Interventional Time Series Priors for Causal Foundation Models
New framework generates synthetic time series data with ground-truth causal graphs and interventions.
A key bottleneck in developing AI for causal inference in time series data has been the lack of high-quality, synthetic training data that includes interventions. While benchmarks exist for observational data with known causal structures, they lack the paired interventional data—showing what happens when you actively change a variable—that is essential for training models to understand cause and effect. Researchers Dennis Thumm and Ying Chen address this with their new framework, CausalTimePrior.
CausalTimePrior is a principled data generator that creates synthetic Temporal Structural Causal Models (TSCMs). It produces both observational time series (showing normal correlations) and corresponding interventional time series (showing the effects of specific changes). The framework is highly configurable, allowing control over the underlying causal graph structure, nonlinear autoregressive mechanisms, and even regime-switching dynamics. It supports multiple intervention types, including hard, soft, and time-varying interventions, providing a rich training environment.
The researchers demonstrate that Prior-data Fitted Networks (PFNs), a promising class of foundation models for tabular data, can be effectively trained on data from CausalTimePrior. After training, these PFNs can perform 'in-context' causal effect estimation. This means the model, given a new, held-out time series dataset it has never seen before, can infer the causal relationships and estimate the effect of potential interventions directly, without needing retraining. This work, presented at the ICLR 2026 Workshop on Time Series in the Age of Large Models, establishes a clear pathway toward building general-purpose foundation models capable of robust causal reasoning in sequential data, which is critical for fields like economics, healthcare, and climate science.
- Proposes CausalTimePrior, a synthetic data generator for Temporal Structural Causal Models (TSCMs) with paired observational and interventional time series.
- Enables training of Prior-data Fitted Networks (PFNs) to perform in-context causal effect estimation on completely new, held-out time series datasets.
- Framework supports configurable causal graphs, nonlinear mechanisms, regime-switching, and multiple intervention types (hard, soft, time-varying) for robust model training.
Why It Matters
Provides the missing training data needed to build AI that can reliably infer cause-and-effect in sequential data like financial trends or patient health records.