Capabilities / Demand and supply-chain intelligence
A useful forecast carries its uncertainty, evidence and owner. The model proposes. The operating team decides what to finance.
Forecast contract / decision bench
Choose a planning situation, then choose how much of the fan to commit. The bench is an illustrative model. It produces no demand result.
Method state · New product
Use product attributes, price position, route to market and comparable launches. Record where each analogue applies and fails. Keep a range until orders provide a local signal.
ReadoutQuantile fan with analogue provenance and a review trigger.
FanWidest fan: no direct sales history.
Method state · Repeat style
Test a seasonal or recent-run-rate reference on identical origins. Add complexity only for repeatable gains. Treat promotions and stockouts as context.
ReadoutChallenger comparison with errors by segment.
FanNarrowest fan: history exists, and a simple reference stays beside it.
Method state · Replenishment
Combine demand quantiles with observed supplier, production and inbound variability. State the service policy before translating the fan into stock. Keep expediting and obsolescence visible.
ReadoutReorder case with a service assumption and an approval owner.
FanSkewed fan: late arrivals stretch the upper band.
Illustrative fan. Index 100 is the origin of the central path. Spreads are assumptions chosen to draw the fan, not data. The bench defines a brief. It does not estimate sales, prescribe purchases or record commitments.
Evidence before apparent precision
A new product lacks a direct sales curve, but not a researchable context. Transfer evidence carefully and show where the transfer breaks.
Cold-start evidence includes attributes, price architecture, channels, interviews, comparable launches and supply constraints. Each source needs provenance, a date and a population. Analogues require a transfer rationale, not copied sales curves.
US$37.3bn to US$49.41bn
US$94.70 and 8.0 pieces
20 to 50%
2025: US$37.3bn, one vendor's estimate of the Africa textile market (R24). Not GFT demand.
CaveatExternal estimates cannot identify the assortment, delivery promise or price points a manufacturer can sell. Regional growth is not factory utilisation. The cited annual growth rate is 4.73%. Not a GFT sales forecast.
Spend: US$94.70 per person on clothing and footwear in 2024. Context, not demand for any proposed product.
CaveatPer-person context is not an entrant's available share. Two measures on separate scales. Not demand for any proposed product or factory.
Reported reduction in forecast errors: 20 to 50%. Sector-level fashion AI evidence, not GFT or Cameroon outcomes.
CaveatRolling-origin tests must compare like with like. Sector-level fashion AI evidence, not GFT or Cameroon outcomes.
Annex qualifications remain attached. Market estimates are contextual. Fashion AI ranges are sector-level findings, not GFT or Cameroon outcomes.
From evidence to release
The pinned fan changes state as work proceeds. Each transition records evidence, comparison, adjustment and authority.
01 Frame · Decision boundary
Name the product, location, channel and horizon. Aggregation must serve the decision, not the available data. Write the loss function before model selection.
02 Form · Quantiles
Request central, lower and upper quantiles. Stockouts, expediting, idle capacity and obsolete cloth carry different losses. A point forecast hides that choice.
03 Challenge · Reference
Keep a naive, seasonal or moving-average reference. Compare out-of-sample errors at identical origins and horizons. Added complexity must earn a decision-relevant gain.
04 Reconcile · Coherence
Product, customer, region and total forecasts should not become unrelated negotiations. Align them into one total and record manual adjustments rather than burying them.
05 Commit · Quantile
Translate the chosen band into stock, capacity and cash cases. A named owner accepts the commitment and its review trigger.
01 Frame · Decision boundary
Name the product, location, channel and horizon. Aggregation must serve the decision, not the available data. Write the loss function before model selection.
P10 46 · P50 127 · P90 208, index at period 10 (illustrative)
02 Form · Quantiles
Request central, lower and upper quantiles. Stockouts, expediting, idle capacity and obsolete cloth carry different losses. A point forecast hides that choice.
P10 46 · P50 127 · P90 208, index at period 10 (illustrative)
03 Challenge · Reference
Keep a naive, seasonal or moving-average reference. Compare out-of-sample errors at identical origins and horizons. Added complexity must earn a decision-relevant gain.
