Store Clustering
“Which stores behave similarly for promotional planning?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which stores behave similarly for promotional planning?
Retailers run 150-300 promotional campaigns per year, typically applying the same offers across all stores or using broad regional groupings. But two stores 5 miles apart can have completely different customer bases, price sensitivities, and category preferences. A suburban family-focused store responds to BOGO deals on bulk items, while an urban store near a college campus responds to single-serve discounts. Mismatched promotions waste 20-30% of promotional budgets ($40-80M annually for a large chain) and can actually decrease margin when high-margin items are discounted at stores where they would have sold at full price.
Quick answer
Store clustering groups locations by behavioral similarity rather than geography. Two stores 5 miles apart can have completely different customer bases, promotional responses, and category mixes. Traditional approaches cluster on static attributes (region, format, square footage), missing the behavioral signals that actually predict promotional response. A relational model generates store embeddings from the full graph of transactions, customer demographics, category performance, and promotional response history, producing 8-12 natural clusters that predict which promotion types work where. On SAP SALT benchmarks, relational approaches hit 91% accuracy vs 75% for XGBoost on segmentation tasks.
Approaches compared
4 ways to solve this problem
1. Geographic and format-based grouping
Group stores by region, urban/suburban/rural format, and square footage. The default in most retail planning systems.
Best for
Initial segmentation when no behavioral data is available. Gets you from 500 individual store plans to 10-15 regional templates.
Watch out for
Two urban stores in the same city can have completely different customer bases. An urban store near a college campus responds to single-serve discounts while one near a financial district responds to premium product promotions. Geography alone misses this.
2. K-means clustering on aggregated metrics
Cluster stores on features like average basket size, category mix percentages, transaction volume, and customer demographics using K-means or hierarchical clustering.
Best for
Teams with clean store-level performance data who need a quick, interpretable segmentation for planning purposes.
Watch out for
Aggregated metrics lose the temporal and promotional response signals. Two stores with the same average basket size may respond completely differently to BOGO vs percentage-off promotions. K-means also forces you to pre-select the number of clusters.
3. XGBoost promotional response models per store
Train individual models for each store to predict promotional lift, then cluster stores by model similarity.
Best for
Retailers with rich promotional testing history across many stores and campaign types.
Watch out for
Training 500 individual models is expensive and data-hungry. Stores with limited promotional history get poor models. No way to transfer learning from data-rich stores to data-poor ones. SAP SALT shows 75% accuracy ceiling.
4. KumoRFM (relational foundation model)
Connects stores to their full behavioral graph: transactions, customer demographics, category performance, promotional response history, and geographic context. Generates store embeddings that capture behavioral DNA.
Best for
Chains with 100+ stores where promotional spend optimization is a major margin opportunity and behavioral differences between stores are significant.
Watch out for
Requires transactional data linked to stores. If your data is aggregated at the regional level without store-level granularity, the clustering will reflect regions, not behaviors.
Key metric: SAP SALT segmentation: relational 91% vs XGBoost 75%. Behavioral clustering improves promotional ROI by 25-40% vs geographic grouping.
Why relational data changes the answer
Traditional clustering uses store-level aggregate metrics: average basket size, revenue, foot traffic. These features cannot capture why two stores behave differently. Store S-14 responds strongly to percentage-off promotions on organic products because its customer base is young professionals with low price sensitivity and high organic affinity. Store S-22 responds to BOGO on bulk items because its customers are families with 3.4-person households. These behavioral drivers live in customer_demographics, category_performance, and promo_response_history tables.
A relational model connects stores to their full behavioral context and generates embeddings that encode this multi-table DNA. Stores S-14 and S-38 cluster together despite being in different states because they share the same customer profile, category mix, and promotional response patterns. This behavioral clustering produces 25-40% higher promotional ROI than geographic grouping because the clusters actually predict which promotion types will work, not just which stores are physically close.
Clustering stores by geography is like grouping students by their seat number in a classroom. Student 14 sits next to Student 15, so they must learn the same way. A relational model is like grouping students by their learning style: visual learners respond to diagrams regardless of where they sit, auditory learners respond to lectures, and hands-on learners respond to projects. The seat number tells you nothing about how to teach them. The store address tells you nothing about which promotions will work.
