Large-scale Models of Consumer Choice and Counterfactual Inference

Опубликовано 17 сентября 2018, 22:27

Recent advances in machine learning have showed great promise in predicting consumer behavior. Yet, many firms need to make decisions about changes to the environment, such as changes in prices, promotions, and introducing new products. These interventions change the joint distribution of observables and thus off-the-shelf predictive models do not provide useful estimates of the impact of the interventions.

For several decades, economists have built models of consumer choice that rely on inferring consumer preferences from their observed choices, with the goal of estimating the impact of interventions such as price changes and firm mergers. However, traditional econometric models have been limited in scale to a relatively small number of consumers and products. In a recent series of papers, we use modern computational techniques to estimate models of consumer choice with many consumers and many products. Applications include consumer shopping basket data and consumer choices of restaurants as measured by GPS data.

We show that our methods provide better and more personalized predictions, as well as better estimates of the impact of price changes or restaurant closures in held-out data. In the baseline model, the analyst is assumed to have prior information about which products are similar to one another and thus substitutes (e.g. different brands of bottled water), but in a more complex model we estimate from the data whether pairs of products are substitutes or complements.

See more at microsoft.com/en-us/research/v...