Econometrics Seminar Series - Klaus Ackermann (Monash University)

Title: Forecasting Synthetic Control

Abstract: We propose Forecasting Synthetic Control (FSC), a framework that estimates causal effects in panel data by training a single global forecasting model on pre-treatment outcomes from all units and using it to predict counterfactuals for treated units. Because only pre-treatment data enter the model, FSC remains valid under a Weaker SUTVA assumption that permits post-treatment spillovers on control units. The framework is agnostic to the forecasting function; we study neural network and classical  instantiations and establish consistency for feed forward neural networks under temporal dependence via beta-mixing theory. In Monte Carlo simulations with positive spillovers, concurrent-control methods (DiD, CausalImpact, ArCo, SCM) miss or severely bias the effect, while FSC recovers it. An empirical application to Nielsen retail scanner data detects a significant promotional sales increase that benchmarks miss.