Chang and Fyffe (1971) assume that a ?rm has a ''long-run sales history of individual seasonal-style-goods SKUs or groups of such SKUs''. They propose to estimate demand by using regression on those historic sales, also based on the ''outcome of some observable variable''. However, they do not explain in detail how that can be done nor do they test the method using real data. It seems dif?cult to apply this method in the apparel industry, as long-run sales histories of very similar products are rare.

Chambers and Eglese (1988) discuss the use of preview demand data that are gathered by sending out a preview catalogue (whichdoes not necessarily include a full product range) to a sample comprised of several thousand regular customers and offering themthe opportunity to order products at a discount before the season starts. They assume that an aggregate forecast for the full product range is given, and propose to forecast the demand for a product line by multiplying the aggregate forecast with the fraction of total preview demand for products in that product line. They further propose a second, slightly more sophisticated forecasting method,which takes into account that the ratio of total demand to preview demand ('the scaling factor') may not be the same for all product lines. These methods are very suitable and, indeed, have been developed for an apparel mail order company.

Thomassey and Happiette (2007) propose a decision-support system based on neural networks, which automatically performs item sales forecasting. The system is designed to deal with many characteristics of the apparel market: large number of items, short lifetimes, substitution of most items with each new collection, long lead times, and in?uence of many external factors like the weather,promotions, fashion, and the economic environment. The proposed system is composed of three steps: obtain prototypes of demand behavior using a clustering procedure on historical demand data, (2) link these prototypes to descriptive criteria (e.g. price, lifespan or materials) using a probabilistic neural network, and (3) assign each new item to a prototype based on the item's descriptive criteria. Forecasts generated by the proposed model on a set of 285 new items from a French apparel distributor have a MAPE of 147%. So, accuracy is low despite the complexity of the method. For this reason, we decided not to include this method in our comparative study. The results in Section 5 will show that the simpler methods that we do consider are all more accurate (for our data set).