Counterfactual Prediction in Epidemic Time Series Dataset

This dataset provides a dynamic benchmark for evaluating counterfactual prediction and causal inference methods in time-series data. Generated using a calibrated, differentiable agent-based model (ABM), it provides realistic factual and counterfactual epidemic trajectories under time-varying policy interventions. The simulation integrates real-world demographic data, mobility signals, and epidemiological records to model COVID-19 transmission across 158 U.S. counties over a 168-day period (October 26, 2020 – April 11, 2021).The dataset is structured around two primary evaluation settings. The single-policy benchmark isolates the causal effect of time-varying occupation-based interventions while holding school policies constant. Expanding on this, the multi-policy benchmark captures the joint interaction effects of simultaneous, time-varying school and occupation interventions, providing data across all policy permutations (Factual/Factual, Counterfactual/Factual, Factual/Counterfactual, and Counterfactual/Counterfactual). To support robust causal inference, both settings include comprehensive static confounding variables, such as population size and age distributions, as well as dynamic time-series covariates, including daily mobility signals and compartmental disease states like the number of exposed and deceased individuals.Version 1.1: Added a CSV data dictionary. Version 1.2: Clarified the type of treatment being used in the data dictionary.