Pith. sign in

REVIEW 3 major objections 4 minor 38 references

AutoCause claims that recording expert decisions yields reproducible causal graphs, and that majority-supported links are more precise on synthetic benchmarks but not on real river topology.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 01:01 UTC pith:RWFCKGDW

load-bearing objection Useful workflow integration, honestly documented; the headline CausalRivers claim rests on six and seven links and should be reframed. the 3 major comments →

arxiv 2608.00198 v1 pith:RWFCKGDW submitted 2026-07-31 cs.LG

AutoCause: A Python framework that automates expert decisions in environmental time-series causal discovery

Marco Ruiz , Miguel Arana-Catania , David R. Ardila , Rodrigo Ventura This is my paper
classification cs.LG
keywords causal discoverytime seriesenvironmental modelingmulti-method consensusassumption diagnosticsopen-source frameworkreproducibilityreference graph
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

AutoCause is an open-source workflow that records the six decisions environmental analysts make when applying causal-discovery methods to time series, and grades every discovered link by how many of four methods agree on it. The paper's central empirical claim is that majority-supported links are more precise than single-method links on two synthetic benchmarks, but that this ordering reverses on a real river network where the reference graph is the physical topology. That reversal shows that consensus support is a useful precision filter only when the reference captures all genuine dependencies. The framework is not a new causal algorithm; it is an auditable decision-and-evaluation layer around established methods, with pre-discovery diagnostics, adaptive test selection, non-causal baselines, and reproducibility logs.

Core claim

AutoCause formalizes six expert decisions—method choice, conditional-independence test, lag horizon, sample-size adequacy, multiple-testing control, and evidence interpretation—into a single Python workflow. On 145 datasets, links detected by at least three of four causal methods (Tier-1) have higher precision than single-method links on DGP-Atlas (0.848 vs 0.126) and TimeGraph (0.694 vs 0.300), but on CausalRivers Tier-1 precision is 0.425 while Tier-2 and Tier-3 are 0.667 and 0.571. The paper attributes this inversion to the reference graph: the river topology omits rainfall, snowmelt, reservoir operations, and tributary inflows that can create real statistical dependence. AutoCause theref

What carries the argument

The central mechanism is method-count consensus tiering combined with a pre-discovery causal-audit module. The audit converts six assumption-violation risks into method, preprocessing, and conditional-independence-test recommendations; the consensus step then counts how many of the four methods detect each link and assigns Tier-1 (≥3 methods), Tier-2 (exactly 2), or Tier-3 (exactly 1) support. The reference graph supplied for evaluation is the object that makes precision measurable and determines whether consensus support appears valuable.

Load-bearing premise

The CausalRivers evaluation counts any link not in the physical river topology as a false positive, even though the topology omits rainfall, snowmelt, reservoir operations, and tributary inflows that can create genuine statistical dependence between stations.

What would settle it

Add a measured common driver, such as catchment-wide precipitation, as an observable variable and rerun AutoCause on the same river subgraphs; if Tier-1 precision against a hydrologically expanded reference rises above the Tier-2 and Tier-3 estimates, the inversion is an artifact of the incomplete reference. Alternatively, on synthetic data with a hidden common cause, compare Tier-1 precision with and without the hidden variable; a similar drop would explain the river result as latent confounding.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single logged workflow makes environmental causal-discovery analyses reproducible and comparable across datasets, since every configuration choice is recorded and can be overridden by domain knowledge.
  • Majority-support tiering is a reliable precision filter only when the reference graph is complete; on real-world data with unobserved drivers, Tier-1 links should be read as repeated detection for further inspection, not as validated causal edges.
  • Audit-triggered preprocessing can rescue methods: on irregular-sampling DGP-Atlas data, PCMCI+ rises from F1 0.00 to 0.78 after interpolation, and deseasonalization recovers all skeleton edges on a trend-seasonal TimeGraph case.
  • No single method dominates across all regimes; a per-dataset oracle gains +0.104, +0.153, and +0.048 F1 over PCMCI+ alone on the three benchmarks, supporting multi-method ensembles as a robustness check.
  • The support tiers are descriptive labels, not posterior probabilities or causal validity scores; they prioritize which links deserve further physical interpretation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The CausalRivers inversion implies that consensus algorithms can share blind spots: if all methods detect a common forcing or an indirect path, agreement may reflect a shared error rather than a direct causal link. A synthetic test that adds a hidden common driver and measures Tier-1 precision collapse would directly probe this mechanism.
  • The paper leaves open whether cross-family agreement carries information beyond vote count; an equal-vote comparison of same-family versus cross-family pairs would settle this, and the paper itself flags it as untested.
  • The competitive performance of the lagged-correlation baseline on river data hints that for some environmental questions, association screening combined with strong domain constraints may be nearly as informative as causal discovery, especially when the reference graph is incomplete.
  • The auditable-decision protocol could transfer to other application domains by adding new method families (e.g., continuous optimization, neural causal discovery) and new reference types, without changing the recorded-decision structure.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents AutoCause, an open-source Python workflow for time-series causal discovery that wraps VAR-Granger, VARLiNGAM, transfer entropy, and PCMCI+, adds non-causal correlation and random-forest references, records six expert configuration decisions, and assigns consensus-support tiers by method count. It evaluates on 145 datasets from DGP-Atlas, TimeGraph, and a topology-derived CausalRivers reference, reporting method F1, complementarity, and tier-level precision. The headline empirical claim is that majority-supported (Tier-1) links are more precise than single-method links on the synthetic benchmarks (DGP-Atlas 0.848 vs 0.126; TimeGraph 0.694 vs 0.300), but that the ordering reverses on CausalRivers (Tier-1 0.425 vs Tier-2 0.667 and Tier-3 0.571).

