Counterfactual Prediction

Counterfactual prediction asks what would happen to a person, population, or system under an alternative treatment, policy, or input, supporting decisions that ordinary forecasting cannot answer. Approaches include structural causal models, longitudinal treatment-effect models, generative methods such as diffusion, graph neural networks, and synthetic control, often combining learned patterns with causal or physical structure. Work spans clinical treatment and survival, epidemic policies, time-series interventions, and physical or market systems, with evaluation using simulated interventions, clinical cohorts, and coverage or calibration tests. The central challenge is that accurate factual predictions do not ensure valid counterfactuals: hidden confounding, limited intervention data, uncertain causal structure, and shifts between training and deployment can undermine reliability.