Pith. sign in

REVIEW 4 major objections 5 minor 24 references

Causal Inference in Finance: An Expertise-Driven Model for Instrument Variables Identification and Interpretation

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a financial causal knowledge graph built from expert texts can supply valid instrumental variables: two text-mined chains pass two-stage least squares with significance.

desk verdict The classification half is solid, but the headline causal claim rides on an untested graph-separation assumption that does not establish the exclusion restriction. read the letter →

arxiv 2411.17542 v1 pith:D7UO2TNP submitted 2024-11-26 econ.GN q-fin.EC

Ying Chen , Ziwei Xu , Kotaro Inoue , Ryutaro Ichise This is my paper · ORCID
classification econ.GNq-fin.EC
keywords causalinferenceinstrumentalvariablesknowledgegraphfinancetwo-stageleastsquaresinterpretabilityfeatureselectiontextmining
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

This paper seeks to establish that expert textual knowledge, organized as a causal knowledge graph, can generate valid instrumental variables for causal inference in finance when hand-crafted instruments are scarce. The proposed model ETE-FinCa extracts cause-effect chains from the financial knowledge graph FinCaKG, scores them as instruments, and tests whether standpoint-specific causal claims hold in broad company data. The authors report that two such chains—economic exposure to EBITDA to corporate governance, and oil to operating profit to securities—yield statistically significant 2SLS coefficients, with the first stage satisfying relevance and exclusion diagnostics. If the approach is sound, it offers a general route from text to testable causal claims in economics.

What carries the argument

The mechanism is the depth-first search over a financial causal knowledge graph, together with a three-hop connectivity rule: a node $Z$ qualifies as an instrument for the pair $A \to B$ when $R(Z,A)$, $R(A,B)$, and $\neg R(Z,B)$ all hold within three hops. This rule operationalizes the IV conditions—relevance ($Z$ and $A$ are associated) and exclusion ($Z$ affects $B$ only through $A$)—by treating graph distance as a proxy for causal association. Quality scores (edge node, edge weight $\geq 5$) filter the mined triples, and 2SLS with industry and year fixed effects then tests whether the implied causal paths hold in real data.

What would settle it

Run a regression of the outcome $B$ on the instrument $Z$ and the treatment $A$ in the same sample: if the coefficient on $Z$ is statistically significant while controlling for $A$, the exclusion restriction is violated and the claimed instrumental validity would collapse; alternatively, a Hansen J test on an overidentified specification would provide a direct statistical check.

Watch

Extended reading notes

Core claim

The central discovery is that the three-hop distance rule in FinCaKG can act as a proxy for the instrumental-variable exclusion restriction. Algorithm 1 searches for triples ($Z$, $A$, $B$) such that $Z$ reaches $A$ within three hops, $A$ reaches $B$ within three hops, and $Z$ does not reach $B$ within three hops; these triples are then scored as instruments. In the empirical test, two standpoint-specific chains survive 2SLS: the first uses foreign-exchange exposure as instrument for the effect of EBITDA on corporate governance (coefficient 26.1, $p<0.01$), and the second uses crude-oil price exposure as instrument for the effect of operating profit on marketable securities (coefficient 0.34, $p<0.01$), with Anderson LM and Cragg-Donald statistics supporting relevance. The paper concludes that these high-quality instruments identify causal relationships with statistical significance.

Load-bearing premise

The entire causal interpretation rests on the assumption that having no path from $Z$ to $B$ within three hops in the knowledge graph guarantees that $Z$ affects $B$ only through $A$.

Editorial extensions

If this is right

  • Features derived from the causal knowledge graph outperform a cosine-similarity baseline by up to 11.7% in accuracy and 23.0% in F1-score for the company-document classification task.
  • The approach mines 19,678 instrument-consistent triples from the full graph, with 870 instruments exclusive to the stakeholder subgraph and 546 exclusive to the shareholder subgraph.
  • Two standpoint-specific causal chains, validated with 2SLS, show significant coefficients (26.1 and 0.34, both $p<0.01$) and pass first-stage relevance diagnostics.
  • The method suggests that expert texts, when structured as a causal graph, can supply interpretable instruments where synthetic or hand-crafted instruments are unavailable.
  • The significant results on the full sample imply that these expertise-driven causal relationships generalize beyond the subgraph standpoints.

