No history is a forecasting condition, not a forecasting ban

Missing historical data is not an anomaly in clothing manufacture; it is a defining structural condition of the sector. Academic reviews of apparel supply chains document that fashion demand combines taste volatility, seasonal cycles and acute lack of historical observations for new styles.[S01] Because clothing items have short commercial lifecycles, supply-chain scheduling and retail allocation decisions must be taken before a garment accumulates a multi-year sales series.[S0101] Even established apparel brands continuously launch silhouettes, fabrics and colourways that did not exist in prior seasons.[S01][S02] What distinguishes a greenfield manufacturer is not novelty, but the depth of the transfer gap it must cross when constructing its initial expectation.

Forecasting theory addresses the zero-observation condition directly rather than treating it as a mathematical void. Comprehensive reviews of the discipline note that when no historical observations exist for the item of interest, structured judgemental methods, reference-class analogies and scenario design become the primary instruments of prediction.[S08][S0102] When a planner estimates the completion time of a new facility or the initial uptake of an unlaunched product, the absence of an internal time series shifts the engineering task from parameter fitting to evidence curation.[S08] The forecaster must still specify the target variable, bound the plausible range, document why chosen comparators are relevant and separate objective prediction from commercial target-setting.[S0102]

Treating cold start as an evidence-design problem changes the first questions a technical team asks. Instead of asking which neural architecture handles short series best, the team asks where transferable signal resides and what distortions contaminate it. Prior information for a new garment venture arrives in four forms: historical launch trajectories of analogous products from external or partner catalogues, macro-level trade and demographic bounds, structured elicitation from domain buyers, and choice experiments conducted with prospective customers.[S02][S29][S31] None of these sources equals an owned, stockout-corrected transaction log, yet each carries partial information that narrows an otherwise arbitrary planning range.

Exhibit 1 illustrates the progression that organises this paper. Pre-launch, the venture relies on analogue distributions and expert priors to form an outside view.[S08][S28] Before committing full factory capacity, it refines those priors through calibrated choice research and small, availability-aware pilot releases that generate the first owned purchase observations.[S31][S32] Only after clean transaction and stockout records accumulate does the venture shift weight toward statistical smoothing, attribute regressions and cross-series models.[S02][S06][S08] Crucially, Exhibit 1 is an operational maturity sequence rather than a guarantee that uncertainty shrinks monotonically; if a pilot reveals that an analogue pool was poorly matched to local buyers, the right response is to widen the interval and revise the assortment before scaling.

Separating forecast construction from target setting is essential during this transition. Guidance on judgemental forecasting warns that when staff responsible for commercial quotas also author the demand forecast without independent rules, optimism bias and political anchoring corrupt the numbers.[S0102] A business plan states the volume required to amortise machinery, whereas a demand forecast estimates what customers will buy under stated price, channel and availability conditions.[S08][S0102] Maintaining a hard wall between break-even targets and empirical demand priors is the first governance rule of a greenfield launch.

Within GFT's three-part Cameroon textile programme, prepared on 31 January 2024 with source refreshes flagged through 2026, demand forecasting is architected as a Phase 2 digital capability rather than a day-one software purchase.[R06][R07] Across the programme's feasibility study, performance-management system design and digital strategy, Phase 1 focuses on foundational instrumentation: reliable enterprise resource planning records, standardised stock-keeping unit (SKU) identifiers, shop-floor quality logs and clean commercial transaction capture. A machine-learning model fitted before the factory can distinguish an unfulfilled order from a cancelled invoice, or a stockout from a zero-demand week, only automates the propagation of measurement noise.[S06][S37]

Waiting for Phase 2 to deploy complex algorithms does not mean that pre-launch and Phase 1 decisions occur without quantitative support. Before a plant trains its first cross-series model, management must choose initial garment categories, book yarn and fabric capacities, negotiate institutional uniform tenders and size opening cut-and-sew batches. This white paper specifies the pre-Phase-2 and transition machinery that bridges that gap: writing the demand contract during project preparation, assembling an auditable prior from analogue and primary research, running small launch pilots that earn clean history, and governing graduation from simple baselines to global probabilistic models in Phase 2.[S02][S21][S28][S32]

The macroeconomic context in Cameroon explains why disciplined demand engineering carries high stakes for local textile industrialisation. According to the inherited sector register, Cameroon produced 330,000 tonnes of seed cotton in 2022, 314,455 tonnes in 2023 and about 340,000 tonnes in 2024, with 2025 projections exceeding 350,000 tonnes.[R01][R02] Yet roughly 95 percent of Cameroonian cotton fibre is exported raw while only about 5 percent is processed locally, alongside a national SND30 target to reach 50 percent local processing by 2030.[R06][R07] Large industrial initiatives reflect this policy push, including the announced SODECOTON and CICAM merger targeted by 2030 and the Camtext SA integrated project announced on 12 November 2025 at US$300 million (180 billion XAF) to process 12,000 tonnes a year toward optimal capacity around 2033.[R08][R09][R10]

