What is store clustering and how does AI improve it?
Store clustering groups doors with similar customers and demand so each group gets the right assortment and depth. How machine learning builds clusters, what data it uses and where the method has limits.
KEY TAKEAWAYS Summary by the editors
- Store clustering groups retail doors with similar demand patterns so that assortment, size curves and stock depth can be planned per group rather than for every store individually or for the chain as one.
- Machine learning clusters stores on what customers actually buy, such as sales mix by category, price band, colour and size, rather than only on store size or region.
- Algorithms such as k-means require planners to choose the number of clusters, so business judgement remains part of the method.
- A 2019 study of two fashion retailers at the University of Porto found that the weight given to store attributes changes cluster results, and recommended a small number of clusters, roughly three to six.
- Clusters must be reviewed regularly and tested against sales results, because customer demand and store networks change over time.
Store clustering is the practice of grouping stores with similar demand so that each group receives an assortment, size mix and stock depth suited to its customers. AI improves it by clustering on detailed sales behaviour, such as the mix of categories, price points, colours and sizes each store sells, instead of relying only on size, region or turnover. The result is a manageable number of store profiles that sit between one assortment for all and a different plan for every door.
Why do fashion retailers cluster stores?
A chain with hundreds of doors cannot plan every store individually, yet one national assortment ignores obvious differences: a city flagship sells differently from an outlet, a coastal store from an alpine one. Clustering creates a middle layer. Planners build assortments, size curves and allocation rules per cluster, then adjust for individual stores where needed.
Research on assortment optimisation shows why local fit matters. Fisher and Vaidyanathan, in Management Science in 2014, developed a method to choose which products each store should carry while capping the number of distinct store assortments, accounting for customers who substitute when their first choice is missing. Implemented recommendations raised sales by 5.8% in tyres and 3.6% in car care products, the two categories where they were put into practice. The study was not in fashion, but the principle of a limited number of tailored assortments is the same one clustering serves.
How does AI build store clusters?
The usual starting point is a profile per store, described by many variables. Clustering algorithms then group stores whose profiles are close to each other. A common method is k-means, which, as the scikit-learn documentation notes, requires the number of clusters to be specified in advance and assumes clusters are roughly compact and evenly shaped, assumptions that real store networks do not always meet. Other techniques handle irregular shapes or weight variables differently.
| Variable type | Examples | What it reveals |
|---|---|---|
| Sales mix | Share of sales by category, price band, colour family | What customers in this store prefer |
| Size profile | Share of sales by size per category | Body size and fit preferences of local customers |
| Store attributes | Selling space, format, channel, concept | Physical constraints on assortment |
| Location and climate | Region, urban or rural, weather zone | Seasonal timing and climate-driven demand |
| Customer data | Loyalty demographics, tourist share | Who shops there, where data allows |
| Performance | Sell-through, full-price share, returns | How well current assortments fit |
What does research on fashion store clustering show?
A 2019 master's dissertation at the University of Porto's engineering faculty, carried out with a retail technology consultancy, tested clustering for two fashion retailers: an international chain present in 32 countries and a single-country retailer. It compared statistical and machine learning methods for weighting store attributes, then clustered stores with k-means and neural gas algorithms.
- For the international retailer, sales channel was the most important attribute, followed by country.
- Different weighting methods produced different clusters, and one method gave almost no weight to attributes with many values, such as country.
- Clustering on sales at a more detailed product level produced better clusters than at a broad level.
- For the single-country retailer, cluster quality was poor, with region the most relevant attribute.
- The author recommended a small number of clusters, roughly three to six, while noting that retailers wanting more differentiation may need more.
The lesson for practitioners is that clustering is sensitive to design choices. Which variables are included, how they are weighted and at what product level sales are aggregated can change the outcome as much as the algorithm.
How many store clusters should a retailer have?
There is no single answer. Too few clusters hide real differences; too many recreate the complexity clustering was meant to remove and stretch the range into quantities too small to buy efficiently. Statistical measures such as the silhouette and Calinski-Harabasz indexes, both used in the Porto study, help compare how well separated alternative cluster sets are. The final number is a business decision that balances local fit against buying and allocation complexity.
Online data adds a further dimension. Sales from e-commerce orders delivered to postcodes around each store can reveal demand that the store itself never captured, for example for sizes or colours it did not stock. Where data protection rules and customer consent allow, combining store and online demand by catchment area gives a fuller picture of local preferences than store sales alone.
How are store clusters used in fashion planning?
- Assortment: deciding which styles and colours each cluster carries.
- Size curves: setting the size split sent to stores in each cluster.
- Depth and allocation: deciding initial quantities per store, and how much to hold back for replenishment.
- Timing: launching seasonal categories earlier or later by climate cluster.
- Markdowns: tailoring clearance timing to clusters with different sell-through patterns.
Allocation is where clusters meet day-to-day operations. Zara's research with academics, published in Operations Research in 2015, showed the value of combining store-level forecasts with decisions on how much to send initially and how much to hold back: in a field experiment with 34 articles, season sales rose by about 2% and end-of-season unsold units fell by about 4%.
What are the limits of AI store clustering?
Clusters are built from past sales, which reflect past assortments. If a store never received petite sizes or a premium line, its history will not show demand for them, and the model may perpetuate the gap. Clusters also age: new stores, refurbishments and changes in local customers shift profiles. Retailers should review clusters at least seasonally, test changes on a subset of stores, and compare cluster-based plans with actual sell-through rather than trusting the statistical fit alone.
Frequently asked questions
What is store clustering in retail?
Store clustering is the grouping of stores with similar customer demand and characteristics, so that assortments, size mixes and stock depth can be planned per group. It sits between planning every store individually and treating the whole chain as one.
Which algorithm is used for store clustering?
K-means is the most common starting point, but it requires the number of clusters to be chosen in advance and assumes compact, evenly shaped clusters. Hierarchical clustering, neural gas and other methods are also used, often combined with techniques that weight store attributes.
How often should store clusters be updated?
Clusters should be reviewed at least once per season and whenever the store network changes significantly. Customer demand, store formats and local competition shift over time, so clusters built on older sales data gradually lose accuracy.
What data is needed for AI store clustering?
The core input is detailed sales history by store, broken down by category, price band, colour and size. Store attributes such as space, format and location add context, and customer or climate data can improve clusters where it is available and permitted.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- Universidade do Porto (FEUP): Store Attribute Weighting for Clustering in Fast Fashion (Andrade, 2019)
- scikit-learn documentation: Clustering
- IDEAS/RePEc: A Demand Estimation Procedure for Retail Assortment Optimization with Results from Implementations (Fisher and Vaidyanathan, Management Science 2014)
- IDEAS/RePEc: Initial Shipment Decisions for New Products at Zara (Gallien et al., Operations Research 2015)
