A big-box retailer tests whether giving a product more shelf space lifts its sales, whether that lift spills over into complementary products, and whether placing complements next to each other amplifies the effect. The case walks through hypothesis testing (regression, A/B tests, change detection, exponential smoothing), a tracking-camera data source with ethical complications, and three approaches for grouping and placing products to maximise total sales: clustering, community detection (Louvain), and optimisation.
Difficulty: Intermediate
Prerequisites: Basic statistics (hypothesis testing, regression), introductory analytics or data science concepts. Familiarity with optimisation terminology is helpful but not essential.
Retailers have a fixed amount of shelf space and need to decide how to divide it among product types to maximise profit. This case study sits at the intersection of retail analytics and operations research: you start with three plausible hypotheses about shelf space, sales, and product adjacency, then apply the analytics toolkit (hypothesis testing, regression, A/B testing, change detection) to validate each one. Once validated, you use clustering and optimisation to turn those findings into actionable store-layout recommendations. The case also introduces a real-world ethics thread around in-store customer tracking. If you have covered regression, hypothesis testing, and basic optimisation earlier in the course, this ties those tools together around a single business problem.
Shelf space allocation
The decision of how many linear feet (or facings) of shelf to assign to each product type. In simple terms, this means choosing which products get more room on the shelf and which get less.
Complementary products
Products whose sales are positively linked: buying one makes a customer more likely to buy the other (e.g. chips and salsa). Think of it as items that "go together" in a shopping basket.
Adjacency effect
The hypothesis that placing complementary products next to each other on the shelf amplifies the cross-selling benefit beyond what you would see if they were in separate aisles.
Point-of-sale (POS) data
Transaction-level records of what was purchased together in a single shopping trip. This is the raw material for testing whether products are truly complementary.
A/B testing
An experimental design where you compare two treatments (here, two shelf-space amounts) by alternating them and measuring the difference in sales.
Change detection
A statistical technique for identifying whether a shift in a time series (such as sales) is genuine rather than random noise, used here after a one-time increase in shelf space.
Exponential smoothing
A time-series method that weights recent observations more heavily, useful for stripping out seasonality and trend so you can isolate the effect of a shelf-space change.
Clustering
Grouping products so that items frequently bought together land in the same cluster. The distance metric is inversely related to co-purchase frequency.
Community-finding model (Louvain algorithm)
A graph-based method that partitions nodes (products) into communities (groups) by maximising the density of edges (co-purchase links) within groups. In simple terms, it finds natural "neighbourhoods" of products that tend to be bought together.
Optimisation model (for shelf-space and layout)
A mathematical programme that maximises an objective (profit or sales) subject to constraints (minimum/maximum shelf space per product, total available space). Used both for deciding how much space each product gets and for deciding where in the store each product group sits.
The entire case rests on three hypotheses, tested in sequence.
Hypothesis 1 (shelf space drives sales): More shelf space given to a product leads to more sales of that product.
Hypothesis 2 (complementary cross-sell): More sales of one product lead to more sales of its complementary products.
Hypothesis 3 (adjacency amplifier): When complementary products are placed next to each other on the shelf, the cross-sell effect is larger than when they are separated.
The retailer's constraints are straightforward: each product type has a minimum and maximum shelf-space allowance, and each store has a fixed total amount of shelf space determined by its size. Promotions, slotting fees, and other complicating factors are set aside for this case.
Data used: Point-of-sale records showing which products were sold together in the same transaction.
Method: Hypothesis testing to confirm that the sale of one product is associated with the sale of its believed complement. Simple regression estimates the magnitude of the effect.
Result: The retailer's intuition about which products are complementary was confirmed by the data.
Data used: Point-of-sale data combined with whatever product-location data was available.
Method: Hypothesis testing confirms that adjacency increases complementary sales. Regression estimates the size of the effect, though with more error because product-location data was incomplete.
This is the hardest hypothesis to test cleanly because of confounders (seasonality, trends, external factors). Three approaches are considered:
Approach 1, A/B testing: Vary shelf space day by day between two amounts. Compare sales under the two conditions directly.
Approach 2, change detection: Increase shelf space once (or gradually) and look for a statistically significant change in the sales time series.
Approach 3, exponential smoothing (requires years of data): Use exponential smoothing to decompose the time series into seasonal, trend, and baseline components. Then test whether the baseline shifted after the shelf-space change.
Key difficulty: All three approaches risk showing correlation rather than causation. Seasonality, promotions running elsewhere in the store, and external market shifts can all confound the result.
To move from correlation to causation for Hypothesis 1, the retailer introduced in-store tracking cameras. The cameras recorded whether each customer:
Walked past a product type
Stopped at a product type
Purchased that product type
This gives a richer causal chain (exposure to shelf, then attention, then purchase), but it introduced two practical problems:
Matching difficulty: Linking a camera observation to a specific person (and then to their transaction record, credit card, name, and store path) proved technically hard.
