Triplet Embeddings for Demand Estimation

We propose a method to augment conventional demand estimation approaches with crowd-sourced data on the product space. Our method obtains triplets data ("product A is closer to B than it is to C") from an online survey to compute an embedding—i.e., a low-dimensional representation of the latent product space. The embedding can either replace data on observed characteristics in mixed logit models, or provide pairwise product distances to discipline cross-elasticities in log-linear models. We illustrate both approaches by estimating demand for ready-to-eat cereals; the information contained in the embedding leads to more plausible substitution patterns and better fit.