Counterfactual Predictions in Shared Markets: A Global Forecasting Approach with Deep Learning and Spillover Considerations

Talk by Klaus Ackermann (Monash University).

We introduce a novel forecasting method employing global deep learning models for estimating the causal effects of interventions across multiple units, incorporating counterfactual and synthetic control for policy evaluation in shared markets. This approach addresses potential spillover effects and leverages time series data for identification. We redefine causal effect estimation as predicting outcomes without intervention, first estimating counterfactual outcomes using high-dimensional time series data.

This process utilizes cross-correlation in time series, employing an autoregressive recurrent neural network with parameter sharing. The second stage estimates and tests the average treatment effect on the target variable for statistical significance. Demonstrated through simulations and empirical studies, our method uniquely estimates effects using pre-treatment data in scenarios where traditional control unit assumptions fail. An empirical example estimates the impact of promotional deals on US grocery store sales, showcasing the method's applicability and contribution to existing literature.