The forecast target must be demand, not a convenient sales extract

Defining the target event solves what is measured; defining the forecast grain solves the structural resolution at which that event is indexed and aggregated.[S0202] In apparel manufacturing, a product is a nested hierarchy of technical and commercial decisions: product family (woven tops), style or silhouette (short-sleeve workshirt), fabric construction (100 percent cotton twill at 180 grams per square metre), colourway or dye lot (indigo or khaki), size (S, M, L, XL), and sales channel (direct e-commerce, Douala retail counter, regional wholesale distributor or institutional contract).[S02][S03][S04]

Machine-learning research on unseen fashion items shows why this attribute and grain hierarchy must be locked in the contract before data capture begins.[S02][S03] Attribute-based demand models infer the sales trajectory of a new item by learning how clothing attributes (category, fabric, colour, cut) and merchandising factors (selling price, product age, store context, promotion) drove demand across prior items.[S02][S03] If an enterprise table concatenates style, colour and size into an unparsed text string, or reuses SKU codes across seasons when fabric weight or price changes, the model loses the structural bridges needed to share statistical strength across variants.[S02][S17][S22]

To preserve cross-product learning and hierarchical coherence, the forecast contract enforces four schema rules across the product and channel hierarchy:[S02][S03][S11][S17]

A forecast is actionable only if it arrives before the physical lead time of the decision it guides.[S08][S0202] Forecasting theory and practice emphasise that method selection, temporal aggregation and evaluation must be tied directly to the planning horizon and decision context.[S08][S0202] Short-term forecasts support personnel scheduling, machine loading and transport dispatch; medium-term forecasts determine raw-material purchasing and batch sizing; long-term forecasts guide capacity and market strategy.[S0202] Evaluating a sixteen-week fabric-procurement forecast using one-week-ahead point accuracy is a methodological error: a model that achieves low one-week error by echoing last week's sales provides zero utility when yarn must be ordered four months in advance.[S08][S12][S33]

In an apparel manufacturing venture, four operational clocks run in parallel, and the forecast contract assigns each clock its own target grain, forecast horizon, update cadence, information cut-off and decision owner.[S01][S08][S0202]

Exhibit 7 maps five apparel commitments across two conceptual dimensions from 0 to 100: product-and-channel grain specificity on the horizontal axis and decision lead-time horizon on the vertical axis.[S02][S03][S06][S08][S0202] In the upper-left quadrant sits fabric and yarn booking (illustrative coordinates 22 and 85), reserving cotton yarn counts and greige fabric weights 16 to 26 weeks ahead at a fabric-family grain.[S01][S0202] In the upper-right quadrant sits the seasonal assortment plan (coordinates 64 and 74), committing style counts and colourways 12 to 20 weeks before launch.[S02][S08] Near the centre sits institutional tender scheduling (coordinates 38 and 62), reserving sewing capacity 8 to 16 weeks ahead.[S08][S0202] In the lower-right quadrant sit initial cut-and-sew lot sizing (coordinates 78 and 42, 4 to 8 weeks ahead at style-by-colour-by-size grain) and weekly size replenishment (coordinates 92 and 16, 1 to 3 weeks ahead at SKU-by-channel grain).[S03][S06]

What Exhibit 7 establishes is that one horizon cannot serve all five clocks; a forecast service must publish multi-horizon distributions reconciled across the hierarchy.[S16][S17][S24] What the matrix cannot establish is exact factory lead times without a supply-chain audit, as lead times depend on whether the maker buys finished fabric, contracts local dyeing, or integrates spinning and finishing.[R09][R10][S01]

Every horizon in the contract requires an explicit forecast origin, an information cut-off timestamp, and a production freeze window.[S13][S37] Machine-learning reproducibility research defines data leakage as a spurious relationship between independent variables and the target arising from collection, sampling or preprocessing strategies that expose information unavailable at prediction time.[S37] In multi-horizon apparel forecasting, leakage occurs whenever an analyst evaluates a 6-week-ahead cutting decision using inventory, price, weather or return records from the intermediate five weeks.[S13][S37]

Physical factories also impose freeze windows: if cutting and sewing require two weeks from fabric issue to garment packaging, a replenishment forecast issued today cannot alter finished goods arriving in weeks one or two.[S06][S08] Its actionable target is cumulative demand after the two-week freeze window across the review cycle, evaluated across the nine-quantile structure of the M5 Uncertainty competition (or a contracted subset such as the 0.10, 0.25, 0.50, 0.75 and 0.90 quantiles).[S14][S16]