Pith. sign in

REVIEW 6 major objections 6 minor 14 references

Causal-TS: A Python Library for Causal Discovery in High-Dimensional and Nonstationary Time Series

T0 review · 6 major / 6 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Causal-TS packages nonstationarity-aware algorithms, GPU CI tests, and regime discovery into one end-to-end Python library for causal graphs from multivariate time series.

desk verdict Solid library announcement with real engineering value; the high-dim preference claim outruns the evidence, and one flagship method is still an unposted preprint. read the letter →

arxiv 2607.24673 v1 pith:XNZALHFB submitted 2026-07-27 cs.LG stat.ME

Mohammad Fesanghary This is my paper
classification cs.LGstat.ME
keywords causaldiscoverytimeseriesnonstationarityGPUaccelerationconditionalindependenceregimePythonlibrary
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Causal-TS is an open-source Python library built so researchers can recover causal structure from high-dimensional, nonstationary multivariate time series without stitching several tools together. It ships four specialized discovery algorithms plus familiar baselines, all driven by a shared conditional-independence layer that can run on GPU. A regime module finds structural breaks, runs discovery inside each regime, and aggregates the graphs; a CLI and optional effect-estimation bridge carry the workflow from raw series to causal estimates. The claim is practical: one installable package can make temporally ordered, nonstationary discovery usable at scale where existing libraries leave gaps.

What carries the argument

A shared modular conditional-independence test layer (eight tests with PyTorch GPU paths) that every algorithm and the regime pipeline call, so skeleton discovery, orientation, and per-regime runs stay under one API.

What would settle it

Run the same high-dimensional nonstationary synthetic and real series through Causal-TS and through existing temporal libraries, and check whether Causal-TS recovers known edges more accurately or faster at scale while the regime pipeline correctly splits and aggregates graphs.

Watch

Extended reading notes

Core claim

The paper presents Causal-TS as a unified, pip-installable library that combines nonstationarity-aware temporal causal discovery, regime-conditional runs, GPU-accelerated conditional independence tests, and an end-to-end path from CLI or Python API to effect estimates, filling a niche that related Python libraries do not jointly cover.

Load-bearing premise

The library is ready and preferable for high-dimensional nonstationary practice mainly because the algorithms share one CI API and a short GPU timing table plus a feature checklist, without reported accuracy or scaling benchmarks on shared tasks.

Editorial extensions

If this is right

  • Practitioners can move from raw multivariate series to a causal graph and effect estimates inside one installable package and one CLI command.
  • Nonstationarity and regime shifts become first-class inputs rather than preprocessing afterthoughts.
  • GPU-backed CI tests lower the cost of nonlinear conditional independence checks that previously limited high-dimensional runs.
  • Domain knowledge can be injected as edge and ancestral path constraints during skeleton and orientation phases.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the shared CI layer stays stable, new discovery algorithms could be added as thin wrappers without rewriting test infrastructure.
  • Finance, climate, and neuroscience pipelines that already segment regimes could plug the same detectors into causal discovery without custom glue code.
  • Absent public accuracy and scaling tables, adoption will likely hinge on whether tutorial notebooks reproduce clean graphs on users' own series.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 6 minor

Summary. The manuscript describes Causal-TS, a pip-installable Python package for causal discovery in multivariate time series. It exposes four author-linked algorithms—CDNOTS, CDNOTS+, CEDAR, and GRACE—plus wrappers for GES, Granger causality, LASSO-VAR, and LGES, a shared GPU-enabled conditional-independence-test layer, regime-aware discovery with pluggable changepoint detectors, synthetic-data generators, a CLI, and DoWhy/Tigramite bridges. The evidence consists mainly of API sketches, an architecture diagram, feature tables, software-engineering claims, and a two-row timing experiment. The basic claim that a unified library exists is plausible, but the manuscript provides little evidence that its algorithms or CI implementations are statistically correct, accurate, or advantageous in the advertised high-dimensional nonstationary regime.

