XYZ analysis classifies references not by what they weigh, but by how they behave. The criterion is the coefficient of variation of demand, that is, the standard deviation over the mean. X items are regular and predictable, Y variable but readable, Z erratic. This dimension is independent of value, and it determines what can reasonably be expected of a forecast.
Why it matters
Classification by value answers one question: where the financial stakes lie. It leaves untouched the next question, which nonetheless drives most technical choices: is this reference predictable? Two items can carry exactly the same annual figure and yet call for opposite treatment, depending on whether their demand arrives steadily or in unpredictable bursts.
XYZ analysis measures precisely that. The coefficient of variation relates the dispersion of demand to its average level, which makes it comparable across references of very different sizes. A low value signals consumption that repeats; a high value signals demand with no stable rhythm, on which any fine forecast will be illusory.
The first panel of the signature shows what those figures actually cover. The three profiles share the same mean, and yet only the first can be anticipated with confidence. The third swings so much that the mean, while remaining mathematically correct, no longer describes anything of operational reality.
This makes the dimension a technical prerequisite, not merely one more segmentation. It says on which references to invest in forecast quality, on which to give up forecasting and protect yourself another way, and which belong to methods designed for intermittent demand.
The financial consequence is direct and reads in the third indicator of the signature. Safety stock is proportional to the standard deviation of demand, hence, at equal mean demand, proportional to the coefficient of variation. A Z reference does not need slightly more buffer than an X reference: it needs several times more, to hold exactly the same service level.
This overturns a common intuition. Many organisations concentrate forecasting effort on high-value references, which seems logical, but the effort pays off best where variability is reducible. On an X reference, improving the forecast lowers the buffer; on a Z reference, no model will make an irregularity that comes from demand itself disappear, and money spent on statistical sophistication would sit better in shorter lead times or pooling.
The expert lesson is therefore to treat regularity as a property to exploit rather than a mere descriptor. An X reference allows larger lots, more spaced-out reviews and thin buffers at no risk. A Z reference imposes the opposite: give up precision, accept more modest service, or pay dearly to hold it. Confusing the two regimes costs capital one way and service the other.
The mechanism
The calculation is direct: on clean history and at a relevant time bucket, estimate the mean and standard deviation of demand, then divide the second by the first. The result is a dimensionless number, which allows a reference sold by the thousand to be compared with one sold by the unit. That comparability is what makes the indicator useful.
The choice of time bucket deserves attention, because it changes the result. Demand that is perfectly stable monthly can look highly irregular daily. The bucket chosen must therefore be that of the decisions the classification feeds: if you replenish weekly, weekly variability is what counts, not daily variability.
Thresholds between classes are not standardised, and that is a good thing. Common markers place the X boundary around half a unit and the Z boundary around one, but the second panel of the signature shows why they must be calibrated on your own portfolio: it is the shape of the distribution, not a convention, that indicates where genuine breaks in behaviour lie.
One last technical point, often neglected. The coefficient of variation loses meaning when demand contains many zero periods, which is common on spare parts. In that case a reference can show a high coefficient without being genuinely erratic, simply because it is rare. Those profiles belong to methods dedicated to intermittency rather than to a direct reading of the ratio.
The traps
A coefficient computed daily on a chain that replenishes weekly describes variability nobody experiences. The bucket must be that of the decision, otherwise the classification files perfectly manageable references under Z and inflates their buffers for nothing.
On references that move only a few times a year, the coefficient of variation becomes high by construction, without demand being genuinely unstable. Confusing them with erratic references leads to oversized buffers when the problem belongs to intermittency-specific methods.
Concentrating modelling effort on the most irregular references is an allocation error. Part of their variability is irreducible and will yield to no model. Forecasting effort pays on X and Y items; on Z items, lead time, pooling and decoupling are what deliver.
The rollout
Compute the coefficient at the time bucket where decisions are actually taken, day, week or month. This choice precedes everything else and conditions the validity of the ranking.
Remove stockouts, exceptional promotions and unrepresentative orders before measuring. An unexplained spike inflates the standard deviation and wrongly downgrades a regular reference.
Spot references with many zero periods and handle them separately, with methods designed for intermittency, rather than applying a direct reading of the ratio.
Plot the portfolio distribution before fixing boundaries, and place them where behaviour genuinely changes rather than at conventional values.
Attach a policy and a forecasting effort to each class: careful models and thin buffers on X, protection and simplicity on Z. Without that translation the measurement has no effect.
Neighboring concepts
From knowledge to action
Regularity governs the buffer, the choice of model and the service achievable. Our Inventory & distribution file measures your real distribution and allocates forecasting effort where it pays.