DoTime: A Synthetic Benchmark Generator for Interventional and Counterfactual Time Series

Research Question and Approach

Most existing benchmarks for causal inference in time series are limited to observational data, which prevents the rigorous evaluation of interventional (do-calculus) and counterfactual reasoning. To address this, the authors introduce DoTime, a scalable generator of multivariate temporal structural causal models (TSCMs). The framework allows researchers to generate synthetic datasets with exact ground truth for interventional and counterfactual queries. The generator supports complex dynamics, including continuous-time intervention windows, regime-switching mechanisms, and non-stationary dynamics, providing a robust testbed for causal foundation models.

The authors release four frozen evaluation suites, including a 100,000-trajectory training set and specific suites for identification structures, regime switching, and continuous-time dynamics. By testing a prior-fitted network (PFN) on these suites, the study establishes a clear performance gap: models trained with interventional data consistently outperform observational models of identical capacity. This result holds across all tested identification structures, trajectory lengths, and random seeds, confirming that interventional training is essential for accurate causal effect estimation in temporal systems.

Causal reasoning over time is critical for high-stakes fields like healthcare, climate science, and policy evaluation, where decision-makers must predict how a system will evolve under specific interventions. By providing a standardized, open-source benchmark with exact ground truth, DoTime bridges the gap between static causal benchmarks and complex, time-varying real-world systems. It enables the development and validation of causal foundation models that can move beyond simple observational correlation to true causal inference.