Significance. If the results hold, the contribution is a valuable reproducibility and auditing layer around established causal-discovery methods, not a new discovery algorithm. The strengths of the manuscript are substantial: a released, versioned Python package with pinned dependencies, a file manifest tracing every reported number, machine-readable configuration logs, bootstrap intervals for method-level F1, inclusion of non-causal reference models, and unusually candid discussion of the limits of consensus tiering and of the CausalRivers topology reference. The paper also avoids overclaiming algorithmic novelty. However, the central consensus-precision claims need stronger statistical support and a clearer separation between development and independent evaluation before they can serve as the paper's main empirical message.

major comments (3)
  1. [§4.4, Fig. 9] The CausalRivers tier-precision inversion rests on Tier-2 and Tier-3 estimates computed from n=6 and n=7 links, with no confidence intervals or bootstrap. One true/false reclassification changes Tier-2 precision by 0.167 and Tier-3 by 0.143; Wilson 95% intervals for the observed 4/6 and 4/7 counts overlap substantially with the Tier-1 interval (0.425, n=287). The paper acknowledges in §2.4 that the topology reference omits rainfall, snowmelt, reservoir operations, and tributary inflows, so links outside the topology may be genuine couplings scored as false. The abstract's 'but not against river topology' therefore overstates the support. I request uncertainty quantification on all tier-precision estimates and a softened or properly conditioned claim (e.g., 'inconclusive on this reference').
  2. [§3.4, §4.1, §4.4] DGP-Atlas is the development ground for the pre-discovery extensions and decision thresholds, yet its 97 datasets are used as primary evidence for the 'synthetic benchmarks' half of the headline claim. The paper states that the three extensions 'emerged during DGP-Atlas validation' (§3.4), that the thresholds are 'implementation defaults' informed by DGP-Atlas (§4.1), and that family-level agreement in Table A.2 is an 'internal consistency check rather than independent external validation.' Under these conditions, DGP-Atlas tier-precision numbers cannot serve as independent confirmation of the workflow. The only independent synthetic evidence is TimeGraph, with n=49, 31, 20 for the three tiers and no uncertainty intervals. I recommend separating development and evaluation sets, or explicitly recasting DGP-Atlas as an illustrative stress test and basing the synthetic claim on independentl
  3. [§4.4, Table 11] Even for the supported orderings, precision differences between tiers are reported as point estimates only. For TimeGraph, Tier-1 precision 0.694 (n=49) vs Tier-3 0.300 (n=20) is suggestive but not quantified; for DGP-Atlas, Tier-1 0.848 (n=1291) vs Tier-2 0.344 (n=209) is a large difference, but bootstrap intervals across datasets (as already used for method F1 in Fig. 3) would make the claim rigorous. I request dataset-level bootstrap or Wilson intervals for the consensus-support precision table, and the same for the vote-threshold comparison.
minor comments (4)
  1. [Fig. 1 vs §3.1/Table 6] Stage 5 in Figure 1 is labeled 'regime-weighted ensemble scoring,' but the text and Table 6 describe an unweighted method-count tier rule. Please align the figure with the implemented rule or explain the weighting if it exists.
  2. [§A.2] The sentence 'PCMCI+ has mean F1=0.814 on the same skeleton endpoint' appears twice in consecutive sentences. One repetition should be removed.
  3. [Fig. 9] The dashed 'chance' line is said to represent a 'common reference prevalence,' but the definition is not given. Please state explicitly how the reference prevalence is computed and whether it is the same quantity across the three benchmarks.
  4. [§4.1] The operational sample-size thresholds (50/200/75/120) are described as 'informed by published guidance and observed estimator behavior'; citing the specific sources in the text or appendix would help readers judge the safety margins.