How KumoRFM solves this
Relational intelligence built for retail and e-commerce data
Kumo builds a relational graph connecting stores to their transaction patterns, customer demographics, product category performance, promotional response history, and geographic context. The model generates embeddings that capture each store's behavioral DNA. Stores S-14 and S-38 cluster together because they share similar customer demographics, promotional response patterns, and category mix, even though they are in different states. Kumo surfaces 8-12 natural clusters that respond predictably to specific promotion types, replacing 500+ store-level rules with a small number of targeted strategies.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
STORES
| store_id | name | region | format | sqft | demo_index |
|---|---|---|---|---|---|
| S-14 | Union Square Market | West | Urban | 42,000 | Young Professional |
| S-22 | Midtown Grocery | Northeast | Suburban | 55,000 | Family |
| S-38 | SoHo Fresh | Northeast | Urban | 38,000 | Young Professional |
CATEGORY_PERFORMANCE
| store_id | category | revenue_share | growth_yoy | avg_basket |
|---|---|---|---|---|
| S-14 | Organic & Natural | 28% | +18% | $42 |
| S-14 | Ready Meals | 22% | +12% | $18 |
| S-22 | Bulk & Family Pack | 35% | +8% | $65 |
PROMO_RESPONSE_HISTORY
| store_id | promo_type | avg_lift | avg_margin_impact | best_category |
|---|---|---|---|---|
| S-14 | % Off | +22% | +$1,200 | Organic |
| S-14 | BOGO | +8% | -$400 | Snacks |
| S-22 | BOGO | +35% | +$2,800 | Bulk Items |
CUSTOMER_DEMOGRAPHICS
| store_id | avg_age | avg_income | household_size | price_sensitivity |
|---|---|---|---|---|
| S-14 | 31 | $92,000 | 1.8 | Low |
| S-22 | 42 | $78,000 | 3.4 | Medium |
| S-38 | 29 | $88,000 | 1.6 | Low |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT MOST_COMMON(STORES.CLUSTER_ID) FOR EACH STORES.STORE_ID
Prediction output
Every entity gets a score, updated continuously
| STORE_ID | NAME | CLUSTER | CLUSTER_LABEL | TOP_PROMO_TYPE |
|---|---|---|---|---|
| S-14 | Union Square | C-1 | Urban Health-Forward | % Off Organic |
| S-38 | SoHo Fresh | C-1 | Urban Health-Forward | % Off Organic |
| S-22 | Midtown Grocery | C-3 | Suburban Family | BOGO Bulk |
Understand why
Every prediction includes feature attributions — no black boxes
Store S-14 (Union Square Market)
Predicted: Cluster C-1: Urban Health-Forward
Top contributing features
Organic category revenue share
28%
28% attribution
Customer demographic profile
Young Pro, 1.8 HH
24% attribution
Promotional response pattern
% Off > BOGO
22% attribution
Price sensitivity index
Low
15% attribution
Store format and size
Urban, 42K sqft
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about store clustering
How many store clusters should a retailer have?
Most retailers land on 8-12 behavioral clusters, regardless of total store count. Fewer than 6 clusters are too broad (you are still using one-size-fits-all promotions). More than 15 clusters create operational complexity that planning teams cannot manage. The right number emerges from the data: when splitting a cluster further does not produce meaningfully different promotional response patterns, you have found the natural granularity.
Can store clustering improve promotional ROI?
Yes, significantly. Retailers typically waste 20-30% of promotional budgets on mismatched promotions: running BOGO deals at stores where customers prefer percentage-off, or discounting premium items at stores where they would sell at full price. Behavioral clustering matches the right promotion type to the right store cluster, recovering $40-80M in wasted spend annually for a 500-store chain. The ROI comes from eliminating negative-margin promotions, not just amplifying positive ones.
How often should store clusters be updated?
Quarterly is the standard cadence. Customer demographics shift slowly, but promotional response patterns can change with competitive entries, format renovations, or neighborhood changes. A store near a new apartment complex may shift from a 'suburban family' cluster to an 'urban mixed' cluster within 6-12 months. Relational models detect these shifts automatically because they track the behavioral signals, not just static attributes.
Bottom line: Increase promotional ROI by 25-40% by targeting store clusters with the right promotion types, recovering $40-80M in wasted promotional spend for a 500-store chain.
Related use cases
Explore more retail & e-commerce use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
