Skip to content
Predict.ai

Decisions

Lead times: the delay is the series

Suppliers, customs, labs, and work orders all have a duration. That duration moves. Forecasting lead time is how you stop treating '14 days' as a constant in an MRP that has not been true since March.

Updated Aug 26, 2026·7 min read

14 days is a rumor

Suppliers, customs, labs, and work orders all have a duration. That duration moves. Forecasting lead time is how you stop treating a frozen cell in MRP as physics. A perfect demand forecast with a fantasy lead time still stocks out.

'14 days' is a rumor

Supplier lead time with a tail eventfrozen 14
Median lead time is a promise. The tail is the buffer. A frozen 14 in MRP will stock out the week a lane slips — even if demand was perfect.

The tail sets the buffer

Median lead time is a promise date. A high quantile is the buffer. Same logic as demand p90. If you only track the average, you will be surprised by the week a lane slipped — and you will blame demand.

p10 · p50 · p90

Prediction interval fan chartp90p50p10
The band is the plan. It should widen as you look further out. A skinny interval at week 12 is usually a model that has not been allowed to be unsure.

Port, season, vendor

Volume on a lane, a holiday in the origin country, a particular vendor, freight rates: these can lead a slip. They are drivers if you will have them at origin. A delay you only see when the ASN fails is not a feature. It is the target.

Duration as a signal

Ingest actual lead times as a signal. Predict.ai can find what leads a slip and keep a quantile on the duration — the number MRP should have been using instead of a frozen 14. Alerts when realized lead time leaves the band are how procurement notices the sneak before the stockout.

FAQ

Should I forecast demand or lead time first?
Both, if both move. A perfect demand forecast with a fantasy lead time still stocks out. Many teams freeze lead time because it is in a table. Unfreeze it.
Is median lead time enough?
For a promise date, maybe. For a safety buffer, no. You want a high quantile of duration, the same way you want p90 on demand.

Keep going

Try it on your data.

Connect an outcome and its history. Predict.ai finds the drivers, races the models, and keeps the forecast live — in the workspace, over the API, and through agents.