Variability analysis and decision-based segmentation
6 min read
Lesson objectives
- Calculate spread and compare demand regularity without asserting an untested distribution.
- Rank consumption value while separating criticality from financial categories.
Please sign in to use all the tools
Once you sign in, the reading tools are yours: larger or smaller text, typeface, line spacing, focus mode, highlighting and bold, and notes saved to your account and gathered in My study.
Two items can both average ten units of weekly demand while behaving very differently. One sells in most weeks; the other has quiet weeks followed by large bursts. Buying the same quantity for both merely because their means agree ignores how demand arrives. Describe volume, spread and intermittency, then connect those descriptions to a particular decision. Segmentation might direct manual forecast review, identify where to qualify an alternative supplier, or focus working-capital attention. Those are different tasks, so a single category should not automatically determine every policy. Start with equal-length periods, consistent units and a demand definition that distinguishes genuine requests from sales constrained by stock availability.
The mean is the sum of observations divided by their count. For a sample of n periods, sample variance is the sum of squared deviations from the mean divided by n−1. Standard deviation is its square root, returning spread to the original units. NIST's statistical handbook gives these scale definitions. For a positive mean, a coefficient of variation describes standard deviation relative to that mean. This ratio can compare relative spread across scales, but it neither establishes a normal distribution nor directly specifies safety stock. Lead time, the service objective and changes in the generating process still matter.
S is more variable within this window, but four weeks do not reveal next year's demand pattern. Investigate whether the two bursts arose from project orders or recurring consumption and whether the window covers only one season. Sequence matters beyond standard deviation. Rearranging S as 0, 20, 0, 20 preserves its mean and variance while changing the observed timing pattern. Keep the ordered series rather than retaining only two summary columns. When the mean is zero, do not manufacture a coefficient of variation by dividing by zero or replacing the result with an arbitrary large number. Route the item for a demand-state review and preserve the reason the ratio is undefined.
Intermittency can be decomposed into occurrence and size when demand occurs. For S, positive demand appears in two of four weeks, giving an observed occurrence rate of 0.5. Mean size in positive weeks is 20. Their product, 0.5×20, reproduces the overall mean of ten. Dropping zero weeks and forecasting 20 every week would double the observed rate without justification. Keeping only ten would conceal the burst pattern. This descriptive decomposition does not prove next week's occurrence probability or size: requests may follow a maintenance schedule or a project. Missing records also prevent reliable occurrence measurement, because an absent row may mean absent recording rather than zero demand.
Spend segmentation answers a different question: where is consumption value concentrated? Use annual consumption quantity multiplied by unit cost in this case, not sales revenue or the current stock valuation. Rank items from largest to smallest and calculate individual and cumulative value shares. Teams may choose local A, B and C boundaries for a stated purpose; an 80/20 pattern is not a law guaranteeing every portfolio's shape. State whether the item crossing a boundary belongs to the higher category or the next one, because that convention changes classifications near the threshold. Also reconcile currencies and costing dates before combining values from different systems.
| Item | Annual consumption | Unit cost | Consumption value | Approximate share |
|---|---|---|---|---|
| A | 400 | 50 | 20000 | 64.31% |
| B | 1000 | 10 | 10000 | 32.15% |
| C | 100 | 10 | 1000 | 3.22% |
| D | 1 | 100 | 100 | 0.32% |
| Total | — | — | 31100 | 100% |
Make a segment lead to a specific action rather than merely attaching letters to items. For high-spend, stable-demand A, review order size and working capital. For intermittent S, investigate the events producing demand and avoid unsupported forecasts between bursts. For critical D, check whether the spare is usable and how quickly an alternative could be obtained. These are candidate priorities requiring operational confirmation, not automatic outcomes of an A/B/C label. A top spend category does not imply that every associated customer deserves higher service, and a low spend category does not eliminate the need for stock accuracy. Consider monitoring effort, error consequences and replacement time before setting review frequency.
Test sensitivity before adopting the rule. Doubling D's cost after an emergency purchase raises its annual value from 100 to only 200. It may remain last even though its sourcing problem has worsened, demonstrating that spend alone is an incomplete risk signal. Annualizing a promotion week by multiplying it by 52 can similarly inflate a temporary item without sufficient support. Use an appropriate actual year, or label a projection and explain why its observation window represents future consumption. Do not rank some items by annual consumption and others by monthly consumption in the same table. Consistency of exposure matters as much as arithmetic accuracy.
Lesson summary
Mean, spread, intermittency, spend and criticality are different axes; connect each with a decision and its evidence limits.