We develop methodology for estimation and inference using machine learning to enrich economic models. Our framework takes a standard economic model and recasts the parameters as fully flexible nonparametric functions, to capture the rich heterogeneity based on potentially high dimensional or complex observable characteristics. These "parameter functions" retain the interpretability, economic meaning, and discipline of classical parameters. In contrast to common implementations of machine learning in economics, these functions need not be predictions. We show that deep learning is particularly well-suited to structured modeling of heterogeneity in economics. First, we show how the network architecture can be easily designed to match the global structure of the economic model, delivering novel methodology that moves deep learning beyond prediction. Second, we prove convergence rates for the estimated parameter functions. These parameter functions are then the key input into the finite-dimensional parameter of inferential interest. We obtain valid inference based on a novel orthogonal score or influence function calculation that covers any second-stage parameter and any machine-learning-enriched model that uses a smooth per-observation loss function. No additional derivations are required and the score can be taken directly to data, using automatic differentiation if needed to obtain the components: the researcher need only define the original model and define the parameter of interest. A key insight is that we need not write down the influence function in order to evaluate it on the data. We apply this after deep learning, but our result can be used for any first-step estimator. Our framework covers, as special cases, well-known examples such as average treatment effects and partially linear models, but we also seamlessly deliver new results for such diverse examples as price elasticities, willingness-to-pay, and surplus measures in binary or multinomial choice models, average marginal and partial effects of continuous treatment variables, fractional outcome models, count data, heterogeneous production function components, and more. Across all these contexts inference can be made as automated as is currently available in special cases. We illustrate the utility of our framework with an application to a large scale advertising experiment for short-term loans. We show how economically meaningful estimates and inferences can be made that would be unavailable without our framework.
