Learning Product Characteristics and Consumer Preferences from Search Data

A key idea in demand estimation is to model products as bundles of characteris- tics. In this paper, we offer an approach for jointly learning latent product characteristics and consumer preferences from search data in order to predict demand more accurately. We combine data on consumers' web-browsing histories and hotel price/quantity data to test this method in the hotel market. In two distinct applications, we show that closeness in latent characteristic space predicts competition, and parameters learned from search data substantially improve postmerger demand predictions.