Clustering is a method companies use to reap the benefits of a centralized approach to retail planning while responding to variations in consumer behavior. Clusters of stores or categories within stores are identified to enable retail units with similar behaviors to be managed together.
In an era of fast increasing consumer expectations (especially in fulfilment and convenience) and instant price comparison, a business’s profitability and survival depend upon the effective use of tools such as clustering.
In this whitepaper we will explore a variety of approaches to clustering; the benefits of using behavioral clustering within the category management process; and how to tailor assortment and space in accordance with customer demand.
Where Does Behavioral Clustering Sit Within the Category Management Process?
Behavioral clustering encourages the different disciplines within category management to work more closely together and abandon the silo approach that holds back some businesses. The data and analysis generated will inform strategic decision-making by Merchandising Space, Range Assortment, Pricing, Marketing, Promotions and Seasonal teams.
Given this, behavioral clustering yields the greatest value when driven early in the business process in the common progression of: financial analysis → customer insight → ROM (rest-of-market) analysis → store development initiatives → marketing → assortment → merchandising → store implementation. Typically, it sits between financial analysis and customer insights.
What is Behavioral Clustering?
Let’s start by looking at the two ends of the spectrum:
1. One Size Fits All
At one end you have a centralized, single-assortment approach in which all stores, capacity allowing, generally receive the same items. This does allow for great economies of scale and for the central buying operation to trade bulk order quantities for discounts. However, because it doesn’t allow for store-to-store variations in consumer behavior, some of those items will be placed in stores where they have little or no market, ultimately stealing shelf space from lines that do.
At the other end of the spectrum, each store is treated as its own individual cluster, with buying responsibilities relegated entirely to each store. This, however, is labour intensive and makes it harder to negotiate price, build a brand identity that lets customers anticipate what can be found in-store, and run an efficient supply chain.
A clustered approach sits somewhere between these two extremes. Clustering itself can be executed in a variety of ways, ranging from the more common top-down methods to demand-driven and behavioral approaches. Let’s explore clustering methods in further detail:
2. Top-down Store Clustering
Traditionally, retailers used “top-down” criteria to define clusters — geography, for example. Geography is easy to define, fits neatly around DCs and may, in a secondary effect, reflect factors like demographics or regional variations.
Another common cluster method is store size. There is a logic to producing an assortment tailored, for example, for stores under 3000 sq. ft (300 m2), another for those over 10,000 sq. ft and a third for those in between.
Other “top-down” attributes include sales volumes, type of store location (city center, suburban, small town, industrial area, rural etc.) and demographics (older shoppers, students, families, income, ethnicity, etc.)
This approach makes assumptions about customer behavior based on top-down characteristics rather than rendering analysis that uses actual demand-level data. Those who employ this approach risk missing out on key customer trends and preferences, ultimately driving a non-optimal customer offering to their stores.
Example – One of our clients historically grouped their stores via a mix of geographic and location, and affluence information. They discovered that one of their stores had been identified as an “economy” store due to location and level of affluence, but that it sat on a busy commuter road often used as an alternate route to the highway. In reality, its customers held more premium preferences than they had assumed — and shopped regularly for dinner on their way home from work.
3. Bottom-up Store Clustering
A bottom-up approach considers store demand patterns as the starting point for cluster analysis. In this approach, clusters are driven primarily through category-level sales at an individual store level. Typically, sales value, volume and profit is used; where available, a best practice approach is to utilize long-term demand forecast information to support the analysis.
The analysis of sales contribution and variation in customer demand — either across all stores in the estate or by chain / format — ultimately drives a more customer-centric approach. Stores which demonstrate “like” behaviors across total category contribution are grouped together to reflect a total store-level view of similarities in customer demand.
Location attributes, demographics and geography are still key components of the analysis, but in this approach they form overlays to the analysis that unlock a more granular view of trends and customer profiles. For example, stores in Cluster A may demonstrate similar demand patterns of over-indexing demand in “confectionary,” “chips & snacks” and “soft drinks,” but the added data overlays highlight that they are generally located near movie theaters, universities and high schools.
By providing insight into customer trends and preferences at a total-store level, this approach to clustering can provide a laser focus to target:
- Store Formats / Macro Space – Stores with similar customer demand patterns across categories are clustered together to drive insights for macro space allocation across store formats & sizes, allowing space to be used flexibly in-line with store clusters and varying demand.
