaction, with the original quantity retained. If that factor is unknown, isolate the affected periods. Treating an old case count as an each count can create a demand jump caused by the unit mapping rather than by consumption.

Calculate variation on a complete, comparable series

NIST defines the coefficient of variation as standard deviation divided by the mean. Use one declared standard-deviation convention across the cohort. A population calculation describes the selected observations; a sample calculation uses a different denominator and gives a different result. Record which convention you used instead of accepting an unexplained spreadsheet default.

Assume these four weeks are the complete population for an illustrative calculation, with all demand recorded in units. Part A has demand of 10, 10, 10 and 10: its mean is 10, standard deviation is 0 and coefficient is 0. Part B has 0, 20, 0 and 20: its mean is also 10, population standard deviation is 10 and coefficient is 1. The same total of 40 units gives different variability.

Illustrative itemWeekly unitsMeanPopulation standard deviationCoefficient
Part A10, 10, 10, 101000
Part B0, 20, 0, 2010101

Removing the zero weeks from Part B leaves 20 and 20. The shortened series has zero variation and a different mean, so it answers how large active weeks are rather than how demand behaves across time. Retain genuine zeros in the full series. Four weeks here explain the arithmetic (they do not establish a sufficient planning history for a seasonal product).

A zero mean makes the ratio undefined. Route all-zero items to a separate review instead of dividing by zero or assigning X because the standard deviation is also zero. Keep new items, discontinued items and negative net-consumption series in explicit exceptions. Returns may need separate treatment before the calculation, with the original transactions preserved for review.

Set class boundaries that remain visible in the result

X represents lower variation, Y intermediate variation and Z higher variation within the chosen policy. SAP’s documentation allows thresholds for variability classification; there is no single boundary that makes all businesses comparable. State each interval, its inclusive endpoint and the effective policy version. The reader should be able to classify a coefficient exactly on a boundary without guessing.

For an illustrative policy, assign X at or below 0.25, Y above 0.25 through 0.50, and Z above 0.50. Under that declared rule, Part A is X and Part B is Z. These example boundaries are not a recommendation for your inventory. Review the resulting groups against forecast errors, replenishment constraints and service outcomes before adopting a policy.

Store the mean, standard deviation, coefficient, number of buckets and zero-bucket count beside the class. Include the period and source measure in exports. A bare Z label cannot tell you whether an item has annual seasonality, a short history, sporadic orders or a data gap. Those causes require different decisions even when the coefficient is similar.

When a refresh changes a class, retain the previous series and threshold version. Recalculate the earlier series under the new thresholds to identify a policy-driven change, then compare the refreshed series under that same policy. Investigate corrected source rows or a shifted window as separate inputs to the comparison. If several inputs changed together, report those changes instead of attributing the entire class movement to demand. The planner can then choose whether the evidence warrants a replenishment review.

Keep the evidence beside the demand class
Series

Measure, calendar, units

Calculation

Mean, variation, policy

Review

Zeros, stockouts, seasonality

A class becomes reviewable when its calendar and exceptions travel with it.

Check a class against its demand history

Bring one approved consumption extract and the calendar used to classify it.

Use the class to choose the next investigation

Inspect the time series before changing safety stock. A repeating seasonal pattern can have a high coefficient across the full year while remaining predictable at the correct season. Sporadic demand has different gaps and order sizes. SAP’s XYZ strategy discusses zero periods and their relevance to sporadic consumption. Keep that distinction in the planning review rather than treating every Z item as equally unpredictable.

Combine the variability class with the existing value class only after confirming matching item keys, locations and windows. An AX label can prioritize a high-value, steady-use item for close replenishment control. A CZ label can trigger a review of buying frequency, minimum order quantities and the cost of holding intermittent stock. These are review paths, not automatic stocking instructions.

For items that need a buffer, the safety stock calculation also needs the demand and lead-time assumptions appropriate to the policy. A coefficient alone omits lead-time uncertainty, supplier constraints and the consequence of a missed order. Review those inputs with the service owner before releasing a changed parameter to the ERP.

How Permute keeps classification evidence reproducible

Permute is the connected-data platform we sell for applications and analysis across business records. We can connect approved NetSuite records and Google Sheets planning extracts, reconcile item and location identities, and preserve links back to the source evidence. Unmatched records and missing periods remain exceptions for the planner to resolve.

We define the consumption measure, calendar, standard-deviation convention and class thresholds as declared business rules. An inventory planner can ask which items changed class and inspect whether the change came from demand, a new window or a revised rule. The answer can retain the underlying series and extraction time so another reviewer can reproduce it.

Keep forecasting and stock-policy approval separate

This analytical workflow does not establish a forecasting model, choose service targets or authorize a purchase. Access depends on the connected sources and granted permissions. Missing historical demand or availability evidence stays unknown, and any change to operational stock parameters needs its own supported, approved execution path.

Accept a class only with its calculation context

XYZ analysis is useful when each class preserves the demand series, calendar and threshold policy that produced it. Keep genuine zeros, isolate missing records and review seasonality before interpreting a high coefficient. Combine the class with value, lead time and service consequences when deciding what to investigate. Release replenishment changes only after the owner has reviewed those inputs.

Review the evidence behind your inventory classes

Trace one disputed class through demand records, calendar choices and stock-policy assumptions.

Questions about classifying variable demand

How often should XYZ classes be refreshed?

Choose a review cadence that matches the planning cycle and the rate at which demand changes. Preserve the prior class and policy version when refreshing, so a seasonal change can be distinguished from a data correction. A new classification is a reason to review the item, not automatic approval to change its stock settings.

Can warehouse-level classes differ from a company-wide class?

Yes. Combining locations changes the demand series and can smooth variation that remains visible at one warehouse. Compute at the level where the replenishment decision is made, and label any company-wide rollup with its included locations.

What should happen to a new item with little history?

Keep the short history visible and use an explicit new-item exception until the planning owner accepts the available evidence. Substituting a similar item may support a planning assumption, but that assumption should remain separate from the new item’s observed coefficient.