Store Clustering for Apparel Brands

What store clustering solves

Store clustering is the practice of grouping retail locations into clusters based on shared demand characteristics — so you can plan localized assortments at the cluster level rather than at each individual store.

The alternative approaches both have problems:

• →One assortment for all stores: Simple to plan but ignores real demand differences. Your urban flagship and your suburban outlet have different customers, different price sensitivity, and different size distributions. Treating them the same guarantees excess in one and stockouts in the other.
• →Unique assortment per store: Perfectly localized but impossible to plan at scale. If you have 20 doors, you can't build 20 unique assortment plans, 20 OTB budgets, and 20 allocation models.

Clustering sits in the middle: group stores with similar demand patterns together, plan at the cluster level, then allocate within clusters based on volume.

This guide is the apparel deep dive. The cross-vertical treatment — how the clustering dimension changes across footwear, home and furniture, outdoor, beauty, toys, baby and juvenile, jewelry and the rest, how many clusters a fleet can carry, and the six ways clusters break — is in Store Clustering and Localization by Vertical.

Clustering approaches

Approach 1: Volume tiers (store grading, not clustering)

Group stores by sales volume into 3 tiers:

• →Tier A (top 20% by revenue): Widest assortment, deepest inventory, first to receive new styles
• →Tier B (middle 50%): Core assortment, moderate depth
• →Tier C (bottom 30%): Narrower assortment, shallow depth, last to receive seasonal newness

This is store grading rather than clustering. What a volume tier properly sets is how deep each door goes and who receives newness first, not what each door carries; the assortment differences in the bullets above are what happens when the tier is left to decide the option list as well, which hands every smaller door a scaled-down copy of the biggest door's assortment. Volume tiers are a reasonable starting structure and become a failure mode only when they are used as a substitute for clustering; see the failure modes in Store Clustering and Localization by Vertical.

Pros: Simple, easy to implement, requires no customer data analysis.

Cons: Ignores why stores perform differently. A Tier B store in an urban market and a Tier B store in a suburban market may have the same revenue but need very different assortments.

Approach 2: Demographic clustering

Group stores by customer demographics:

• →Urban professional: Higher price point tolerance, smaller sizes, trend-forward
• →Suburban family: Mid-price, full size range, classic styles
• →College/young adult: Lower price point, trend-driven, fast-turning fashion
• →Tourist/resort: Seasonal peaks, broader size range, impulse purchases

Pros: Captures real differences in customer behavior.

Cons: Requires customer data analysis. Demographic assumptions can be wrong.

Approach 3: Demand-pattern clustering (most accurate)

Group stores by actual sales data patterns:

1. →Pull sell-through by category by store for the last 2–4 seasons
2. →Identify which stores sell similar category mixes at similar rates
3. →Group stores with correlated demand patterns

Two stores with correlated demand — even if they're in different cities — should be in the same cluster. A store in downtown Seattle and a store in downtown Portland may share more demand characteristics with each other than either shares with a suburban store in its own metro area.

Pros: Based on actual behavior, not assumptions.

Cons: Requires data analysis (but it's not complicated — correlation analysis in Excel works fine for 10–30 stores).

Approach 4: Hybrid clustering

Combine demand pattern with volume tiers and strategic importance — clustering decides the assortment, the volume tier decides the depth:

This is the most practical approach for growing brands with 10–30 stores.

How clustering changes your planning

Assortment width by cluster

Illustrative figures, not benchmarks — the shape matters, the percentages are a starting point:

• →Flagship: 100% of assortment
• →Urban core: 80% — exclude extreme price points and niche categories
• →Suburban core: 70% — focused on proven categories and core styles
• →Outlet: 40% — carryover styles, markdown inventory, core basics only

Size curve by cluster

Size demand varies by location. Urban stores typically skew smaller (XS/S/M). Suburban stores have a flatter distribution. Outlet stores may skew toward extreme sizes (what's left from other clusters).

Build cluster-level size curves from sell-through data. The mechanism is direct rather than statistical: sizes a cluster does not sell stop being sent there, so they stop ending the season on the markdown rack, and the sizes that cluster does sell stay on the floor for longer. The size curve is where cluster-level planning usually pays for itself first, because it changes what ships without changing what is bought.

Depth by cluster

Depth follows volume, but breadth and depth trade against each other inside a cluster's budget — so the widest cluster is not automatically the deepest:

• →Flagship: Wide assortment, moderate depth (showcase breadth)
• →Suburban core: Narrower assortment, deep inventory (focus on what sells)

Allocation logic

Initial allocation follows cluster rules. In-season reallocation (transfers between stores) should happen within clusters first — a transfer from one urban core store to another urban core store is more likely to sell than a transfer from urban core to outlet.

Getting started with 10–20 stores

Step 1: Pull data

Export sell-through by category by store for the last 2 seasons. You need: units sold, units received, sell-through rate, and average selling price — by category, by store.

Step 2: Visual grouping

Before running any analysis, look at the data. Can you see obvious groupings? Do certain stores consistently sell more outerwear? Do certain stores have higher sell-through rates? The human eye is surprisingly good at spotting patterns in a well-organized table.

Step 3: Formal clustering

For a more systematic approach:

1. →Calculate category contribution by store (what % of each store's sales comes from each category)
2. →Rank stores by similarity of category contribution
3. →Group stores with similar profiles

In Excel, you can do this with a simple correlation matrix. Most brands at this size settle on 3–5 clusters, and the test for whether that is the right number is not statistical fit but the merge test: take any two clusters and ask what would actually differ in next season's plan if they were one. If the answer is only a different depth, that is store grading rather than clustering.

Step 4: Validate and adjust

Test your clusters against one season of actual data. Did stores within the same cluster perform similarly? If one store consistently deviates from its cluster, it may need to be reassigned.

