ABSTRACT
In retail, whenever an item is unavailable, customers may decide to purchase other items instead. We define the substitution rate between product A and B as the probability that, if product A is unavailable, a customer who originally wanted product A would purchase product B instead. Normally, estimating substitution rates requires both inventory and Point-of-Sale (POS) data. However, many retailers may not have inventory data with the expected granularity and reliability. Our approach for estimating substitution rates uses only POS data, produces confidence estimates and can be implemented at scale using methods such as Stochastic Variational Inference (Hoffman, Blei, Wang & Paisley, 2013). We test our approach using simulated POS data and find that our estimates have sufficient accuracy without using inventory data. We further demonstrate that the resulting substitution rates can be used to simulate changes in inventory and assortment policies and gauge impact on revenue and profit.