P10 43 · P50 127 · P90 212, index at period 10 (illustrative)
04 Reconcile · Coherence
Product, customer, region and total forecasts should not become unrelated negotiations. Align them into one total and record manual adjustments rather than burying them.
P10 66 · P50 127 · P90 189, index at period 10 (illustrative)
05 Commit · Quantile
Translate the chosen band into stock, capacity and cash cases. A named owner accepts the commitment and its review trigger.
P10 94 · P50 127 · P90 161, index at period 10 (illustrative)
Uncertainty needs an owner
A forecast proposes a distribution. An authorised person decides which uncertainty to finance.
The fan should widen when evidence weakens. That is information for the decision, not a modelling defect to be designed away.
Wide fanReduce the commitment and stage the release. Do not force agreement on a central estimate.
Narrower fanCommit to a middle band and name the review trigger that would change it.
Narrow fanA tighter band is earned only after rolling-origin tests against a simple reference.
Fan width at the gate: 196 index pointsIllustrative fan, not data
A narrow band can be spurious when history is short, stockouts censor demand or a supplier changes. Review these conditions before trusting precision. Where evidence is weak, reduce the commitment rather than forcing agreement on a central estimate.
Overrides should state evidence, affected quantile, consequence, owner and expiry. The planning meeting can accept, reduce or reject a proposed release. The model does not sign that commitment, and seniority does not turn an override into evidence.
Planning is a commitment system
Forecasting matters when it changes replenishment, production or capacity. The operating routine keeps that translation explicit and reviewable.
idle capacity and obsolete cloth
financed buffer
stockouts and expediting
Owner · review trigger · expiry
Override needs evidence, a reason and an expiry
Arrival window: late arrivals stretch the upper tail
Inventory · Service policy leads
Use demand and lead-time distributions to test safety stock, reorder points and review periods. Show how order minima, pack sizes and cash limits change the result. Inventory is a financed buffer, not proof of forecast quality.
Lead time · Lead-time spread leads
Separate supplier processing, production, transport, customs and receiving where behaviour differs. Preserve late and early arrivals. A single average can make a reorder point unreliable.
S&OP · Cash exposure leads
Bring the reconciled fan, capacity, procurement, cash and commercial assumptions into one review. Minutes record accepted risk, actions and owners, not agreement on a number.
Authority · Commitment owner leads
Overrides require evidence, a reason and an expiry. Release authority stays with the designated role. Later orders, stockouts and errors test model and override.
Cameroon textile / conditional application
The method would connect market evidence to product commitments without treating regional growth as a factory order book.
None recorded
Trial orders should test a defined assortment, not an undifferentiated national total.
For GFT's Cameroon textile programme, the question is which product and customer combinations merit testing. Market context can frame interviews, but a range still needs evidence on sizes, quality, delivery, margins and substitution. Cold-start forecasts remain provisional until local orders and fulfilment data challenge them.
Point at a tier to read it
Context only. Sourced market estimates frame interviews. They are not an order book.
Evidence requiredThe source, its date and the population it describes.
Context only. No order is recorded here.
Capacity and OEE analysis can test whether a service commitment is deliverable.
Supply-chain design must be tested beside demand. Imported inputs, local processing, changeovers and distribution alter lead time and feasible service. Line analysis should use the same scenarios, with utilisation and availability as constraints, not assumed outcomes.
The promise sits inside productive time on this schematic line.
Time that makes good product at the planned pace. This is what a service promise can draw on.
Schematic proportions, not a measured line. Test line losses in the OEE tool
Commercial, operations and finance should record who can release stock, capacity and expenditure.
Digital design can specify records and decision rights; it cannot establish demand.
No release has a named owner yet. The plan cannot release stock, capacity or expenditure.
A structure to record, not a record of GFT roles. Name the owner for each release.
Continue with the decision
Examine the method, then test the decision logic. Point at a thread to see how it changes the fan. Each route keeps assumptions visible and commitment authority with people.
Wide fan, gate at P80
Wide fan. Analogue evidence is thin, so the gate sits at a cautious band.
PaperForecasting from a cold start
Examine analogue selection, evidence transfer and forecast evaluation without direct sales history.
Follow this threadNothing is recorded or sent.
One commitment, owned by a person.
Illustrative fan, not data