Assortment Cannibalization Detection

The core statistical objective is to measure the true incremental revenue generated by a new SKU, net of the sales it diverts from existing items. This requires establishing a synthetic control group using historical baselines or holdout geographies to estimate what sales would have been without the introduction. The difference between observed total category sales and the predicted baseline reveals the net lift, distinguishing a genuine market expansion from a zero-sum redistribution.

This quantifies the rate at which customers switch from Product A to Product B when B is introduced or promoted. A high cross-elasticity indicates strong cannibalization risk. Detection systems model this using discrete choice theory, calculating the probability of a user selecting a substitute based on shared attributes:

• Attribute Distance: Measuring similarity in price, brand, size, and color.
• Co-View Analysis: Tracking which items are compared during the same session.
• Sequential Purchase Patterns: Identifying if users who bought A historically now buy B.

A Product Affinity Graph maps co-purchase and co-view relationships as weighted edges between product nodes. Cannibalization manifests as a measurable structural shift in this graph. When a new item is introduced, the system monitors for edge weight decay on existing complementary or substitute relationships. A sudden drop in the co-purchase frequency of two previously associated items signals that the new entrant has disrupted an established purchasing pattern, capturing demand that previously belonged to an incumbent.

Cannibalization is often localized. A new product in a specific Micro-Merchandising Zone may steal sales from an existing item only within that cluster. Detection requires Geospatial Demand Clustering to compare sales velocity shifts in test zones against control zones. This isolates whether a decline in Product A is due to the introduction of Product B or an external factor like a local competitor's promotion. The system maps demand transference vectors to visualize where sales are migrating geographically.

Initial cannibalization rates are often inflated by a novelty effect, where customers trial a new item temporarily before returning to their habitual purchase. A robust detection system distinguishes transient trial from permanent share shift by applying time-series decomposition. It separates the signal into trend, seasonality, and residual components, monitoring whether the cannibalization coefficient stabilizes or decays after a defined maturation period, typically 4-8 weeks post-launch.

Simple pre-post sales comparisons are confounded by seasonality and promotions. Cannibalization detection employs Difference-in-Differences (DiD) analysis, a quasi-experimental technique. The system compares the change in sales of the incumbent product in a 'treated' group (exposed to the new SKU) against a 'control' group (not exposed) over the same period. The DiD estimator isolates the causal impact of the new introduction, controlling for time-variant external factors affecting both groups equally.