Significance. If the implementation is as described, Causal-TS would occupy a useful niche by combining nonstationarity-aware time-series discovery, regime segmentation, nonlinear CI tests, GPU support, effect-estimation bridges, and a CLI in one package. The public repository, pip installation claim, CI across Python 3.10–3.12, documentation/notebooks, and reusable CI-test layer are tangible strengths. The present evidence does not yet substantiate the stronger high-dimensional or comparative-practice implications.

major comments (6)
  1. [§3.2, Table 2; §5] The only performance evidence is KCI/SplitKCI wall-clock time at T=1000, d=20, averaged over five runs, showing 2.9x and 2.0x speedups. This does not support the paper's 'high-dimensional' positioning or Table 1's GRACE range of d<=100. Please add scaling results in d and T, including the advertised regime, memory use, GPU overhead/warm-up, precision, and clear CPU/GPU baselines.
  2. [§3.1, Table 1; §4, Table 3] No adjacency recovery, orientation accuracy, calibration, or runtime comparison is reported for CDNOTS, CDNOTS+, CEDAR, or GRACE against Tigramite, causal-learn, or the package's own GES/Granger/LASSO-VAR/LGES wrappers. Table 3 therefore records claimed capabilities rather than demonstrated usefulness. Standard synthetic experiments with data-generating processes, seeds, parameter settings, and appropriate graph-recovery metrics are needed.
  3. [§3.2] The claim that all eight CI tests share GPU-accelerated PyTorch implementations is central, but there is no statistical validation of these ports. Please report finite-sample size at nominal alpha, power on representative alternatives, p-value calibration, and numerical agreement with established CPU implementations, particularly for KCI, SplitKCI, DFCIT, RCOT, CMIknn, GCMI, and the newly introduced LinSig variant.
  4. [Table 1; References, [Fesanghary, 2026]] CEDAR is one of four flagship entry points, but its sole reference reads 'arXiv ID to be inserted upon posting.' Reviewers and users therefore cannot inspect the method behind run_cedar from the cited literature. The revision should provide a complete citable reference or a sufficiently self-contained algorithm specification and validation in this manuscript.
  5. [§4, Table 3] Several negative or check-mark entries are definition- and version-sensitive, including the distinctions between temporal discovery, nonstationarity-aware discovery, regime-aware discovery, and effect estimation. The matrix gives no software versions, inclusion criteria, citations to API documentation, or runnable examples. Please make each row operational and substantiate it, or narrow the comparison to what is directly demonstrated.
  6. [§3.3 and §5] The paper identifies the submitted software as v0.25.0 but does not provide a release tag, commit hash, permanent DOI, or archived artifact. The installation command, package name, repository, Python versions, and dependency versions should be tied to an immutable release so that the reviewed implementation can be distinguished from later repository changes.
minor comments (6)
  1. [Figure 1] Figure 1 contains the typesetting artifact 'CI T ests'; check for similar spacing problems elsewhere.
  2. [Abstract; §3.1; Table 1] The names CEDAR/GRACE are inconsistently capitalized as Cedar/Grace in the Abstract and body; use one convention throughout, including Table 1.
  3. [Abstract; §2; Figure 1] The Abstract emphasizes optional DoWhy integration, while Figure 1 lists 'DoWhy/Tigramite' for effect estimation. Please clarify exactly which integrations are optional dependencies and which functions they expose.
  4. [§3.2, Table 2] Table 2 should state the CPU model, PyTorch/CUDA versions, numerical precision, synthetic generator, number of CI-test calls, and seeds in addition to the NVIDIA L4.
  5. [§3.3] The installation command uses `causalts`, while the repository and library are named `causal-ts`; please state the PyPI project name and import-package name explicitly.
  6. [§3.3] The report of 22 pytest modules and six workflows would be more informative if accompanied by coverage, supported operating systems, and a link to the passing CI run for the reviewed release.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: software-library paper with no fitted-input predictions or load-bearing self-definitional chain.

full rationale

Causal-TS is a systems/library description paper. Its load-bearing claims are architectural and engineering (unified CI-test API, four algorithm entry points, regime pipeline, CLI, DoWhy hooks, pip installability), not numerical laws or first-principles predictions derived from fitted constants. Table 2 reports wall-clock CI-test timings; Table 3 is a feature checklist—neither is a prediction forced by a fit. Citations to CDNOTS (Sadeghi et al. 2025), CEDAR (Fesanghary 2026), and GRACE (Fesanghary & Havaldar 2026) name which methods the package wraps; that is normal attribution for a library paper and does not make the existence or API of the library equivalent to those citations by construction. There is no uniqueness theorem imported to forbid alternatives, no ansatz smuggled in as a derived form, and no renaming of a known empirical law presented as unification. Missing arXiv IDs and thin comparative benchmarks are documentation/evidence gaps, not circularity. Derivation chain is self-contained as software description; score 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 5 invented entities

As a library paper, load-bearing premises are standard causal-discovery assumptions inherited from wrapped algorithms, plus engineering choices about API unification and GPU ports. No physical constants or fitted scientific parameters underwrite a nature claim; free parameters are user hyperparameters. Invented entities are named software modules/algorithms largely pointed at companion papers.

