Demand forecasting in apparel

Basics can be forecast from their own history. Fashion cannot, so planners forecast at a higher level, category or class, and then allocate down to styles using attributes: silhouette, color family, price point, and the performance of comparable items in prior seasons. The error at style level stays large even when the category forecast is good.

Size forecasting sits underneath and carries its own error, which is where a category-level miss becomes a size-level problem, as described in the size curve.

Time series methods work for replenishable items with stable patterns. Attribute-based and analog methods work for new items. Machine learning models mostly improve the attribute matching rather than removing the fundamental problem, which is that the future demand for a new item is not contained in the past.

The realistic goal is not accuracy. It is a forecast whose error is small enough, and understood well enough, that the commitment can be sized around it.

Because the commitment happens before the season, forecast error is not an analytical outcome. It is physical goods in a warehouse, and it stays there until someone buys it or it is written down. That translation is the whole reason the surplus chain exists, and it is described in where unsold apparel goes.