Circularity Check

0 steps flagged

No load-bearing circularity; the central claims rest on external benchmarks and the in-sample DGP-Atlas development is explicitly labeled internal.

full rationale

AutoCause's central empirical claims are benchmark evaluations, not derivations that reduce to their inputs. The headline tier-precision ordering is shown on TimeGraph (Ferdous et al., 2025), an external benchmark, as well as on DGP-Atlas; the CausalRivers inversion uses an external topology-derived reference (Stein et al., 2025). Thus the main 'majority precision is reference-dependent' claim does not reduce to any fitted parameter or self-defined quantity. The closest issue is that three causal-audit changes 'emerged during DGP-Atlas validation' (Section 3.4) and the thresholds are 'implementation defaults rather than universally validated cutoffs' (Section 4.1), so DGP-Atlas results are partially in-sample. The paper explicitly confines DGP-Atlas recommendation accuracy to 'an internal consistency check rather than as independent external validation' (Section 3.4) and states the experiments 'do not independently validate' the automated decision rules (Section 4). This is transparent in-sample optimism, not a prediction forced by construction. The self-citations to Ruiz et al. (2026) for causal-audit and DGP-Atlas are code- and data-released and are not used to forbid alternatives or to define the target result. The small CausalRivers Tier-2/Tier-3 denominators (n=6, 7) and the incompleteness of the topology reference are statistical validity threats, not circularity. Accordingly, no circular step is identified.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claims rest on benchmark ground truths, wrapped-method assumptions, and hand-set thresholds. The most important are the DGP-Atlas self-benchmark, the CausalRivers topology reference, and the fixed tau_max/risk thresholds.

free parameters (4)
  • tau_max = 5 = 5
    Fixed maximum lag across all 145 datasets in Section 4.1; chosen for cross-method comparability and to cover TimeGraph's true lag of 2 and a 30-hour river travel time. It bounds every method's search and therefore all F1/precision numbers.
  • causal-audit risk thresholds = low <0.30, moderate 0.30-0.60, high >=0.60
    Used by the decision tree for method/CI-test/preprocessing routing; developed with DGP-Atlas in view and not independently validated (Sections 3.4, 4.1).
  • sample-size adequacy thresholds = ParCorr 50, CMIknn 200, VAR-Granger 75, VARLiNGAM 120
    Implementation defaults determining warnings/fallbacks; chosen by hand from published guidance and estimator behavior (Section 3.3).
  • consensus tier cutoffs = Tier-1 >=3, Tier-2 =2, Tier-3 =1
    Definitional vote thresholds; alternatives are explored in Table 11, so they are not hidden, but they determine the precision comparisons in Section 4.4.
axioms (6)
  • domain assumption Causal Markov condition and faithfulness hold for the constraint-based output (PCMCI+).
    Required for the CI-test pruning logic the workflow delegates to Tigramite (Section 3.3; Runge et al. cited).
  • domain assumption Causal sufficiency: all relevant common drivers are observed for the evaluated methods; LPCMCI was excluded by wall-time.
    Wrapped methods assume no unobserved confounders except LPCMCI; Section 5.4 says no default can establish this.
  • domain assumption DGP-Atlas VAR(1) coefficient matrix is the true causal graph.
    Used as ground truth for 97 of 145 datasets (Section 2.4); generated by the same authors' atlas.
  • domain assumption CausalRivers directed topology is a valid reference for false-positive scoring.
    The paper acknowledges it omits shared rainfall, snowmelt, reservoirs, and tributaries (Section 2.4); still used to compute FDR.
  • standard math Standard statistical diagnostics (RESET, distance correlation, Shapiro-Wilk, BH, IAAFT surrogates) have valid operating characteristics at T=500-1460.
    No calibration or sensitivity analysis for these; relies on standard literature (Section 3.3, Appendix A.4).
  • domain assumption TimeGraph ground-truth graphs correspond to the evaluated method outputs.
    Structured benchmark with lagged/contemporaneous edges; used as an external synthetic reference (Section 2.4).