free parameters (4)
  • num_lags / max_lag = example default 3
    User-chosen lag depth in run_cdnots / run_cedar / run_cdnots_gated examples; discovery output depends on this choice.
  • CI test hyperparameters (kernel bandwidths, k for CMIknn, significance level α)
    Not fixed by the paper; each of the eight CI tests carries implementation choices that affect edge decisions.
  • regime aggregation rule (union/majority/weighted/intersection) and detector settings (PELT penalty, etc.)
    Pluggable detectors and merge rules are chosen by the user and change the returned graph set.
  • Hard Concrete / L0 gate hyperparameters in GRACE
    Neural sparsification depends on temperature and regularization strength inherited from Louizos et al. 2018 style gates.
assumptions (6)
  • domain assumption Temporal precedence: causes precede effects, so edges are oriented with time order.
    Stated in §1 as the structural advantage of time series; underpins all algorithms.
  • domain assumption Causal Markov and faithfulness (or equivalent CI-implication) conditions so that conditional independence tests recover the skeleton.
    Implicit in PC/PCMCI-style CDNOTS/CDNOTS+ and pairwise CI methods; standard but unproven on real data.
  • domain assumption Nonstationarity captured by a (possibly multidimensional) context variable C yields additional identifying power for direction.
    §1 cites Huang et al. 2020 and Sadeghi et al. 2025; CDNOTS/CEDAR rely on this.
  • domain assumption Observed multivariate series are sufficiently sampled; no resolution of latent confounding or selection bias by the library alone.
    Explicitly acknowledged in Conclusion.
  • ad hoc to paper PyTorch GPU ports of KCI, SplitKCI, and other CI tests are statistically equivalent to CPU reference implementations for practical use.
    Table 2 reports speed only, not distributional equivalence tests across seeds and dimensions.
  • standard math Standard results on PELT changepoint optimality and related detectors correctly segment causal regime shifts when used as plugged in.
    Killick et al. 2012 cited for PELT; regime module composes these detectors without new proof.
invented entities (5)
  • Causal-TS library (unified CI layer, CLI, regime pipeline, DoWhy bridges) independent evidence
    purpose: Single installable artifact exposing multiple discovery algorithms and end-to-end workflow.
    Primary contribution of the manuscript; existence is empirically checkable via GitHub.
  • CDNOTS+
    purpose: PCMCI+-style skeleton variant of CDNOTS aimed at hub/scale-free graphs.
    Presented as a specialized algorithm in Table 1; detailed theory deferred.
  • CEDAR
    purpose: Pairwise CI method with partial distance correlation lag selection and AR-aware testing.
    Cited as Fesanghary 2026 with ‘arXiv ID to be inserted’; not independently established in this text.
  • GRACE (Hard Concrete L0 gated refinement of CDNOTS skeleton)
    purpose: High-dimensional edge refinement with neural sparse gates.
    Cited as Fesanghary & Havaldar 2026; this paper only wraps it.
  • LinSig CI test
    purpose: Linear-kernel variant of signature kernel CI testing inside the shared test layer.
    Described as ‘our’ variant of Manten et al. 2024; no separate validation section.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal-TS: A Python Library for Causal Discovery in High-Dimensional and Nonstationary Time Series." pith.science (2026). https://pith.science/paper/XNZALHFB

@misc{pith2026260724673,
  author       = {Pith},
  title        = {Pith review of: Causal-TS: A Python Library for Causal Discovery in High-Dimensional and Nonstationary Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XNZALHFB}},
  note         = {Machine review of arXiv:2607.24673}
}
read the original abstract

We describe Causal-TS, an open-source Python library for causal discovery in high-dimensional and nonstationary multivariate time series. Causal-TS provides four specialized algorithms-CDNOTS, CDNOTS+, CEDAR, and GRACE-along with wrappers for GES, Granger, LASSO-VAR, and LGES, all sharing a unified conditional independence (CI) test layer with GPU acceleration via PyTorch. A regime discovery pipeline detects structural breaks via pluggable changepoint detectors and runs discovery per regime with regime-specific parameters. A command-line interface, synthetic data generators, and optional DoWhy integration provide an end-to-end pipeline from raw time series to causal effect estimates. The library is pip-installable, tested on Python 3.10--3.12, and available at https://github.com/bloomberg/causal-ts.

Figures

Figures reproduced from arXiv: 2607.24673 by the authors.

