A pipeline that modernizes empirical demand estimation by representing each product with fine-tuned multimodal transformer embeddings built from its text description, image, and tabular covariates, capturing demand-relevant attributes (quality, branding, visual style) that hand-coded variables miss. These embeddings improve prediction of prices and sales ranks and, crucially, serve as high-dimensional confounder controls inside a debiased/double machine-learning estimator within a dynamic panel / difference-in-differences design. This yields more credible estimates of the price elasticity of demand and reveals strong heterogeneity in elasticity driven by product-specific features, demonstrated on Amazon.com toy-car data.
This paper modernizes empirical demand analysis by using AI-generated multimodal product representations as controls in a causal-inference pipeline. Using a dataset of toy cars on Amazon.com (text de…