Inventory segmentation (ABC-XYZ)

A real café stocks hundreds of things: beans, milk, cups, syrups, napkins, a rare single-origin coffee. Nobody can tune every item by hand, and they don’t all deserve the same care. Running out of milk is a disaster; running out of cinnamon is a shrug. Segmentation sorts items into groups so each group gets a sensible policy.

ABC sorts by money. Rank items by annual value (usage × unit cost). A handful of items usually carry most of the value: the A items (roughly the top 70–80% of value), then B, then the long tail of C items that are many in number but small in money.

XYZ sorts by predictability, using the coefficient of variation, $CV = \sigma / \text{mean}$. A common cut-off is X for CV below 0.5 (steady, like milk), Y for 0.5 to 1 (some swings), and Z above 1 (erratic, like the rare coffee that sells 0 kg one week and 5 kg the next).

Crossing the two gives a 3 × 3 grid, and the corners tell you what to do:

• AX (valuable, steady): run lean with frequent review and high service. Predictability keeps the buffer small anyway.
• AZ (valuable, erratic): the danger zone. Put your best forecasting effort here, talk to customers, or make to order.
• CX (cheap, steady): automate it. Order big batches rarely and keep a generous buffer; it costs almost nothing.
• CZ (cheap, erratic): question whether to stock it at all. Make to order or delist, because this is where dead stock piles up.

Step 1: ABC by annual value

1. For each item, compute annual value = annual usage × unit cost.
2. Sort from highest to lowest and compute the cumulative share of total value.
3. Draw the lines. A common split is A up to about 80% of value, B up to about 95%, and C for the rest. The cut-offs are a choice, not a law.

Step 2: XYZ by predictability

For each item, compute the coefficient of variation of demand over a sensible time bucket, usually weeks or months: $CV = \sigma / \mu$. Then group: X below 0.5, Y from 0.5 to 1, Z above 1.

Step 3: a policy for each cell

The grid decides more than safety stock. It sets the forecasting method, how often each item is reviewed, who looks at it, and whether it should be stocked at all.

The service-target paradox

A common rule gives A items the highest service targets because they matter most. Cost logic often says the opposite. Each extra point of service on an A item needs expensive safety stock, while a C item can be pushed to 99.5% for very little. Customers count missing napkins as a failed order too. So a sensible policy often sets higher targets for cheap items and slightly lower, tightly managed targets for expensive ones, reaching the same overall fill rate with less money. Research on ABC and service levels (Teunter, Babai and Syntetos, 2010) found that ranking items by a cost criterion that includes shortage cost and order size, instead of by value alone, lowers total inventory cost.

Measure variability properly

• Use forecast error, not raw demand. Ice cream has a high CV because of its summer peak, yet a seasonal forecast predicts it well. The CV of forecast error (RMSE divided by mean demand) measures how surprising an item is, which is what buffers protect against.
• Mind the time bucket. Weekly demand always looks more variable than monthly demand. Use the bucket that matches your review period and lead time.
• Intermittent items need two measures. For items with many zero periods, the Syntetos–Boylan scheme uses the average interval between demands (ADI) and the squared CV of the non-zero demand sizes ($CV^2$), with cut-offs at 1.32 and 0.49. That sorts items into smooth, erratic, intermittent and lumpy, each with a suitable forecasting method.

Other ways to classify

• VED (vital, essential, desirable) ranks by criticality. It is used for spare parts and hospital supplies, where a cheap part can stop a production line.
• FSN (fast, slow, non-moving) ranks by movement, which helps find dead stock.
• Lifecycle stage: launch, growth, maturity and phase-out items need different forecasting and buying rules.
• Margin or customer: ABC on profit or on strategic customers instead of on cost.

Keep it alive

Re-run the classification regularly (quarterly is common), but add a tolerance band so items near a boundary don’t flip classes every quarter and send planners chasing changes. Then check the result with common sense: a C item that is the only part keeping a key customer’s line running is really an A.