Machine learning researchers increasingly rely on benchmarks to evaluate causal inference methods, yet most time-series benchmarks are observational, small-scale, or domain-specific. This leaves interventional and counterfactual estimation under-served in critical applications such as healthcare, policy evaluation, and climate science, where the ability to reason about interventions and alternative scenarios is paramount.
In response, we introduce DoTime, an open, scalable, and theoretically grounded generator for multivariate temporal structural causal models (TSCMs) with interventions. Released as the dotime PyPI package, DoTime ships with four frozen evaluation suites, designed to support rigorous and reproducible research.
DoTime advances beyond prior generators by incorporating several novel capabilities:
- Continuous-time intervention windows, enabling realistic, time-varying interventions.
- Counterfactual sampling modes with a positivity guard to ensure valid inference.
- Regime-switching SCMs, which generalize interrupted time-series models.
- Non-stationary dynamics by construction, with switching SCM parameters.
- Deterministic ramp and sinusoidal intervention profiles, placing trends and structural breaks inside the evaluation window.
These features make DoTime particularly well-suited for modeling complex, real-world dynamics.
We further demonstrate DoTime's utility as a prior for a causal foundation model reference implementation. The released suites include a training-scale snapshot of 100,000 trajectories and eight named identification structures, each with exact ground truth. Paired interventional trajectories come from the same SCM throughout, and the continuous-time suite provides shared-noise counterfactuals.
To facilitate adoption, we provide reference baseline implementations along with an evaluation harness. We also pose a falsifiable claim: interventional training yields a measurable direction-accuracy advantage over an observational model of identical capacity. Tested across three training seeds per arm, and under structure-matched evaluation on held-out episodes, the interventional prior-fitted network (PFN) demonstrates a positive gap in every structure, trajectory length, and seed tested.
With its comprehensive design and rigorous validation, DoTime fills a critical gap in the benchmarking landscape for interventional and counterfactual time-series analysis.
via ArXiv LG