Cluster-level size curves in practice

The size curve is the part of clustering that most reliably changes the season's outcome, and it is also the part most often done on bad evidence. A size curve describes shape — the share of units by size — and it is a cluster property, while the number of units is a door property that allocation sets. Keeping those two apart is the whole discipline: a cluster gets a curve, a door gets a quantity, and the door's quantity is spread across the cluster's curve.

The fitting condition is strict and it is where most cluster curves go wrong. A curve may only be fitted on door-weeks in which the full run stood on the floor. Once the core sizes sell through, the sizes still selling are the ones nobody wanted, so a curve measured across broken runs describes the residue rather than the demand. In practice that means pulling receipts and on-hand alongside sales, marking the weeks each door held a complete run, and fitting on those weeks only. Where a cluster has too few complete door-weeks to fit on, it inherits the fleet curve — and the inheritance is recorded, because an inherited curve that is never re-tested quietly becomes a measured one by forgetting.

Two apparel specifics sit on top of this. Extended sizes carry their own curve rather than a stretched version of the core one, and a door that has never been allocated the extended range has no history for it at all, which a history-only cluster will read as no demand and then confirm by allocating none. Planning an extended size range covers the band's own logic. And kidswear age bands behave the same way: a curve borrowed across bands is not evidence about either.

The practical consequence for buying is the pack. A cluster curve that cannot be expressed in the pack configurations the supplier actually ships will be rounded at the DC, which means the cluster's plan and the cluster's receipts differ before the season starts. Case packs and planned depth and size and pack optimization cover where the rounding happens and how much of it a curve can absorb. A cluster whose curve requires singles from a supplier that only ships pre-packs is a cluster that will be served somebody else's curve.

Climate clusters for outerwear and transitional weight

Outerwear is the category where climate clustering stops being a refinement and becomes the plan. The week outerwear demand actually opens moves by weeks across a national fleet, and a single flow date is early in the warm doors and late in the cold ones. The cost is asymmetric: arriving early in a warm market fills the floor with weight nobody is buying and pushes the door into markdown before its season starts, while arriving late in a cold market misses the full-price weeks the category depends on. Climate clusters decide dates before they decide content, and the dates are worth more than the assortment differences underneath them.

The clustering attribute is not the region name. It is the week demand for the category historically turned on in that door, anchored to weather rather than to the retail calendar. A prior year re-read on its own weather anchor tells you the shape of the season; the same prior year read on calendar weeks tells you when the weather happened to arrive that year, which is not a plan. Doors that share a season-start week belong together even when they sit a long way apart.

Weight ladder follows from the same attribute. A cold cluster needs the full ladder — transitional, mid-weight, heavyweight, and the insulated tier at the top — while a mild cluster may need only the first two and a longer transitional window rather than a heavier one. Sending a shortened version of the cold ladder to a mild door gives it the wrong end: it receives fewer transitional options, which are the ones it actually sells, and a token heavyweight presence that marks down.

Markdown cadence differs by climate cluster too, which is easy to miss because it is set centrally. A cold cluster is still selling full-price weight in weeks when a mild cluster has been on promotion for a month, and a single fleet-wide markdown date takes margin out of the doors that did not need it. Markdown and exit strategy by vertical covers the exit side. The shoulder seasons are where the two clusters diverge most, and they are where the calendar is usually set by whoever is loudest rather than by the doors.

DTC and wholesale doors are different clustering problems

An own-door fleet and a wholesale account base look like the same clustering exercise and are not, because the data underneath them is different in kind.

Own doors produce the full record: sales, receipts, on-hand, transfers, traffic where it is captured. Availability can be reconstructed, which means censored history can be corrected before any grouping is done, and clusters can be fitted on demand rather than on what happened to be in stock. That is the ideal case, and everything in the sections above assumes it.

Wholesale accounts rarely produce that record. Most of what the brand sees is sell-in — what the account bought — which is a record of the buyer's forecast and the account's cash position, not of consumer demand. Clustering accounts on sell-in groups them by how they buy: the account that front-loads its season looks strong and the account that reorders conservatively looks weak, regardless of what sold at retail. Where accounts report sell-through, those are the only accounts whose record can be clustered directly. Non-reporting accounts are assigned by proxy from reporting accounts with comparable trade areas and door types, and the proxy is written down so the first full season reads as a test of it.

The clustering unit is different too. A multi-door account can behave as one cluster if it buys centrally and distributes internally, and the brand's cluster map then stops at the account boundary because the brand does not control what happens after. An account that lets the brand plan by door is a different case and can be clustered at door level like an own-door fleet. Assortment planning for wholesale brands covers account-level assortment work, and order fill rate for wholesale apparel covers what happens when the cluster's plan and the shipped order diverge.

Outlet doors belong in their own cluster in both channels, never as the bottom rung of a full-price ladder. Their price architecture, their markdown cadence and their carryover intake are structurally different, and mixing them into a full-price cluster drags the cluster's sell-through and markdown read toward a number that describes neither. Planning an outlet channel sets out the separation.

When clustering doesn't help

• →Fewer than 8 stores: Not enough data or variation to form meaningful clusters. Plan individually.
• →Pure DTC brand: No physical stores means no location-based clustering (though you can cluster by customer segment for marketing purposes).
• →Pop-ups and temporary locations: Too short-lived to cluster. Plan each as a one-off.

Further reading

• →Localized Assortment Planning — from clustering to cluster-specific assortments
• →Allocation and Replenishment Best Practices — allocating within clusters
• →Size and Pack Optimization — cluster-specific size curves
• →Omnichannel Assortment Planning — clustering across channels
• →Customer-Centric Merchandise Planning — using customer data to inform clustering