Assessing the Importance of an Attribute in a Demand SystemStructural Model versus Machine Learning

Firms can prioritize among the product attributes based on consumer valuations using market-level data. However, a structural estimation of market demand is challenging, especially when the data are updating in real-time and instrumental variables are scarce. We find evidence that Random Forests (RF)—a machine-learning algorithm—can detect consumers’ sensitivity to product attributes similar to the structural framework of Berry-Levinsohn-Pakes (BLP). Sensitivity to an attribute is measured by the absolute value of its coefficient. We check the RF’s capacity to rank the attributes when prices are endogenous, coefficients are random, and instrumental or demographic variables are unavailable. In our simulations, the BLP estimates correlate with the RF importance factor in ranking (68%) and magnitude (79%), and the rates increase with the sample size. Consumer sensitivity to endogenous variables (price) and variables with random coefficients are overestimated by the RF approach, but ranking of variables with non-random coefficients match with BLP’s coefficients in 96% cases. These estimates are pessimistically derived by RF without parameter-tuning. We conclude that machine-learning does not replace the structural framework but provides firms with a sensible idea of consumers’ ranking of product attributes.