We describe Causal-TS, an open-source Python library that provides a common interface and lagged-graph representation for causal discovery (structure learning) from stationary and nonstationary multivariate time series. It learns lagged and contemporaneous causal graphs with constraint-based (CDNOTS, CDNOTS+, Cedar), score-based (LGES, TGES, GES, LASSO-VAR, VARLiNGAM, Granger) and hybrid (GRACE, REAP) methods. The constraint-based methods share 17 linear and nonlinear conditional independence (CI) tests, nine with PyTorch implementations. It also provides candidate-regime detection with per-regime discovery, latent-confounder correction (LUCID), missing- and mixed-data support, a data-driven configuration recommender, a command-line interface, and an optional DoWhy bridge from a discovered graph to causal effect estimates and counterfactuals. The library is pip-installable, tested on Python 3.10-3.14, and available at https://github.com/bloomberg/causal-ts.
翻译:暂无翻译