- Category Variation Analysis – Understanding demand patterns across all categories within the store estate allows quick identification and ranking of categories which show the largest variation in demand across the estate. Those categories would therefore benefit most from the implementation of a more localized / store-level range & space customer offer to maximize ROI.
Note: Although this approach reflects customer demand at its core, it requires a further level of behavioral clustering to unlock customer demand at a category level for high variation categories — in other words, recognizing that not all categories are shopped in the same way within a single store. For example, customers shopping for “Wine” may be more price conscious and opt for the latest discounts whereas “Baked Beans” shoppers may be more brand conscious, opting always for their favorite brand regardless of price.
4. Bottom-up Category Clustering
Bottom-up category clustering provides the most granular and behavioral approach to clustering; it groups stores according to similar customer demand across the total estate or chain for the individual category. This approach uses sales or, ideally, demand forecast data at an SKU/Store level along with available attribute information (e.g. product brand, flavor, price, packaging type etc.) to reveal the factors contributing to customer decision-making processes and preferences.
Just as with bottom-up store clustering described previously, this approach requires the application of data overlays such as store attributes, demographics and geographical location to provide deeper insight into cluster trends and shopper profiles.
Once the clusters have been identified, shopper profiles are created to represent shopper preferences within the category i.e. over or under indexing brands, flavors, price and packaging types along with typical shopper profiles like high or low affluence, income levels, age groups etc. This allows category managers to better target range and merchandising according to shopper demand.
What is the Optimal Number of Clusters?
The question then becomes “how many clusters?” In an earlier age, top-down clustering was the only practical approach given the huge labor demands of the alternatives. These days, however, advances in automation and optimization technology mean that much of the workload can be passed to a good assortment and space optimization system. Computers are peerless when it comes to spotting the patterns that create effective clusters. However, management time is still a bottleneck, and cluster numbers must reflect the supply chain team’s ability to manage them.
Optimal clustering finds the sweet spot between “one-size-fits-all” and “micro-managing-everything.” The best trade-off is between efficiency and effort for every store and category type. It’s worth noting that managing every sub-category individually can result in missing the insights and advantages that come from spotting trends in similar categories.
At RELEX we’ve found it exceptionally useful to begin with a blue-sky approach, letting the system suggest clusters without predetermined criteria. Our customers are almost always taken aback by the results.
One supermarket chain we work with typically divides its stores into three categories that reflect shopper missions: local convenience outlets, medium-sized town or city-center outlets and large out of town stores. In a number of instances, the convenience stores behaved like city-center stores and vice versa.
Another customer divides its stores according to income bracket based on location demographics. In more than one case, stores in lower income areas behaved quite differently than expected because they were on routes to higher income areas and were shopped by commuters heading home.
Takeaway: Let the data speak for itself.
Though clustering can seem like a vast amount of extra work, take heart. Many categories are relatively uniform. Variations across stores are relatively low and, once identified, remain quite static. Breakfast cereals might exhibit a high degree of variation, but trash bags probably don’t. You need bags that don’t split, are the right size and are stocked in the right quantities — that’s pretty much it. Or, to consider an example from the hardware space, customer taste in light fittings doubtless varies with income and age, but taste in hinges and screws probably does not.
Retailers might routinely operate 100 categories, but of those, only 20 or even just 10 might exhibit a degree of variation across stores that warrants clustering. In general, the Pareto principle (better known as the 80/20 rule) applies; 20% of a store’s assortment produces 80% of the profits. By focusing efforts on 20% of a store’s products, 80% of the possible overall benefits can be achieved.
Identifying which categories need close management leads to a better understanding of what to organize into clusters. It’s simple prioritization: leave trash bags to the system and focus on cookies.
Clustering Theory – Best Practices
RELEX’s behavioral clustering solution takes a best practice approach: it supports bottom-up store and category clustering through granular category and product attribute variation analysis. The RELEX solution uses 3 key factors — cohesion, separation and population — to create, compare and rank a wide number of cluster schemes, ranking clusters by accuracy to determine the optimal mix:
Clustering Key Principles
Ensuring that the stores within a cluster are as close together as possible in terms of behavior.
