A model-agnostic framework to extract the model-implied causal structure from any trained temporal predictor.
March 27, 2026
Original Paper
Causal-INSIGHT: Probing Temporal Models to Extract Causal Structure
arXiv · 2603.25473
The Takeaway
Practitioners can now systematically 'interrogate' black-box time-series models to see which features and lags they actually rely on for predictions. This provides a rigorous post-hoc way to validate temporal models against known domain causal structures.
From the abstract
Understanding directed temporal interactions in multivariate time series is essential for interpreting complex dynamical systems and the predictive models trained on them. We present Causal-INSIGHT, a model-agnostic, post-hoc interpretation framework for extracting model-implied (predictor-dependent), directed, time-lagged influence structure from trained temporal predictors. Rather than inferring causal structure at the level of the data-generating process, Causal-INSIGHT analyzes how a fixed,