pith-pipeline@v1.3.0-alltime-deepseek · 34514 in / 15550 out tokens · 146151 ms · 2026-08-04T01:01:04.884258+00:00 · methodology

0 comments
read the original abstract

Environmental time-series causal discovery requires expert decisions about method choice, conditional-independence tests, lag horizons, sample-size adequacy, multiple-testing control, and evidence interpretation. Applied inconsistently across datasets, these choices yield graphs that cannot be compared, reproduced, or audited. We present AutoCause, an open-source Python workflow that records each decision, derives defaults from an extended causal-audit module, and admits domain-informed overrides. The workflow wraps four established causal-discovery methods from three families, adds non-causal reference models, and grades links by method-count support. On 145 datasets from DGP-Atlas, TimeGraph, and a topology-derived CausalRivers reference, the methods recover complementary parts of the reference graphs. Majority-supported links are more precise than single-method links on the synthetic benchmarks but not against river topology. AutoCause converts inconsistent expert practice into an auditable, repeatable analysis; causal interpretation remains with the analyst. Available at https://github.com/marcoruizrueda/autocause.

Figures

Figures reproduced from arXiv: 2608.00198 by David R. Ardila, Marco Ruiz, Miguel Arana-Catania, Rodrigo Ventura.

Figure 1
Figure 1. Figure 1: AutoCause pipeline. A complete call runs Stages 0–7. Optional stages can be selected subject to their input dependencies. 6 data_df = df , 7 output_dir =" results / my_experiment ", 8 ) 9 print ( result [" consensus "][" tier1_edges "]) # majority - supported candidate links Marco Ruiz et al.: Preprint submitted to Elsevier Page 8 of 33 [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pre-discovery summary for a CausalRivers subgraph comprising five Bavarian gauging stations on the Donau and Abens. The panels show the standardized discharge series; six assumption-violation risk scores with 95% bootstrap uncertainty intervals, grouped as low (< 0.30), moderate (0.30–0.60), and high (≥ 0.60); variable-specific autocorrelation functions with 95% significance bands and inset estimates of 𝑇e… view at source ↗
Figure 3
Figure 3. Figure 3: Mean F1 score per method per benchmark with dataset-level bootstrap 95% intervals from 2000 resamples. Lagged correlation and the random-forest predictive baseline are non-causal references. Source: make_figures.py and the per-dataset metric files specified in Section 4.1. (Hyvärinen et al., 2010; Shimizu et al., 2006). PCMCI+ reaches F1 = 0.87 on the clean VAR family F1 and F1 = 0.91 on F8. Preprocessing … view at source ↗
Figure 4
Figure 4. Figure 4: Mean F1 score for each method across the 10 DGP-Atlas families, including the lagged-correlation and predictive reference models. The preprocessing stage applies interpolation to F3 and F9 and removes seasonality from F6. VARLiNGAM achieves the highest mean F1 in six families, while the lagged-correlation baseline remains slightly below the best-performing causal method in each family. The connection betwe… view at source ↗
Figure 5
Figure 5. Figure 5: Effect of audit-driven preprocessing on two cases. Both rows run PCMCI+ (ParCorr, Benjamini–Hochberg FDR); the two right-most columns differ only in whether the audit-recommended preprocessing is applied. Top row, TimeGraph C1 (trend-seasonal): without deseasonalization three of four skeleton edges are recovered; with it all four are recovered, with one contemporaneous link oriented against the equivalence… view at source ↗
Figure 6
Figure 6. Figure 6: F1 score for each method across the 18 TimeGraph categories, evaluated on unordered skeleton pairs. Lagged correlation and the predictive model are included as non-causal reference methods. The figure was generated with make_figures.py from the TimeGraph per-family result files. Random connected Root-cause chain Confounder structure 0.0 0.2 0.4 0.6 0.8 1.0 Rate PCMCI+ TPR PCMCI+ FDR VAR-Granger TPR VAR-Gra… view at source ↗
Figure 7
Figure 7. Figure 7: Skeleton TPR (solid) and graph FDR (hatched) for CausalRivers, grouped by subgraph topology, with bootstrap 95% intervals across subgraphs. The three classes are the CausalRivers sampling strategies: randomly connected five-station subgraphs, root-cause chains in which the longest directed path contains all five stations, and confounder structures in which at least one station has two or more downstream st… view at source ↗
Figure 8
Figure 8. Figure 8: Method-count support for one DGP-Atlas instance from the non-Gaussian F8 family (dgp_002), containing six variables and nine reference connections. All results are evaluated on the unordered skeleton. Panel (a) shows the reference graph. Panel (b) shows Tier-1 links detected by at least three methods, recovering 8 of 9 reference links with precision 0.89. Panel (c) shows Tier-2 links detected by two method… view at source ↗
Figure 9
Figure 9. Figure 9: Consensus-support precision per benchmark. On DGP-Atlas and TimeGraph, majority-supported Tier-1 exceeds Tier-2 and Tier-3. On CausalRivers the ordering reverses; the Tier-2 and Tier-3 estimates use only six and seven edges. The dashed line shows a common reference prevalence. Sources: recompute_tier_metrics.py and verify_causalrivers_tiers.py. 5.3. Falsification and the role of non-causal baselines IAAFT … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 7 canonical work pages · 2 internal anchors

