Planning on a number nobody believes
Most planning systems use a lead time held in a supplier master, entered when the supplier was onboarded and rarely revisited. Everyone knows some suppliers run late, and that knowledge lives with individual planners rather than in the system.
The data to do better is almost always already there: purchase order dates, promised dates and goods receipt dates, going back years.
What drives variation
- The supplier and the specific item - lead times vary hugely within one supplier's range
- Order size, which can trigger a different production run or shipping mode
- Time of year, including factory shutdowns and national holidays in the source country
- Whether the item was in stock or made to order
- Route and mode, where these differ between orders
- How busy the supplier is, which sometimes shows in their recent performance
Seasonal shutdowns are worth special mention because they are knowable in advance and frequently absent from planning systems. A supplier closing for two weeks is not a prediction problem; it is a calendar nobody entered.
Predict a distribution, not a date
A single predicted date is the wrong output. What planning needs is the shape: most likely arrival, and how bad the tail is.
Illustrative example: a supplier whose orders arrive between 18 and 24 days with most at 20 needs different safety stock from one averaging 20 days but occasionally taking 45. The averages match; the risk does not.
| Output | Use |
|---|---|
| Median lead time | Normal planning |
| 80th percentile | Safety stock sizing |
| 95th percentile | Critical items, risk review |
| Spread over time | Supplier performance conversations |
Using it in supplier management
A by-product of this work is an evidence base for supplier reviews. 'Your quoted lead time is 21 days and your last two years average 29, with a quarter of orders beyond 35' is a different conversation from a general complaint.
It also supports dual-sourcing decisions on the right grounds - reliability rather than only unit price. A cheaper supplier with a long tail can cost more once the safety stock it requires is counted.
A quoted lead time is a promise. Your receipt history is evidence.
Where to be careful
Receipt dates can be misleading. A receipt booked in on Monday for goods that arrived Friday adds noise, and a partial delivery raises the question of what counts as arrival.
Decide these definitions before modelling and apply them consistently. As with most business data problems, the definitions matter more than the algorithm, and getting them wrong quietly biases everything downstream.