Causal Inference with Neural Network Models and Advanced Time Series Forecasting Techniques

This thesis addresses the challenges of causal effect estimation in complex real-world scenarios by proposing a global forecasting model (GFM) with deep neural networks (DNNs) for policy evaluation. The research develops an autoregressive causal GFM-DNN estimator to forecast counterfactual outcomes and estimate the average treatment effect (ATE) or heterogeneous treatment effect (HTE) without relying on strong assumptions. Additionally, it incorporates probabilistic forecasting to assess the distribution of the ATE and employs clustering and partitioning techniques to infer HTE. This approach enhances the interpretability of treatment analysis by identifying variables responsible for heterogeneous impacts and depicting the non-uniform effects of target interventions.