  1. [1]

    Learninghigh-dimensionaldirectedacyclicgraphswithmixeddata-types,in:ProceedingsofMachine Learning Research, PMLR

    Andrews,B.,Ramsey,J.,Cooper,G.F.,2019. Learninghigh-dimensionaldirectedacyclicgraphswithmixeddata-types,in:ProceedingsofMachine Learning Research, PMLR. pp. 4–21. Arpit, D., Fernandez, M., Feigenbaum, I., Yao, W., Liu, C., Yang, W., Josel, P., Heinecke, S., Hu, E., Wang, H., Hoi, S., Xiong, C., Zhang, K., Niebles, J.C.,

    2019

  2. [7]

    Datasets and Benchmarks Track, pp. 1–45. URL:https://openreview.net/forum?id=ck7euvU7V3. neurIPS 2025 Datasets and Benchmarks Track Poster. Hyvärinen, A., Zhang, K., Shimizu, S., Hoyer, P.O.,

    2025

  3. [8]

    Journal of Machine Learning Research 24, 1–8

    Python package for causal discovery based on LiNGAM. Journal of Machine Learning Research 24, 1–8. URL:https://jmlr.org/papers/v24/21-0321.html. Kraskov,A.,Stögbauer,H.,Grassberger,P.,2004. Estimatingmutualinformation. Phys.Rev.E69,066138. doi:10.1103/PhysRevE.69.066138. Krich, C., Runge, J., Miralles, D.G., Migliavacca, M., Perez-Priego, O., El-Madany, T...

    doi:10.1103/physreve.69.066138 2004

  4. [15]

    doi:10.48550/arXiv.2407.13054,arXiv:2407.13054

    Comprehensive review and empirical evaluation of causal discovery algorithms for numerical data. doi:10.48550/arXiv.2407.13054,arXiv:2407.13054. Nowack, P., Runge, J., Eyring, V., Haigh, J.D.,

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.13054

  5. [19]

    Econometrica 89, 955–980

    Local projections and VARs estimate the same impulse responses. Econometrica 89, 955–980. doi:10.3982/ECTA17813. Ramsey,J.B.,1969. Testsforspecificationerrorsinclassicallinearleast-squaresregressionanalysis. JournaloftheRoyalStatisticalSociety:Series B (Methodological) 31, 350–371. doi:10.1111/j.2517-6161.1969.tb00796.x. Ramsey, J.D., Zhang, K., Glymour, ...

    arXiv 1969

  6. [22]

    Nature Reviews Earth & Environment 4, 487–505

    Causal inference for time series. Nature Reviews Earth & Environment 4, 487–505. doi:10.1038/s43017-023-00431-y. Marco Ruiz et al.:Preprint submitted to ElsevierPage 32 of 33 AutoCause: Automating expert decisions in environmental time-series causal discovery Runge, J., Nowack, P., Kretschmer, M., Flaxman, S., Sejdinovic, D., 2019b. Detecting and quantify...