Ensuring that the clusters are as far apart from one-another as possible in terms of behavior.
Ensuring that there are as many stores within a single cluster as possible.
The graph above shows a sample cluster analysis for a “Yogurts” category. Each of the blue circles represents a store, and each store is mapped according to its sales contribution levels using a value/volume/profit mix for the selected attributes. The graph reveals the “Sub-Category” (Health, Organic, Children’s, Greek etc.) and “Price-Level” (product price-bands) product attributes to be significant drivers in the customer decision making process, thus the wide variation across the category.
Four clusters of stores show similar shopper patterns for the category, so they would benefit from a range and space strategy tailored to reflect this. Take note, though, of the individual store that’s been separated into its own cluster. This store serves as an excellent example, as it is the only store in the estate located in a train station. It follows that this store’s shopper mission differed greatly from the others’, and as such, its range and space offering needed tweaking to reflect this.
A question we hear often in behavioral clustering is “how should we determine the attributes that drive clusters?” The Customer Decision Tree (CDT) for the category is a good starting point. However, RELEX’s suggestion is to utilize all available attribute data and allow the solution to highlight key attributes that reflect a high level of variation across the estate. This approach allows category managers to see whether there are attributes not currently included in the CDT for the category that should be considered in the range & space context moving forward.
What Insight Will Clustering Deliver?
Once the solution has determined the most accurate cluster scheme through demand data, the next step is to further analyze the category and the performance of its product groups. This is also when overlaying information attained through census data, other market data channels or customer loyalty information (for example geographic location, store attributes and demographics) should be added.
The clusters can be mapped to Google Earth, a useful visual tool for identifying geographical variations:
This example highlights in different colours the boundaries surrounding the clusters. Here, we can see a regional trend for one cluster which is concentrated in the South region whereas the other clusters overlap in terms of geographical distribution.
You can utilize over and under-indexing cluster analysis to quickly build shopper profiles and preferences per cluster. The table below represents a brief example of key locational and shopper trends for 3 selected clusters. As stated above, having all available attribute data to analyze here allows the system to highlight key considerations for the cluster while minimizing analysis work for category and insights managers:
At this stage in the analysis process, it is typical to assign a naming schema to clusters. In the example above, Cluster A would likely represent a more premium customer offering, so you might label it as Premium.
Align the supply chain team with assortment and space decision-making by incorporating wastage and markdown data into cluster profiling. RELEX clients identify and rank clusters based upon wastage indices and identify an over-index in products sold at markdown price. This takes cluster analysis to the next level by prioritizing store groupings that incur higher costs through spoilage or mark-down to help identify where to focus efforts first in terms of category and store group priorities that reflect the largest opportunity area through tailoring range and space with a more targeted customer offering.
Recognize the Opportunity
In many respects, clustering is simply part of a wider process. However, by defining the differentiation strategy so that the customer sits at the core, you maximize insights across a variety of business areas right at the start of the process.
The table below shows a high-level example of insights to drive range, space pricing and promotional strategies across 3 identified clusters:
You may be placing ten or twenty categories in one of three, four or five groups, but there are additional factors that may converge with the clustering analysis to determine the assortment, including Core Range, Choice, Regional & Local Products, Media Activity, Fresh Food Wastage, Availability and Logistical Efficiency & Restraints. Any store can be further balanced through proper space analysis to ensure that shelf space allotted to any given category reflects its importance within the local mix.
Clustering offers additional opportunities to teams involved in seasonal and promotions planning and new product introductions. For instance, one RELEX client has 1800 stores and trials products in six stores in each of its six clusters — a total of 36 stores or just 2% of its stores — and yet they are confident that the returns from those stores will be representative of the chain as a whole.
When stores apply these techniques and RELEX technology to category management, we routinely see growth in sales and profit and stock turns, as well as decreased spoilage, especially with fresh food.
The RELEX behavioral clustering solution is highly effective at identifying opportunities to adjust space and assortment in individual categories, resulting in increased sales.
If you want to explore how you can make best use of the principles outlined in this whitepaper, RELEX offers both the technology and the expertise. For further insights, read “Five pitfalls to avoid when creating store clusters” by Michael Falck, President, RELEX North America.