← Back to Stories

Story · Case study

A demand forecast the planners are allowed to argue with

A demand forecast is only useful if the person who disagrees with it can say what they disagree with. Northwind's planners can now point at the driver.

The company

Company
Northwind Foods
Industry
Consumer goods

The starting point

Northwind Foods plans chilled and ambient lines across three regions. The demand forecast was a spreadsheet with eleven contributors and one author, and its authority was entirely personal: it was right because the person who built it was usually right.

That is a fine arrangement until the week it disagrees with a regional planner. Then the argument has nowhere to go, because the assumptions are inside formulas rather than on the screen, and the only available move is to overwrite the number.

What changed

What they described

Northwind's planning lead wrote down what a forecast is in their business: which drivers move a line, which of them are seasonal, which promotions count as demand and which as pull-forward, and what a planner is allowed to override. The last clause turned out to be the important one — an override is a first-class input, not a correction applied afterwards.

The application that came back puts the drivers on the screen, names each one, and keeps every run's inputs beside its outputs in Northwind's own store. A planner who disagrees with the number opens the run and points at a driver. That is the whole of the change, and it is the reason the forecast is now used by people who did not build it.

The results

What did not change

The forecast is still wrong sometimes, and Northwind is careful about saying so. Six per cent mean absolute error on a chilled range is a good number and it is still a number that puts stock in the wrong region most weeks. What changed is the speed of the correction: a planner who spots it on Wednesday can re-run with a different assumption and see the decomposition by Wednesday afternoon, rather than waiting for the author to have a free morning.

Example

SKUs forecast each run

1840

Named drivers on screen

7

each one overridable

Mean absolute percentage error

6.2%

from 9.8% on the workbook

Hours to a published plan

4 hrs

“The argument used to be about whose number was right. Now it is about which assumption is wrong, which is an argument we can actually finish.”

— Tomas Lindqvist, Head of demand planning, Northwind Foods (an example business)

The apps

What it cost to run and what to expect

The part the IT team cared about

Two things, neither of them about forecasting. The first is that the source extracts land in Northwind's own tenant space and stay there — the application reads what it is given and writes what it produces into the same place, and no ledger extract is copied somewhere Northwind cannot point at. The second is that access is granted per person on the application itself: eleven planners can open it, four can publish, and the regional analysts who kept asking for the workbook can read a run without being able to change one.

Connected to this story

Tell me when it's ready

We keep your email only to tell you once, when the app is ready, and delete it then.

What happens next

Thank you. We will write once, when it's ready.