    doi:10.1038/s43017-023-00431-y

  7. [23]

    (Eds.), Proceedings of the NeurIPS 2019 Competition and Demonstration Track, PMLR

    The causality for climate competition, in: Escalante, H.J., Hadsell, R. (Eds.), Proceedings of the NeurIPS 2019 Competition and Demonstration Track, PMLR. pp. 110–120. URL:https://proceedings.mlr. press/v123/runge20a.html. Ryšavý, P., He, X., Mareček, J.,

    2019

  8. [24]

    Science 308, 523–529

    Causal protein-signaling networks derived from multiparameter single-cell data. Science 308, 523–529. doi:10.1126/science.1105809. Schreiber, T.,

    doi:10.1126/science.1105809

  9. [27]

    doi:10.1214/10-STS330

    To explain or to predict? Statistical Science 25, 289–310. doi:10.1214/10-STS330. Slote, K., Fish, J., Bollt, E.,

    doi:10.1214/10-sts330

  10. [28]

    URL:https://github.com/ Center-For-Complex-Systems-Science/causationentropy, doi:10.5281/zenodo.17047565

    CausationEntropy: A Python library for causal discovery. URL:https://github.com/ Center-For-Complex-Systems-Science/causationentropy, doi:10.5281/zenodo.17047565. Smith, S.M., Miller, K.L., Salimi-Khorshidi, G., Webster, M., Beckmann, C.F., Nichols, T.E., Ramsey, J.D., Woolrich, M.W.,

    work page doi:10.5281/zenodo.17047565

  11. [30]

    2nd ed., MIT Press, Cambridge, MA

    Causation, Prediction, and Search. 2nd ed., MIT Press, Cambridge, MA. doi:10.7551/mitpress/ 1754.001.0001. Stein, G., Shadaydeh, M., Blunk, J., Penzel, N., Denzler, J.,

    doi:10.7551/mitpress/

  12. [31]

    CausalRivers - scaling up benchmarking of causal discovery for real-world time-series, in: The Thirteenth International Conference on Learning Representations, pp. 1–16. URL:https://openreview.net/forum? id=wmV4cIbgl6. iCLR 2025 Spotlight. Székely, G.J., Rizzo, M.L., Bakirov, N.K.,

    2025

  13. [35]

    Coarse-to-fine learning of dynamic causal structures, in: The Fourteenth International Conference on Learning Representations, pp. 1–27. URL:https://openreview.net/forum?id=ooqnLFagKq. iCLR 2026 Poster. Yu, G., Guo, C., Luk, W., 2026a. Robust time series causal discovery for agent-based model validation. doi:10.48550/arXiv.2410.19412, arXiv:2410.19412. Yu...

    work page doi:10.48550/arxiv.2410.19412 2026

  14. [36]

    LOCAL: Learning with Orientation Matrix to Infer Causal Structure from Time Series Data

    LOCAL: Learning with orientation matrix to infer causal structure from time series data. doi:10.48550/arXiv.2410.19464,arXiv:2410.19464. Zhang, K., Zhu, S., Kalander, M., Ng, I., Ye, J., Chen, Z., Pan, L.,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.19464

  15. [37]

    doi:10.48550/arXiv

    gCastle: A Python toolbox for causal discovery. doi:10.48550/arXiv. 2111.15155,arXiv:2111.15155. Zheng, X., Aragam, B., Ravikumar, P., Xing, E.P.,

    Pith/arXiv arXiv doi:10.48550/arxiv.2111.15155

  16. [38]

    DAGs with NO TEARS: Continuous optimization for structure learning, in: Proceedings of the 32nd International Conference on Neural Information Processing Systems, Curran Associates Inc.. p. 9492–9503. Zheng,Y.,Huang,B.,Chen,W.,Ramsey,J.,Gong,M.,Cai,R.,Shimizu,S.,Spirtes,P.,Zhang,K.,2024. Causal-learn:CausaldiscoveryinPython. Journal of Machine Learning Re...