Figure 1
Figure 1. Causal-TS architecture. All algorithms share a unified CI test layer. The causal-ts CLI wraps the full pipeline in a single command [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 2 linked inside Pith

  1. [1]

    A distribution-free conditional independence test with applications to causal discovery

    Zhanrui Cai, Runze Li, and Yaowu Zhang. A distribution-free conditional independence test with applications to causal discovery. Journal of Machine Learning Research, 23 0 (85): 0 1--41, 2022

    2022

  2. [2]

    CEDAR : Causal edge discovery for autoregressive processes

    Mohammad Fesanghary. CEDAR : Causal edge discovery for autoregressive processes. arXiv preprint, 2026. arXiv ID to be inserted upon posting

    2026

  3. [3]

    GRACE : Gated refinement for accurate causal edge discovery in high-dimensional time series

    Mohammad Fesanghary and Abhinav Havaldar. GRACE : Gated refinement for accurate causal edge discovery in high-dimensional time series. arXiv preprint arXiv:2606.23880, 2026

    arXiv 2026

  4. [4]

    Causal discovery from heterogeneous/nonstationary data

    Biwei Huang, Kun Zhang, Jiji Zhang, Joseph Ramsey, Ruben Sanchez-Romero, Clark Glymour, and Bernhard Sch \"o lkopf. Causal discovery from heterogeneous/nonstationary data. Journal of Machine Learning Research, 21 0 (89): 0 1--53, 2020

    2020

  5. [5]

    Rebecca Killick, Paul Fearnhead, and Idris A. Eckley. Optimal detection of changepoints with a linear computational cost. Journal of the American Statistical Association, 107 0 (500): 0 1590--1598, 2012

    2012

  6. [6]

    Learning sparse neural networks through L_0 regularization

    Christos Louizos, Max Welling, and Diederik P Kingma. Learning sparse neural networks through L_0 regularization. In International Conference on Learning Representations, 2018

    2018

  7. [7]

    Signature kernel conditional independence tests in causal discovery for stochastic processes

    Georg Manten, Cecilia Casolo, Emilio Ferrucci, S ren Wengel Mogensen, Cristopher Salvi, and Nico Kilbertus. Signature kernel conditional independence tests in causal discovery for stochastic processes. In Advances in Neural Information Processing Systems, 2024

    2024

  8. [8]

    Elements of Causal Inference: Foundations and Learning Algorithms

    Jonas Peters, Dominik Janzing, and Bernhard Sch \"o lkopf. Elements of Causal Inference: Foundations and Learning Algorithms. MIT Press, 2017

    2017

Show all 14 references
  1. [9]

    Sutherland, and Arthur Gretton

    Roman Pogodin, Antonin Schrab, Yazhe Li, Danica J. Sutherland, and Arthur Gretton. Practical kernel tests of conditional independence, 2024. URL https://arxiv.org/abs/2402.13196

    2024

  2. [10]

    Conditional independence testing based on a nearest-neighbor estimator of conditional mutual information

    Jakob Runge. Conditional independence testing based on a nearest-neighbor estimator of conditional mutual information. In International Conference on Artificial Intelligence and Statistics, pages 938--947. PMLR, 2018

    2018

  3. [11]

    Detecting and quantifying causal associations in large nonlinear time series datasets

    Jakob Runge, Peer Nowack, Marlene Kretschmer, Seth Flaxman, and Dino Sejdinovic. Detecting and quantifying causal associations in large nonlinear time series datasets. Science Advances, 5 0 (11): 0 eaau4996, 2019

    2019

  4. [12]

    Causal discovery from nonstationary time series

    Agathe Sadeghi, Achintya Gopal, and Mohammad Fesanghary. Causal discovery from nonstationary time series. International Journal of Data Science and Analytics, 19: 0 33--59, 2025. doi:10.1007/s41060-024-00679-7

    2025 doi

  5. [13]

    DoWhy : An end-to-end library for causal inference

    Amit Sharma and Emre Kiciman. DoWhy : An end-to-end library for causal inference. arXiv preprint arXiv:2011.04216, 2020

    2011 arXiv

  6. [14]

    Causal-learn: Causal discovery in P ython

    Yujia Zheng, Biwei Huang, Wei Chen, Joseph Ramsey, Mingming Gong, Ruichu Cai, Shohei Shimizu, Peter Spirtes, and Kun Zhang. Causal-learn: Causal discovery in P ython. Journal of Machine Learning Research, 25 0 (60): 0 1--8, 2024

    2024

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.