A greenfield apparel maker must treat national processing targets and incumbent metrics with strict analytical care. The gap between 5 percent local fibre processing and a 50 percent national target demonstrates raw-material supply potential, not guaranteed consumer or institutional absorption for any specific garment line.[R06][R07] Historical distress at existing domestic facilities underscores the cost of misjudging market fit and utilisation: public reports indicate that CICAM's domestic market share fell from about 85 percent three decades ago to about 5 percent, with July 2023 capacity utilisation reported at 33.7 percent at Garoua, 48.9 percent at CICAM 1 and 10.83 percent at CICAM 2, and underutilisation losses of about 2 billion XAF in 2020.[R11][R12] Meanwhile, Statista estimates Cameroon clothing and footwear expenditure at about US$94.70 per person (around 8.0 pieces) in 2024 and apparel e-commerce revenue rising from US$20.32 million in 2023 to US$33.61 million by 2027, while McKinsey reports that AI at scale in established fashion firms has been associated with 20 to 50 percent forecast-error reductions and 35 to 40 percent inventory-cost reductions.[R26][R27] None of those inherited benchmarks may be copied into a greenfield plan as an expected GFT factory result.

A demand forecast has no operational meaning until its grain, horizon and target event are bound to a specific physical or financial commitment. In apparel manufacturing, candidate dimensions include style, fabric construction, colourway, size, sales channel, geography and week.[S02][S03][S04] Publishing a forecast at the finest combination of all dimensions months before launch creates an illusion of precision while guaranteeing extreme noise, whereas publishing only an annual factory tonnage figure leaves planners unable to order dye lots or cut size curves.[S08][S11] The published forecast grain must match the finest level at which a named decision-maker can alter an operational commitment.[S08][S0102]

Different clocks govern an apparel venture's commitments, and each clock demands a different forecast structure. Long-range capacity and raw-material booking (reserving greige cotton fabric weights or yarn counts 16 to 26 weeks ahead of delivery) depends on aggregate fabric-family demand across styles.[S01][S0101] Medium-range assortment and initial cut-and-sew lot sizing (6 to 12 weeks before launch) requires style-by-colour volume ranges and a prior size-distribution profile so the cutting room can lay markers efficiently.[S02][S03] Short-cycle channel allocation and replenishment (1 to 4 weeks ahead) operates at the SKU-by-channel-by-week grain, where observed sell-through and stockout flags update the initial prior.[S03][S06]

Exhibit 3 summarises how commitment clocks map to forecast grains, illustrative planning horizons in weeks, and the portion of the predictive distribution that drives the decision.[S03][S06][S08][S16] When booking undyed or standard-dyed fabric 16 to 26 weeks ahead, pooling across styles that share the same base cloth reduces variance, because fabric can still be allocated across silhouettes later in the calendar. When releasing the initial cut-and-sew lot 6 to 12 weeks before launch, the planner commits to specific styles, colours and size ratios, but caps the initial batch using the downside and central quantiles so untested styles do not lock up working capital in finished goods.[S16][S32] Once the launch is live, 1-to-4-week replenishment shifts focus to SKU availability and upper-tail quantiles to prevent stockouts on fast sellers.[S06][S16] While Exhibit 3 provides an operational template, each factory must substitute its actual spinning, dyeing, cutting and shipping lead times into the contract.

Aligning the forecast unit with the decision also requires strict separation between recorded sales and latent demand. Retail forecasting research establishes that estimating demand solely from sales transactions introduces a systematic negative bias whenever products experience stockouts or partial availability.[S06] Stockout-annotated retail datasets confirm that the sales-equals-demand assumption collapses when inventory runs dry, because observed sales are right-censored at the available stock ceiling.[S07] In a cold-start apparel launch, where initial pilot lots are kept small to contain financial exposure, stockouts on popular sizes or colours are frequent.[S06][S32] Feeding raw, uncorrected pilot sales into a replenishment model causes the model to mistake a stockout week for a demand plateau and starve the winning SKU.[S06][S07]

Channel structure further complicates the target definition. An order placed by a wholesale distributor or an institutional buyer is a lumpy commitment at the factory gate, whereas consumer sell-through in direct retail or e-commerce reflects end-buyer uptake day by day.[S04][S06] Treating wholesale sell-in and retail sell-through as interchangeable rows in one time series corrupts both level and timing. Chapter 2 therefore formalises a binding forecast contract that separates order, dispatch, sell-through, return and stockout events before any model runs.[S06][S07][S36]