Reading between the lines

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

  • The paper's exogeneity claim rests on the graph's completeness; if unobserved confounders lie outside the three-hop horizon, the instrument could still affect the outcome through an unseen path, and the reported diagnostics do not test that.
  • A stronger test of the exclusion restriction, such as a Hansen J overidentification test or a direct regression of the outcome on the instrument controlling for the treatment, would be a natural next step.
  • The same text-to-instrument pipeline could be ported to other expert-heavy domains, such as healthcare or public policy, wherever a causal knowledge graph can be assembled from written expertise.
  • The 3-hop cutoff is an assumption borrowed from prior work; an editor's sensitivity analysis would check whether results change materially with a 2-hop or 4-hop window.
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

4 major / 5 minor

Summary. The paper proposes ETE-FinCa, an expertise-driven pipeline for discovering instrumental variables (IVs) from a financial causal knowledge graph (FinCaKG). The method uses a DFS algorithm to find triples (Z, A, B) in which Z is within three hops of A and A is within three hops of B, while Z is not within three hops of B; it then scores these triples as high-, middle-, or low-quality based on graph edge properties. The authors evaluate the graph-derived features on a shareholder/stakeholder classification task and report accuracy and F1 improvements. They further run 2SLS regressions for two selected chains, 'economic exposure → EBITDA → governance' and 'oil → profit → securities', and report statistically significant coefficients, concluding that high-quality IVs identify causal relationships.

Significance. The classification experiments are a credible and well-reported feature-selection contribution: the authors compare multiple corpora, feature styles, and classifiers, and the improvements over the cosine-similarity baseline are substantial. The DFS algorithm is explicit and the graph-mining results are described in useful detail. However, the paper's headline claim—that the mined IVs identify causal effects—rests on a graph-separation criterion that is not a valid substitute for the IV exclusion restriction. The reported 2SLS diagnostics test relevance and weak instruments, not exogeneity, and the selected chains are a tiny post-hoc subset of thousands of candidates. If the causal-identification claim were removed or reframed as hypothesis generation, the classification and graph-mining results could be a useful applied contribution; as it stands, the central claim is not established.

major comments (4)
  1. [Section III.C, Algorithm 1] The paper treats the condition that Z is within three hops of A but not within three hops of B as establishing the IV exclusion restriction. This is neither necessary nor sufficient. A directed path of length greater than three from Z to B, a common ancestor of Z and B, or an unobserved confounder not represented in FinCaKG would violate Cov(Z, ε2)=0 while still passing Algorithm 1. Conversely, the existence of a three-hop chain in a text-mined knowledge graph does not by itself establish a structural causal relation, because the graph edges are extracted from text and are not estimated structural equations. Additionally, the stated condition 'A ⊥ B | Z' in Section III.C is not the IV condition; a valid instrument graph satisfies Z ⊥ B | A (or, properly, exclusion of Z from the outcome equation), not independence between treatment and outcome given the instrument. The central claim that the identified triples are valid instruments is therefore unsupported.
  2. [Section IV.B.4, Table IV] The only diagnostic statistics reported for the 2SLS estimates are the Anderson canonical correlation LM statistic and the Cragg-Donald Wald F statistic. These test relevance and weak-instrument strength, respectively; neither tests exogeneity. The sentence in Section IV.B.4 claiming that these statistics support Cov(Z, ε2)=0 is therefore incorrect. Moreover, each specification is exactly identified, so there is no overidentifying restriction available to probe the exclusion assumption. The significant coefficients in Table IV cannot by themselves establish that the exclusion restriction holds.
  3. [Section IV.B.3] The three scoring rules for 'high-quality' IVs—being an edge node, wz,a ≥ 5.0, and wa,b ≥ 5.0—are arbitrary and are not tied to any econometric condition for IV validity. Edge-node status and edge-weight thresholds do not speak to exclusion or relevance. In addition, the same FinCaKG graph is used both to mine the triples and to score their quality, and the graph is produced by the same research group as the present paper; there is no external validation of the causal content of its edges. The label 'high-quality' is therefore not evidence that the identified instruments are econometrically valid.
  4. [Section IV.B.4] The empirical study reports only two of the many high-quality chains identified in Table III (95 in SH-FinCaKG and 21 in ST-FinCaKG), and these two appear to be selected after inspecting the graph-based results. With 19,678 candidate triples overall, presenting only the two chains that yield statistically significant 2SLS coefficients, without multiple-testing correction or a pre-specified selection rule, makes the p<0.01 results difficult to interpret. The paper should either report results for all high-quality chains or provide a principled, pre-specified reason for selecting these two.
minor comments (5)
  1. [Section III.C, Algorithm 1] The pseudocode removes b from Btemp and then unconditionally adds (z, a, b) to the result set K; please clarify whether the triple should be added before the removal or only when b is not removed.
  2. [Table I] The table is very hard to read because the column headers and row groupings are ambiguous; for example, the 'Num. of Feature' column appears to contain three different values in one row. Please reformat the table so each configuration is a separate row.
  3. [Table I footnote] The footnote contains the typo 'Attentation' instead of 'Attention'.
  4. [Section IV.B.1] The text refers to 'row 1 in CF_inCaKG (weighted)' when comparing to the baseline, but Table I has multiple blocks; please specify the exact row or add row labels to avoid ambiguity.
  5. [Reference [24]] The load-bearing assumption that causal chains lose meaning after three hops is justified only by an 'in press' reference; please provide a citable published version or a fuller explanation of this assumption.

