Synthetic Learner: Model-free inference on treatments over time

Understanding the effect of a particular treatment or policy pertains to many areas interest, ranging from political economics, marketing healthcare. In this paper, we develop non-parametric algorithm for detecting effects over time in context Synthetic Controls. The method builds on counterfactual predictions algorithms without necessarily assuming that correctly capture model. We introduce an inferential procedure detect and show testing controls size asymptotically stationary, beta mixing processes imposing any restriction set base under consideration. discuss consistency guarantees average estimates derive regret bounds proposed methodology. class may include Random Forest, Lasso, other machine-learning estimator. Numerical studies application illustrate advantages method.