There was a time when it was enough to have individual store managers make the decisions of which products to carry and how much to order. As hardlines retailers have grown, their offerings have become more elaborate and competition has become more sophisticated. Customers and their behavior have become more complex as well. It’s no longer possible for people to keep track of all the variations and trends making each location unique.
For the last two or more decades, conventional wisdom has been that clustering stores is the way to address this problem. While, conceptually, grouping stores with similar characteristics makes sense – the reality is that it’s very difficult if not impossible to do this in a consistent and effective manner. From volume clusters to attribute clusters and even more recently devised mechanisms for deriving “intelligent” clusters using BI tools, all suffer from flaws that keep them from achieving the elusive goal of localization. Let’s look at some of these to understand why that is.
Inherent problems with volume based clusters:
The most widely used of these decades-old methods to cluster stores is volume clustering. Typically using historical sales for a group of merchandise presumed to be similar then finding apparently logical breaks as that volume is ranked high to low. The breaks represent boundaries which define which cluster stores fall into. Most commonly the result is a set of 10 or fewer cluster groups (although we’ve witnessed organizations that do many more). The three most common flaws in this process are:
- The clusters are virtually always based on historical sales
- The clusters are not updated frequently enough
- The clusters don’t incorporate any attributes of the merchandise or the stores
Typically a retailer looks at a store’s performance over the last year or compares like for like seasons to group stores that performed similarly together. The problem is that usually its only historical sales that are used to group the stores together. This creates a bit of a self-fulfilling prophecy. If a store underperformed in a department because of an inordinate amount of out-of-stocks, could the store have performed better?
The better option is to use either historical demand, or better yet, a demand forecast. When we refer to demand we look for what would have happened had the store been in stock. This allows us to understand any missed opportunity that may have resulted from out of stock situations. Of course using a true, forward looking demand forecast which includes understanding those missed historical opportunities to cluster would be the best of both worlds. Not only are you incorporating lost sales, but you would be incorporating future trends of store behavior as well.
Another problem is timing of when clusters are used and how often they are evaluated. Store assignment to clusters should be evaluated as often as possible. Customers change their shopping behavior constantly and this changes the behavior of individual stores constantly. As a result, store clusters need to be re-evaluated as frequently as possible using the most current understanding of behavior. At a minimum cluster should be updated prior to each merchandising activity that is consuming them. For example, updates should occur pre-season when an assortment plan is created (which products in which locations), again when a buy plan is made (soft commitments for how much), again when the actual order is placed (time has passed, things have changed), again when commitments to DC shipping locations (DC splits) are confirmed, again if the buy is pre-distributed to commit stock to stores, another when the goods hit the DC, and so on.
Furthermore, clusters should never be used to drive replenishment decisions. Once the goods have hit the stores, any re-allocation/replenishment activities should be based on actual, individual item/store behavior. There is simply no good reason to use clusters after the first allocation – and even then a sound argument can be made as to whether they should be used at all. This is especially true for hardlines and commodity products that don’t have the volatility or size complexity seen in hardlines. Today’s technology is capable of identifying and leveraging the uniqueness of individual products in individual stores, so each store can be treated as it needs to be. This is where true localization can truly begin to provide huge benefits.
Clustering beyond volume
Retailers use more than just volume clusters but as of yet no standards have been adopted. Climate, an attribute of physical location, is one of the most commonly seen attribute clustering criteria. While useful, categorizing a location as one of three or four climate groupings is an inexact science at best. And there are surely many other attributes of location and product that can refine the quality of the result. In traditional execution, however, more attributes leads to more groups needing to be managed. These groups are often nested into artificial hierarchies and quickly become difficult to navigate and virtually useless when complexity reaches the point that execution is impossible.
More recently many have tried to utilize Business Intelligence solutions and statistical analysis to find more refined groupings of stores that behave similarly. One of the most common flaws of this approach is that the resulting clusters have no definition of what makes them similar. While it may be accurate that they have similar behavior, that conclusion alone is useless unless a merchant understands what makes a group unique. Without that there is nothing available to guide a cluster specific decision.
Where does that leave us?
Ultimately, newer technologies that have been focused on solving these problems in retail have refined the utilization of clusters significantly. Better solutions constantly review and update clusters based on current behavior and the processes that consume clusters are able to accept and modify their conclusions accordingly and without unnecessary user intervention. They analyze and update activity across a variety of attributes proven to impact the products and locations within the grouping and offer flexibility to navigate across those without being tied to unmanageable hierarchical relationships. They also derive learning about individual products to the point that the need for clusters is either significantly reduced or eliminated in many processes throughout the management of inventory (such as replenishment). If your processes are not supporting that level of locating management and practical clustering, it’s simply not possible to achieve what is expected when discussing localization in retail.
Consider the following as you think about the quality of your clustering practices:
- Do you cluster only on volume or on other attributes or KPIs too?
- Do you cluster only on historical sales or do you incorporate missed opportunities / lost sales?
- Do you cluster based on a forecast of demand for the periods you’re using the clusters to represent?
- Do you re-evaluate your cluster assignments as often as possible?
- Do you re-cluster in-season?
- Do you cluster the stores as low on the merchandise hierarchy as is reasonable for your business?
- If you change cluster definitions can the systems that use them accept that change cleanly?
- Do you drive re-allocations or replenishment by clusters?
If the answer to any of the above questions is “no”, there is room for some self-evaluation and improvement to your localization strategy.
Look out for next week’s blog on SKU Rationalization.
Sign up to receive updates throughout this series.