Circularity Check

1 steps flagged · score 4.0 of 10

The 2SLS coefficients are not definitionally forced, but the causal interpretation rests on a load-bearing 3-hop cutoff imported from the authors' own prior work rather than on an exogeneity test.

  1. ansatz smuggled in via citation [Section III.C (Interpretation Module), paragraph defining 3-hop associations; implemented in Algorithm 1]
    "Most causal chains lose causal meaning after three hops [24]; thus, we define associations within 3-hop as effective logical connections."

    The 3-hop cutoff is the criterion that converts graph reachability into 'effective logical connections,' which Algorithm 1 then treats as the IV pattern 'A ⊥ B | Z' and Section IV.B.4 treats as exogeneity (Cov(Z, ε2) = 0). The only support offered for this cutoff is reference [24], an in-press paper by two of the present co-authors (Xu and Ichise). No independent validation is given, so the causal-identification claim reduces to a self-cited ansatz. The 2SLS coefficients are still estimated from independent Bloomberg data and could have been insignificant, so the circularity is partial rather than definitional.

full rationale

The classification benchmark is self-contained: feature sets built from Csim and FinCaKG are compared on validation data with fixed baselines, and the reported accuracy/F1 improvements are not artifacts of construction. The two 2SLS case studies use external market and financial data, so the reported coefficients (26.1 and 0.34) are not forced by the graph weights. However, the paper's headline causal claim depends on the validity of Z as an instrument, and that validity is asserted from a 3-hop graph-separation rule whose only cited justification is the same authors' prior work [24]. The statistical tests reported (Anderson canonical LM and Cragg-Donald F) speak only to relevance and weak instruments, not to the exclusion restriction; the sentence claiming these statistics support exogeneity is unsupported. This is a correctness/identification gap, and the self-cited 3-hop rule is load-bearing for it. Because the empirical estimates themselves are independent and not definitionally equal to the graph inputs, the overall circularity is moderate (4), not extreme.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The pipeline rests on the correctness of the self-cited FinCaKG graph, on arbitrary thresholds, and on the assumption that graph separation equals instrument validity. These are not independently verified, and they directly determine which IVs are mined and which causal effects are reported.

free parameters (3)
  • 3-hop distance threshold = 3
    Defines which chains are causally meaningful and which nodes count as unrelated; cited to [24] but not fitted or validated.
  • Csim similarity threshold = 0.55
    Hand-set threshold for including vocabulary in the baseline corpus; no tuning or sensitivity analysis is reported.
  • Edge weight thresholds for IV quality = 5.0
    Hand-set cutoffs for the Z-A and A-B edge weights in the quality score; no justification or robustness check is provided.
assumptions (5)
  • domain assumption FinCaKG edges represent true causal relationships
    The entire IV search and 2SLS interpretation assume the graph's cause-effect edges are correct and complete; the graph is from self-cited prior work [10] and is not externally validated in this paper.
  • domain assumption Within three hops, causal meaning is preserved; beyond three it is not
    Invoked to define R(x,y) in Algorithm 1, citing [24]; no independent evidence is provided.
  • domain assumption Z not associated with B within three hops implies Z is excludable
    Used to turn graph separation into the IV exclusion restriction; graph non-adjacency is not conditional independence.
  • domain assumption Annual report self-description as shareholder or stakeholder maximizing is a valid label
    Labels for classification are constructed from company claims; no manual validation or inter-annotator agreement is reported.
  • domain assumption 2SLS standard assumptions hold for the selected chains
    Relevance is tested via Anderson and CD statistics, but exogeneity is assumed from the graph; weak instrument concerns remain because the CD statistics are barely above 10.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Inference in Finance: An Expertise-Driven Model for Instrument Variables Identification and Interpretation." pith.science (2026). https://pith.science/paper/D7UO2TNP