    2024

  17. [127]

    Marbach, D., Prill, R.J., Schaffter, T., Mattiussi, C., Floreano, D., Stolovitzky, G.,

    doi:10.1016/j.compenvurbsys.2026.102435. Marbach, D., Prill, R.J., Schaffter, T., Mattiussi, C., Floreano, D., Stolovitzky, G.,

    arXiv 2026

  18. [524]

    Marco Ruiz et al.:Preprint submitted to ElsevierPage 31 of 33 AutoCause: Automating expert decisions in environmental time-series causal discovery Granger,C.W.J.,1969

    doi:10.3389/fgene.2019.00524. Marco Ruiz et al.:Preprint submitted to ElsevierPage 31 of 33 AutoCause: Automating expert decisions in environmental time-series causal discovery Granger,C.W.J.,1969. Investigatingcausalrelationsbyeconometricmodelsandcross-spectralmethods. Econometrica:JournaloftheEconometric Society 37, 424–438. doi:10.2307/1912791. Günther...

    arXiv 2019

  19. [1415]

    Pamfil, R., Sriwattanaworachai, N., Desai, S., Pilgerstorfer, P., Georgatzis, K., Beaumont, P., Aragam, B.,

    doi:10.1038/s41467-020-15195-y. Pamfil, R., Sriwattanaworachai, N., Desai, S., Pilgerstorfer, P., Georgatzis, K., Beaumont, P., Aragam, B.,

    doi:10.1038/s41467-020-15195-y

  20. [1995]

    Journal of Econometrics 66, 225–250

    Statistical inference in vector autoregressions with possibly integrated processes. Journal of Econometrics 66, 225–250. doi:10.1016/0304-4076(94)01616-8. Yang, D., Tan, Q., Domeniconi, C., Wang, J., Cui, L., Yu, G.,

    work page doi:10.1016/0304-4076(94)01616-8

  21. [2000]

    Physical Review Letters 85, 461–464

    Measuring information transfer. Physical Review Letters 85, 461–464. doi:10.1103/PhysRevLett.85.461. Schreiber,T.,Schmitz,A.,2000.Surrogatetimeseries.PhysicaD:NonlinearPhenomena142,346–382.doi:10.1016/S0167-2789(00)00043-9. Shimizu, S., Hoyer, P.O., Hyvärinen, A., Kerminen, A.,

    doi:10.1103/physrevlett.85.461 2000

  22. [2005]

    Springer, Berlin, Heidelberg

    New Introduction to Multiple Time Series Analysis. Springer, Berlin, Heidelberg. doi:10.1007/3-540-27752-8. Lyu, W., Lei, Y., Yi, W., Song, Y., Li, X., Dai, S., Qin, Y., Zhao, W.,

    work page doi:10.1007/3-540-27752-8

  23. [2006]

    Journal of Machine Learning Research 7, 2003–2030

    A linear non-Gaussian acyclic model for causal discovery. Journal of Machine Learning Research 7, 2003–2030. URL:http://jmlr.org/papers/v7/shimizu06a.html. Shmueli, G.,

    2003

  24. [2007]

    The Annals of Statistics 35, 2769–2794

    Measuring and testing dependence by correlation of distances. The Annals of Statistics 35, 2769–2794. doi:10.1214/009053607000000505. Thumm, D., Chen, Y.,

    doi:10.1214/009053607000000505

  25. [2009]

    2nd ed., Cambridge University Press

    Causality: Models, Reasoning, and Inference. 2nd ed., Cambridge University Press. doi:10.1017/CBO9780511803161. Peters, J., Janzing, D., Schölkopf, B.,

    doi:10.1017/cbo9780511803161

  26. [2010]

    Proceedings of the National Academy of Sciences 107, 6286–6291

    Revealing strengths and weaknesses of methods for gene network inference. Proceedings of the National Academy of Sciences 107, 6286–6291. doi:10.1073/pnas.0913357107. McKinsey & Company,

    doi:10.1073/pnas.0913357107

  27. [2011]

    NeuroImage 54, 875–891

    Network modelling methods for FMRI. NeuroImage 54, 875–891. doi:10.1016/j.neuroimage.2010.08.063. Spirtes, P., Glymour, C., Scheines, R.,

    doi:10.1016/j.neuroimage.2010.08.063 2010

  28. [2015]

    5th ed., John Wiley & Sons

    Time Series Analysis: Forecasting and Control. 5th ed., John Wiley & Sons. Brouillard,P.,Squires,C.,Wahl,J.,Körding,K.,Sachs,K.,Drouin,A.,Sridhar,D.,2025. Thelandscapeofcausaldiscoverydata:Groundingcausal discoveryinreal-worldapplications,in:Huang,B.,Drton,M.(Eds.),ProceedingsoftheFourthConferenceonCausalLearningandReasoning, PMLR. pp. 834–873. URL:https:...

