All Times EDT
Keywords: Policy Evaluation, Causal Effects, Engel Curve, Supervised Machine Learning, Deep Neural Nets
This paper proposes a hybrid approach to estimating demand with rich observational information and predicting counterfactual outcomes combining deep neural nets with microeconomic theory. Two theoretical assumptions, functional-separability and quasi-homotheticity, decompose price response and demand shifter into separable functions in a linear Engel curve. In doing so, the causal component, price response, can take a separate, flexible functional form from the predictive component, demand shifter, offering a framework that machine learning methods can enhance predictive performance of each component. The proposed theoretical decomposition leads to a novel identification strategy of seasonal demand shifters – minimum expenditures spent within a category/season – that alleviates endogeneity concerns from unobserved confounds (i.e., seasonal pricing) in the absence of good instruments. Synthetic data analyses show that the proposed method provides with stable predictive demand curves with respect to counterfactual price points that have not implemented before, and is robust to correlated demand and price shocks.