Why the average is the wrong target
A forecast that predicts the average is right on average and short about half the time. For anything where running out has a cost, that is not the number to order to.
The usual workaround is an average plus a safety percentage. That works where uncertainty is uniform, and it is not - some products are far more variable than others, and a flat 20% over-orders the predictable ones while still failing on the erratic ones.
Predicting a percentile directly
Quantile forecasting trains the model to predict a chosen percentile rather than the middle. Ask for the 90th percentile and it predicts the level actual demand will fall below about nine times in ten.
The advantage over a flat buffer is that the model learns how much extra each item needs. A stable product gets a small uplift; a volatile one gets a large one, automatically, from its own history.
| Percentile | Meaning | Suits |
|---|---|---|
| 50th | Middle - short half the time | Where shortage costs little |
| 80th | Covered four times in five | General stock holding |
| 95th | Covered nineteen times in twenty | Critical items, high shortage cost |
| 20th | Deliberately conservative | Perishables, where excess is worse |
Choosing the percentile from costs
The percentile is a business decision derived from the ratio between the cost of being short and the cost of holding excess. Where shortage costs far more, aim higher; where excess spoils or must be marked down, aim lower.
Illustrative arithmetic: if being short costs roughly four times what holding a spare unit costs, an 80% service level is a reasonable target. The exact figure matters less than the reasoning being written down and owned by someone.
It can also differ by item. Critical spares justify a much higher percentile than ordinary consumables, and letting the percentile vary by category is usually more valuable than improving the underlying forecast.
A whole distribution is better than one number
Predicting several percentiles gives the shape of the uncertainty. That supports conversations no point forecast can: what would it take to be safe nineteen times in twenty, and what does that cost in stock?
It also exposes asymmetry. Demand that is usually modest but occasionally very large produces percentiles far apart at the top end, which is exactly the situation where an average plus buffer fails worst.
Presenting it without confusing people
Percentiles confuse if presented as forecasts. 'The 90th percentile forecast is 340' invites the question of what happened to the real forecast.
Frame it as a decision instead: 'order 340 to cover demand in about nine weeks out of ten; 280 covers about seven in ten'. That is immediately usable and makes the trade-off visible to whoever owns the money.
Order to the average and you will be short about half the time. That is arithmetic, not bad luck.