    2025

  29. [2017]

    The MIT Press

    Elements of Causal Inference: Foundations and Learning Algorithms. The MIT Press. URL: https://lccn.loc.gov/2017020087. Petersen, A.H.,

    arXiv

  30. [2018]

    Tetrad— a toolbox for causal discovery, in: 8th International Workshop on Climate Informatics, National Center for Atmospheric Research, Boulder, Colorado. pp. 1–4. URL:https://www.atmos.colostate.edu/~iebert/PAPERS/CI2018_paper_35.pdf. Ruiz,M.,Arana-Catania,M.,Ardila,D.R.,Ventura,R.,2026. Causal-Audit:Aframeworkforriskassessmentofassumptionviolationsinti...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.02488 2026

  31. [2020]

    Biogeosciences 17, 1033–1061

    Estimating causal networks in biosphere-atmosphere interaction with the PCMCI approach. Biogeosciences 17, 1033–1061. doi:10.5194/bg-17-1033-2020. Lütkepohl, H.,

    doi:10.5194/bg-17-1033-2020 2020

  32. [2021]

    Knowledge and Information Systems 63, 3041–3085

    Causal inference for time series analysis: Problems, methods and evaluation. Knowledge and Information Systems 63, 3041–3085. doi:10.1007/s10115-021-01621-0. Niu, W., Gao, Z., Song, L., Li, L.,

    work page doi:10.1007/s10115-021-01621-0

  33. [2022]

    CIPCaD-Bench: Continuous industrial process datasets for benchmarking causal discovery methods, in: 2022 IEEE 18th International Conference on Automation Science and Engineering (CASE), IEEE Press. pp. 2124–2131. doi:10.1109/CASE49997.2022.9926420. Mogensen, S.W., Rathsman, K., Nilsson, P.,

    arXiv 2022

  34. [2023]

    doi:10.48550/arXiv.2301.10859,arXiv:2301.10859

    Salesforce CausalAI library: A fast and scalable framework for causal analysis of time series and tabular data. doi:10.48550/arXiv.2301.10859,arXiv:2301.10859. Assaad,C.K.,Devijver,E.,Gaussier,E.,2022. Surveyandevaluationofcausaldiscoverymethodsfortimeseries. JournalofArtificialIntelligence Research 73, 767–819. doi:10.1613/jair.1.13428. Benjamini,Y.,Hoch...

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.10859 2022

  35. [2024]

    (Eds.), Proceedings of the Third Conference on Causal Learning and Reasoning, PMLR

    Bootstrap aggregation and confidence measures to improve time series causal discovery, in: Locatello, F., Didelez, V. (Eds.), Proceedings of the Third Conference on Causal Learning and Reasoning, PMLR. pp. 979–1007. URL: https://proceedings.mlr.press/v236/debeire24a.html. Delwiche,K.B.,Knox,S.H.,Malhotra,A.,Fluet-Chouinard,E.,McNicol,G.,Feron,S.,Ouyang,Z....

    2021

  36. [2025]

    TimeGraph: Synthetic benchmark datasets for robust time-series causal discovery, in: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2, Association for Computing Machinery, New York, NY, USA. pp. 5425–5435. doi:10.1145/3711896.3737439. Gerhardus,A.,Runge,J.,2020. High-recallcausaldiscoveryforautocorrelatedtimeseriesw...

    arXiv 2020

  37. [2026]

    Interventional time series priors for causal foundation models, in: 1st ICLR Workshop on Time Series in the Age of Large Models, pp. 1–15. URL:https://openreview.net/forum?id=JbTgx2L9Z2. iCLR 2026 TSALM Workshop Poster. Toda, H.Y., Yamamoto, T.,

    2026

  38. [2553]

    Runge, J., Gerhardus, A., Varando, G., Eyring, V., Camps-Valls, G.,

    doi:10.1038/s41467-019-10105-3. Runge, J., Gerhardus, A., Varando, G., Eyring, V., Camps-Valls, G.,

    doi:10.1038/s41467-019-10105-3