5. Forecasting Hierarchies

Teaching rewrite — the Objective step: at what level of detail should you forecast? Original wording; numbers are illustrative.

Introduction

Many supply chains — especially manufacturers — forecast 18 months ahead, per country, per month. Is that a deliberate best practice or just a default nobody questioned? The level of detail of a forecast should match the decisions it supports. When it doesn’t, planners spread country forecasts across warehouses with flat splits and stock ends up in the wrong place.

This lesson covers:

The three forecasting dimensions: materiality, geography, temporality (plus horizon)
Moving between levels: bottom-up, top-down, middle-out
Which level to focus on (the decision level)
Which level to build the forecast from (not necessarily the same)

5.1 Three dimensions

Dimension	Examples of levels	Units of measure
Materiality	SKU → product → segment → brand → total	Units, value, weight, raw material needed
Geography	Store/zip code → warehouse → region → country → world; also channel, customer segment, B2B vs B2C
Temporality	Day → week → month → quarter → year

The fourth question — how far ahead (horizon) — is the subject of lesson 6.

A coffee shop could forecast at very different levels:

27 lattes tomorrow (product × day)
5 gallons of milk next week (raw material × week)
$4,800 of drinks next month (value × month)

5.2 Zooming in and out

Technique	How it works	Example
Bottom-up	Forecast the detailed level; sum upward	Sum store forecasts to get a country forecast
Top-down	Forecast the aggregate; split downward with rules	Split a monthly forecast into days evenly (flat split), or a national forecast into stores by historical revenue share
Middle-out	Forecast an intermediate level; sum up and split down	Forecast per country; sum to regions; split to stores

All three keep forecasts reconciled: the levels add up.

flowchart TB
  T[Total / Brand / World / Year]
  M[Product / Country / Month]
  B[SKU / Store / Week]
  B -- bottom-up: sum --> M
  M -- sum --> T
  T -- top-down: split --> M
  M -- middle-out: split --> B

Worked example: top-down split

A country forecast for next month is 9,000 units. Three warehouses serve the country:

Warehouse	Historical demand share (from geographic footprint)	Flat split	Share-based split
North	50%	3,000	4,500
Centre	30%	3,000	2,700
South	20%	3,000	1,800

A flat split sends 1,500 too few units North and 1,200 too many South — a guaranteed misallocation.

5.3 Which level should you focus on?

With unlimited data, time, and computing power you could forecast everything at the finest grain and add it up. In practice:

Review time and computing power are limited.
There is an accuracy glass ceiling: at very fine grain (product × customer × store × hour) noise swamps signal. If an item sold once last year — in December — was that seasonality or luck? Aggregated data reveals patterns that granular data hides.

So pick a focus level — and pick it from the decisions:

Your supply chain makes decisions at specific levels. Forecast demand at those levels.

Decision	Focus level
What to ship from the plant to each regional warehouse	Product × warehouse — based on each warehouse’s geographic footprint, not its historical shipments
Production set-up by packaging format	Packaging level, discussing what drives each format (events, promotions)
Single warehouse supply	Possibly one global forecast is enough; per-channel detail may not repay the effort
Items with very different growth rates or components	Forecast them separately rather than splitting a total
Make-to-order business	Sub-components, mid/long term
Make-to-stock business	Granular, short term

Pro note — forecasting per warehouse

Don’t forecast a warehouse from its shipments. If warehouse A is out of stock, warehouse B (the next closest) serves A’s customers — A’s shipments understate its demand and B’s overstate it. Track demand by the geography the warehouse should serve (e.g. zip codes). This also survives warehouse openings and closures.

Red flag

Hearing “we use a flat split to turn the monthly forecast into weeks” or “we split the country forecast across warehouses evenly” is a sign the forecast level doesn’t match the decision level.

Checkpoint: dimensions and focus

Quiz

Checkpoint — sections 5.1 to 5.3

1 / 4

Which are the three forecasting dimensions?

Pick an answer to see if you are right.

5.4 Which level should you build the forecast from?

The level you care about is not necessarily the level at which you should compute the forecast. A model run on weekly-by-channel data gives a slightly different answer than one run on monthly-by-market data — different history, different information.

flowchart LR
  D[Daily data<br/>school holidays, strike, weather, match day] --> W[Weekly forecast<br/>= level of interest]

Ice-cream truck: you restock every weekend for the coming week, so the weekly forecast is what matters. But building it from a daily forecast lets you use daily information: schools closed Monday, a strike Tuesday, rain on Wednesday (normally the big after-school day), a match on Friday. Summing the daily forecasts can give a more accurate weekly number.

Worked example: daily build-up

Day	Normal daily demand	Known event	Adjusted forecast
Mon	40	School closed (−50%)	20
Tue	40	Strike (−25%)	30
Wed	60	Rain (−50%)	30
Thu	40	—	40
Fri	40	Match day (+50%)	60
Week	220		180

A weekly model that ignores the daily events predicts ~220; the daily build-up predicts 180.

Review frequency

Updating forecasts more often uses new data sooner and can improve accuracy — but too often causes over-reaction (“nervous” plans) and eats planner time.

Test, don’t assume

Granular data in a granular model is not guaranteed to be better. Try several levels and models, measure accuracy at the level of interest, keep the winner — and re-test when data or the business changes.

Checkpoint: building the forecast

Quiz

Checkpoint — section 5.4

1 / 4

You need a weekly forecast. Should the model always run on weekly data?

Pick an answer to see if you are right.

Rank: hierarchies in practice

Quiz

Hierarchy decisions

1 / 2

Rank these from MOST to LEAST noisy (lowest signal-to-noise first).

OrderDrag the handle to rearrange — top should be first.

1Product family × country × month
2SKU × store × day
3SKU × warehouse × week
4Brand × world × year

Drag items into the correct order (or use ↑↓), then check.

Check: forecasting hierarchies

Quiz

Forecasting hierarchies

1 / 5

Summing store forecasts to get a national forecast is…

Pick an answer to see if you are right.

Takeaways

Remember

Forecasts live in three dimensions — materiality, geography, temporality — plus a horizon
Move between levels with bottom-up, top-down, middle-out; keep them reconciled
Focus on the level where decisions are made — not on the habitual “country × month”
Forecast warehouses by geographic footprint, not shipments
The level you build the forecast from can differ from the level you care about — test and measure

Introduction​

5.1 Three dimensions​

5.2 Zooming in and out​

Worked example: top-down split​

5.3 Which level should you focus on?​

Checkpoint: dimensions and focus​

5.4 Which level should you build the forecast from?​

Worked example: daily build-up​

Review frequency​

Test, don’t assume​

Checkpoint: building the forecast​

Rank: hierarchies in practice​

Check: forecasting hierarchies​

Takeaways​

Further Reading​

Introduction

5.1 Three dimensions

5.2 Zooming in and out

Worked example: top-down split

5.3 Which level should you focus on?

Checkpoint: dimensions and focus

5.4 Which level should you build the forecast from?

Worked example: daily build-up

Review frequency

Test, don’t assume

Checkpoint: building the forecast

Rank: hierarchies in practice

Check: forecasting hierarchies

Takeaways

Further Reading