Scalable Demand Transfer Modeling for Massive Retail Assortments
A new restricted logit approach allows retailers to estimate substitution effects across inventories of over one million items.
Retailers can now estimate demand transfer coefficients across item universes exceeding one million products. This method calculates the percentage of demand that shifts to alternative items when a specific target product is removed from the shelf, enabling more accurate forecasting for massive assortments.
Traditional assortment optimization requires a separate demand forecast for every possible combination of items, a process that becomes computationally impossible as catalogs grow. This new approach forecasts item demand independently and then applies adjustments based on the relationships between similar items.
Inventory management systems can use these coefficients to reduce the risk of lost sales during stockouts. By quantifying exactly which products absorb the demand of a missing item, firms can optimize shelf space and prioritize the availability of high-substitution anchors.
Supply chain logistics and procurement models stand to gain from this scalability. The ability to model substitution at this scale allows for more precise safety stock levels, as the impact of a single product's absence is distributed across the broader category rather than treated as a total loss of revenue.
Retailers operating in highly fragmented categories with millions of SKUs can now move away from simplistic substitution assumptions. The empirical results from historical transaction data show that this procedure improves demand forecasting when standard substitution behaviors are present.
This shift toward scalable demand transfer estimation moves assortment optimization from a niche exercise for small product lines to a core operational tool for enterprise-scale retail.