Privacy and ethics: The act of matching tracking data to personal purchase histories raises serious privacy concerns. Customers may not know they are being tracked, and combining behavioural data with financial records amplifies the sensitivity.
The case makes an explicit point: analytics capability does not override ethical responsibility. The power to gather and link data does not automatically justify doing so.
Students often assume that a positive correlation between shelf space and sales automatically proves causation. It does not. Seasonality, promotions, and external trends can all produce the same correlation without a causal link.
Students sometimes confuse "complementary products" with "substitute products." Complements are bought together (chips and dip). Substitutes replace each other (Coke and Pepsi). The cross-sell hypothesis applies only to complements.
The adjacency effect (Hypothesis 3) is sometimes mistaken for the complementary effect (Hypothesis 2). They are distinct: Hypothesis 2 says complements boost each other's sales wherever they sit. Hypothesis 3 says the boost is larger when they sit next to each other.
Students may overlook that the tracking-camera approach has both a technical limitation (matching a camera observation to a transaction) and an ethical limitation (privacy). Both matter, and the ethical concern is not secondary to the technical one.
⚠️ Know which statistical method matches which hypothesis. Regression for estimating effect size (H2, H3). A/B testing, change detection, or exponential smoothing for H1.
⚠️ Be ready to explain why correlation does not equal causation in the shelf-space context and to name at least two confounders (seasonality, external trends).
⚠️ The ethical dimension of in-store tracking is a likely discussion or short-answer topic. Be prepared to articulate why matching camera data to purchase records is problematic, not just technically but ethically.
⚠️ Understand the role of constraints in the optimisation problem: minimum and maximum shelf space per product, total available space per store.
True or false: A/B testing is the best method for testing whether complementary products boost each other's sales.
False. A/B testing is used here for Hypothesis 1 (shelf space drives sales). Hypothesis 2 (complementary cross-sell) is tested with POS data and regression.
Fill in the blank: The key difference between Hypothesis 2 and Hypothesis 3 is that Hypothesis 3 adds the condition of physical _______ on the shelf.
Adjacency.
True or false: Exponential smoothing requires only a few weeks of sales data to isolate a shelf-space effect.
False. Approach 3 explicitly requires "enough years of data" to account for seasonal and trend effects.
Fill in the blank: The in-store tracking cameras recorded three levels of customer engagement: walked past, _______, and purchased.
Stopped at (the product type).
True or false: The retailer's optimisation problem has no upper bound on shelf space per product.
False. Each product type has both a minimum and a maximum shelf-space allowance.
Q: A retailer runs a simple regression of Product B sales on Product A sales using POS data. The coefficient is positive and significant. Which hypothesis does this support, and what does the coefficient estimate?
A: This supports Hypothesis 2 (complementary cross-sell). The coefficient estimates the magnitude of the effect: how much an increase in Product A sales is associated with an increase in Product B sales.
Q: Why might a change-detection approach to testing Hypothesis 1 produce a misleading result?
A: Because confounding factors (seasonality, external market trends, concurrent promotions) can cause sales to shift at the same time as the shelf-space change. The observed change in sales may not be caused by the additional shelf space.
Q: The retailer's tracking-camera project was technically challenging. Name the technical problem and the ethical problem it raised.
A: The technical problem was matching a camera observation to a specific person across multiple data sources (path, transaction, credit card, name). The ethical problem was that linking behavioural tracking data to personal purchase records creates significant privacy risks, especially if customers are unaware of the tracking.
Q: Explain why Hypothesis 1 is tested last, even though it appears first in the list.
A: Hypothesis 1 is the hardest to test cleanly. Hypotheses 2 and 3 can be tested with existing POS and location data using regression and hypothesis tests. Hypothesis 1 requires experimental or quasi-experimental designs (A/B tests, change detection, exponential smoothing) and faces the correlation-versus-causation problem more acutely.
Q: What are the two constraints on the shelf-space optimisation problem?
A: Each product type has a minimum and maximum amount of shelf space it can receive, and the store has a fixed total amount of shelf space determined by its physical size.
This case connects directly to regression analysis: simple regression is the tool for estimating effect magnitude in Hypotheses 2 and 3. If you have covered regression earlier, this is a concrete application of those skills.
The A/B testing and change-detection approaches for Hypothesis 1 connect to experimental design and time-series analysis. Exponential smoothing, in particular, ties back to forecasting methods covered in most introductory analytics courses.
The optimisation component (deciding shelf-space amounts and store layout under constraints) connects to linear programming and operations research. If your course covers constrained optimisation, this case provides a worked example of why you need it.
shelf space optimisation, retail analytics, complementary products, cross-selling, adjacency effect, point-of-sale data, POS data, A/B testing retail, change detection, exponential smoothing, hypothesis testing, simple regression, correlation vs causation, in-store tracking, customer tracking ethics, privacy in analytics, shelf allocation, planogram, retail layout optimisation, constrained optimisation