@misc{pith2026241117542,
  author       = {Pith},
  title        = {Pith review of: Causal Inference in Finance: An Expertise-Driven Model for Instrument Variables Identification and Interpretation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D7UO2TNP}},
  note         = {Machine review of arXiv:2411.17542}
}
read the original abstract

Instrumental Variable (IV) provides a source of treatment randomization that is conditionally independent of the outcomes, responding to the challenges of counterfactual and confounding biases. In finance, IV construction typically relies on pre-designed synthetic IVs, with effectiveness measured by specific algorithms. This classic paradigm cannot be generalized to address broader issues that require more and specific IVs. Therefore, we propose an expertise-driven model (ETE-FinCa) to optimize the source of expertise, instantiate IVs by the expertise concept, and interpret the cause-effect relationship by integrating concept with real economic data. The results show that the feature selection based on causal knowledge graphs improves the classification performance than others, with up to a 11.7% increase in accuracy and a 23.0% increase in F1-score. Furthermore, the high-quality IVs we defined can identify causal relationships between the treatment and outcome variables in the Two-Stage Least Squares Regression model with statistical significance.

Figures

Figures reproduced from arXiv: 2411.17542 by the authors.

Figure 1
Figure 1. The causal graphs with instrument variable (Z) specification. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The workflow of ETE-FinCa model for causal vairables identification and interpretation. Algorithm 1: IV identification based on DFS Input: Graph G, all NPs unique (V), tuple nodes sameid Output: Set of (Z, A, B) triples K Define R(x, y): the distance between x and y in Graph is less than or equal to 3 hops; Initialize empty result set K; for each z in V do Az = {a ∈ V |R(z, a)}; for each a in Az do Btemp = {b ∈ V |R… view at source ↗
Figure 3
Figure 3. An instance of causal variables identification in DFS algorithm. The [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 23 canonical work pages

  1. [1]

    Principal stratification in causal inference,

    C. E. Frangakis and D. B. Rubin, “Principal stratification in causal inference,” Biometrics, vol. 58, no. 1, pp. 21–29, 2002

    work page 2002

  2. [2]

    Estimating causal effects of treatments in randomized and nonrandomized studies,

    D. B. Rubin, “Estimating causal effects of treatments in randomized and nonrandomized studies,” Journal of Educational Psychology , vol. 66, no. 5, pp. 688–701, 1974

    work page 1974

  3. [3]

    Wikidata: a free collaborative knowl- edgebase,

    D. Vrande ˇci´c and M. Kr ¨otzsch, “Wikidata: a free collaborative knowl- edgebase,” Communications of the ACM , vol. 57, no. 10, pp. 78–85, 2014

    2014

  4. [4]

    Causality,

    J. Pearl, “Causality,” Cambridge: Cambridge University Press , 2009

    work page 2009

  5. [5]

    Estimation of regression coefficients when some regressors are not always observed,

    J. M. Robins, A. Rotnitzky, and L. P. Zhao, “Estimation of regression coefficients when some regressors are not always observed,” Journal of the American Statistical Association , vol. 89, no. 427, pp. 846–866, 1994

    work page 1994

  6. [6]

    Deep IV: A flexible approach for counterfactual prediction,

    J. Hartford, G. Lewis, K. Leyton-Brown, and M. Taddy, “Deep IV: A flexible approach for counterfactual prediction,” in Proc. 34th Interna- tional Conference on Machine Learning , 2017, pp. 1414–1423

    work page 2017

  7. [7]

    Auto IV: Counterfactual prediction via automatic instru- mental variable decomposition,

    J. Yuan et al., “Auto IV: Counterfactual prediction via automatic instru- mental variable decomposition,” ACM Trans. Knowl. Discov. Data , vol. 16, no. 4, pp. 1–20, Aug. 2022

    work page 2022

  8. [8]

    (COMET-) ATOMIC 2020: On symbolic and neural commonsense knowledge graphs,

    J. D. Hwang et al., “(COMET-) ATOMIC 2020: On symbolic and neural commonsense knowledge graphs,” in Proc. 35th AAAI Conference on Artificial Intelligence, 2021, vol. 35, no. 7, pp. 6384–6392

    work page 2020

Show all 24 references
  1. [9]

    Treatment effect estimation with data-driven variable decomposition,

    K. Kuang, P. Cui, B. Li, M. Jiang, S. Yang, and F. Wang, “Treatment effect estimation with data-driven variable decomposition,” in Proc. 31st AAAI Conference on Artificial Intelligence , 2017

    2017

  2. [10]

    FinKG: A core financial knowledge graph for financial analysis,

    N. Kertkeidkachorn, R. Nararatwong, Z. Xu, and R. Ichise, “FinKG: A core financial knowledge graph for financial analysis,” in Proc. IEEE 17th International Conference on Semantic Computing, 2023, pp. 90–93

