Foundation models, particularly Prior-data fitted networks (PFNs), have proven effective in tabular causal inference. However, their application to time series is limited by the scarcity of synthetic data generators that provide interventional targets.

CausalTimePrior: A New Approach

To address this challenge, CausalTimePrior has been proposed, a framework for generating synthetic temporal structural causal models (TSCMs). This framework produces both observational and interventional time series, supporting configurable causal graph structures, nonlinear autoregressive mechanisms, and regime-switching dynamics.

Key Features

CausalTimePrior supports multiple intervention types (hard, soft, time-varying). PFNs trained on CausalTimePrior are able to perform in-context causal effect estimation on held-out TSCMs, paving the way for foundation models for causal inference in time series. This approach could have significant implications in areas such as finance, medicine, and engineering, where time series analysis is critical.