By Dr. Luis Wilbert, PhD Graduate at ZLC and Dr. Yasel Costa, ZLC Professor.
Novelty sells, and any retailer whose products are in some sense subject to ‘fashion’ must continually retire lines that are well past their sales peak, and replace them with new offers – even though these may be functionally almost indistinguishable from those they replace.
In some ‘fast fashion’ businesses – Zara being the textbook case – this involves replacing almost every line, every season and often within the season. More commonly, in what might be termed ‘slow fashion’ markets (think of Puma, Nike, Adidas) the season’s ‘collection’ will include items that will only be stocked for a single season, but also a core of NOS (Never Out of Stock) items which will continue to be offered across many seasons. Nonetheless, this NOS assortment also needs to be refreshed. The retailer needs to decide which lines to replace and which to retain through the next season(s). That may select a few hundred new/replacement lines out of several thousand possibilities. To complicate things further, because of production and shipping lead times, these choices have to be made around even months before the start of the new season, when demand forecasts are inherently uncertain.
The aim, obviously, is to maximise profit. Sales of a fashion line (assuming it is successful) typically ramp up from launch to a peak, subsiding to a rather lower (but still profitable) plateau, and ultimately tailing off. Ideally the new line is introduced before sales of the existing item tails off too severely, to maintain a high overall level of sales in that category.
But the assortment will often include several very similar SKUs, and the proposed new line may be very similar in appearance to the one it replaces. The retailer wants to limit product overlap, but needs to consider product substitution and estimate the likely degree of Sales Transferability (ST). This works both ways. If a SKU is withdrawn, what proportion of its demand will transfer to other, fairly similar lines. Or if a new SKU is introduced, how much of the demand for it is genuinely new, and how much has been cannibalised from sales of other similar lines? ST also opens the possibility that the retailer can simplify its offer without significant loss of sales by selectively retiring some of a group of very similar SKUs and showcasing another. This is sometimes termed ‘risk pooling’.
In my thesis towards a PhD, I have been modelling a framework that includes Sales Transferability in this decision-making process. The work has been greatly aided by Adidas who have provide comprehensive data on ‘real life’ assortment changes.
The proposed model consists of three sequential stages. The first is a Sales Transferability calculation. Similar products are grouped by attributes. There is a hierarchy of attributes – are the items shorts or shirts or shoes? Do they group by age group, gender, function and so on? But for our purposes we are interested in the finer detail differences, probably visual, within a group of similar products which would cause a consumer to consider one line an acceptable substitute for another – in this case we grouped by the items’ primary colours, although the same approach could be taken with other features. To capture this, we apply machine learning to 32 features in 18,000 images of the items (which was largely successful although some difficulty was experienced in distinguishing between shoes intended for different genders!), and can then create a Structural Similarity Index Measure, which should provide a measure of Sales Transferability within each ‘cluster’ of similar products.
In the second phase I have created a demand model which reflects both seasonality and Product Life Cycle. Obviously, there is no demand data for the proposed new SKU, but we do have data for the line that is being superseded, and provided we have a good estimate of ST, we can convert that historic data into a reasonably robust forecast for the new line. By merging the historics of one SKU and the forecast for its replacement, we end up with an expected demand that is almost stable, even though it is derived from two different articles. All this feeds into a robust optimisation model which determines an optimal assortment strategy, maximising revenue while reducing the risks arising from demand prediction error.
Courtesy of Adidas we were able to compare the demand forecasts from this model with those produced by the more traditional approaches common in the industry, which it matched or outperformed – for example, showing an 18% improvement over scenarios that depended solely on historic data. That equates to a maximum total revenue of 57.7 million units, which is 78% more than the ‘baseline’ prediction, and only 13% below the season’s best performing assortment.
So, the approach has been successful and there is no reason in principal why it should not be adapted to other businesses, such as supermarkets, and clustering using different, perhaps non-visual, features on different types of article (they do of course have to be features that the customer may take note of when making their comparison). Many retailers don’t want to be offering over-similar products, and this work helps with the challenge of, firstly, identifying what ‘similarity’ is, and then deciding which items in a similarity cluster can be dropped without losing the market, or without that loss outweighing the savings made by not carrying that SKU, savings which can be applied elsewhere.
For more information, contact Dr. Yasel Costa at [email protected]