    2023

  3. [11]

    Constructing a control group using multivariate matched sampling methods that incorporate the propensity score,

    P. R. Rosenbaum and D. B. Rubin, “Constructing a control group using multivariate matched sampling methods that incorporate the propensity score,” The American Statistician , vol. 39, no. 1, pp. 33–38, 1985

    1985

  4. [12]

    Learning counterfactual representations for estimating individual dose- response curves,

    P. Schwab, L. Linhardt, S. Bauer, J. M. Buhmann, and W. Karlen, “Learning counterfactual representations for estimating individual dose- response curves,” in Proc. 34th AAAI Conference on Artificial Intelli- gence, 2020, pp. 5612–5619

    2020

  5. [13]

    Kernel instrumental variable re- gression,

    R. Singh, M. Sahani, and A. Gretton, “Kernel instrumental variable re- gression,” in Proc. Advances in Neural Information Processing Systems , 2019

    2019

  6. [14]

    Conceptnet 5.5: An open multilingual graph of general knowledge,

    R. Speer, J. Chin, and C. Havasi, “Conceptnet 5.5: An open multilingual graph of general knowledge,” in Proc. 31st Conference on Artificial Intelligence, 2017, pp. 4444–4451

    2017

  7. [15]

    Recursive partitioning for heterogeneous causal effects,

    S. Athey and G. Imbens, “Recursive partitioning for heterogeneous causal effects,” in Proc. the National Academy of Sciences , vol. 113, no. 27, pp. 7353–7360, 2016

    2016

  8. [16]

    Analysis of representations for domain adaptation,

    S. Ben-David, J. Blitzer, K. Crammer, and F. Pereira, “Analysis of representations for domain adaptation,” in Proc. Advances in Neural Information Processing Systems , 2007, pp. 137–144

    2007

  9. [17]

    Metalearners for estimating heterogeneous treatment effects using machine learning,

    S. R. K ¨unzel, J. S. Sekhon, P. J. Bickel, and B. Yu, “Metalearners for estimating heterogeneous treatment effects using machine learning,” in Proc. the National Academy of Sciences , vol. 116, no. 10, pp. 4156– 4165, 2019

    2019

  10. [18]

    Causenet: Towards a causality graph extracted from the web,

    S. Heindorf, Y . Scholten, H. Wachsmuth, A.-C. Ngonga Ngomo, and M. Potthast, “Causenet: Towards a causality graph extracted from the web,” in Proc. 29th ACM International Conference on Information & Knowledge Management, 2020, pp. 3023–3030

    2020

  11. [19]

    Approximate residual balancing: debiased inference of average treatment effects in high dimensions,

    S. Athey, G. W. Imbens, and S. Wager, “Approximate residual balancing: debiased inference of average treatment effects in high dimensions,” Journal of the Royal Statistical Society: Series B (Statistical Methodol- ogy), vol. 80, no. 4, pp. 597–623, 2018

    2018

  12. [20]

    Estimating individual treatment effect: generalization bounds and algorithms,

    U. Shalit, F. D. Johansson, and D. Sontag, “Estimating individual treatment effect: generalization bounds and algorithms,” in Proc. 34th International Conference on Machine Learning , 2017, pp. 3076–3085

    2017

  13. [21]

    Doubly robust joint learning for recommendation on data missing not at random,

    X. Wang, R. Zhang, Y . Sun, and J. Qi, “Doubly robust joint learning for recommendation on data missing not at random,” in Proc. 36th International Conference on Machine Learning , 2019, pp. 6638–6647

    2019

  14. [22]

    Commonsense causal reasoning between short texts,

    Z. Luo, Y . Sha, K. Zhu, S.-W. Hwang, and Z. Wang, “Commonsense causal reasoning between short texts,” in Proc. 15th International Conference on Principles of Knowledge Representation and Reasoning , 2016

    2016

  15. [23]

    Guided generation of cause and effect,

    Z. Li, X. Ding, T. Liu, J. E. Hu, and B. Van Durme, “Guided generation of cause and effect,” in Proc. 29th International Joint Conference on Artificial Intelligence, 2020, pp. 3629–3636

    2020

  16. [24]

    Exploring causal chain identification: Compre- hensive insights from text and knowledge graphs,

    Z. Xu and R. Ichise, “Exploring causal chain identification: Compre- hensive insights from text and knowledge graphs,” 26th International Conference in Big Data Analytics and Knowledge Discovery, Springer , 2024, in press

    2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.