Cannibalization Risk Scoring

Quantifies the degree to which two products are interchangeable from the consumer's perspective. This component uses collaborative filtering and market basket analysis to calculate a similarity score between the promoted item and every other SKU in the catalog.

• Cosine Similarity on Embeddings: Compares product description vectors and co-purchase graphs.
• Cross-Visit Analysis: Measures how often users view Product A and Product B in the same session.
• Price Band Proximity: Weighs the risk higher if the promoted item falls within a 10-15% price band of a sibling product.

A high substitutability score is the primary prerequisite for a high cannibalization risk.

A binary classification model that predicts whether a unit sold on promotion represents true incremental revenue or a diverted sale from a full-margin product. It is trained on historical transaction logs labeled via A/B test holdout groups.

• Feature Inputs: Basket composition, user's historical brand loyalty, and presence of a competing promotion.
• Gradient Boosting: Typically implemented using XGBoost or LightGBM to handle high-cardinality categorical features like SKU ID.
• Output: A probability score between 0 and 1, where 1.0 indicates a fully diverted sale.

This classifier directly feeds the final risk score, penalizing promotions that attract existing high-value customers.

Evaluates the timing of the promotion relative to the target product's lifecycle and the purchase cycles of potentially cannibalized items. The risk score is dynamically adjusted based on temporal proximity.

• Replenishment Cycle Modeling: Predicts when a user is due to repurchase a regular-price item; a promotion coinciding with this window triggers a high-risk flag.
• Launch Decay Curves: Applies a time-decay factor to the risk score for new product launches, as initial cannibalization may stabilize into a distinct market segment.
• Seasonality Overlap: Detects if the promotion overlaps with a peak season for a high-margin sibling product.

This component ensures that a promotion for snow boots in January carries a different risk profile than one in November.

Translates the probability of unit diversion into a financial impact forecast. This component does not just count lost units; it calculates the margin dollar erosion by comparing the profit of the promoted item against the profit of the item it replaces.

• Profit per Unit Delta: Margin_Promoted - Margin_Cannibalized. A negative delta indicates a destructive promotion.
• Volume Sensitivity Analysis: Models the expected number of units at risk based on the promotion's predicted uplift and the diversion probability.
• Total Erosion Forecast: Probability_of_Diversion * Expected_Volume * Profit_Delta.

This output allows revenue managers to set a Dynamic Price Floor that prevents promotions from dipping below the point of margin destruction.

Adjusts the base risk score based on the specific customer cohort targeted by the promotion. The model recognizes that cannibalization risk is not uniform across all users.

• Loyalty Tier Weighting: A discount offered exclusively to a 'Platinum' tier member who only buys full-price carries a near 100% diversion risk.
• Basket Size Context: A promotion targeting users with historically large baskets may cannibalize multiple full-price items simultaneously.
• Price Sensitivity Tag: Applies a risk multiplier for users tagged as 'Deal Seekers' by the Price Discrimination Engine, as they are most likely to trade down.

This calibration ensures that a 'Win-Back' campaign for lapsed users is scored differently than a mass coupon drop.

A continuous monitoring system that ingests post-promotion sales data to validate and recalibrate the risk model. It measures the actual cross-elasticity of demand observed during the promotion period.

• Control Group Comparison: Compares the sales velocity of non-promoted sibling products against a synthetic control group.
• Halo Effect Detection: Identifies if the promotion actually drove positive cross-sales of complementary products, partially offsetting the cannibalization loss.
• Model Retraining Trigger: If the observed diversion rate exceeds the predicted rate by a statistically significant margin, it triggers an Online Model Retraining pipeline.

This feedback mechanism closes the loop, transforming the scoring model from a static predictor into a self-correcting system.