{"id":"dc31d28a-cbdc-4c03-bbe7-15997c6fddc5","arxiv_id":"2411.17542","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A knowledge-graph-driven pipeline extracts instrumental variables from causal chains in financial text and reports significant 2SLS estimates for two causal relationships.","lead":"This paper proposes ETE-FinCa, a pipeline that mines instrumental variables from a financial causal knowledge graph and tests them with two-stage least squares regression. It reports that graph-based features improve classification accuracy by up to 11.7%, and that two graph-derived instruments produce statistically significant causal estimates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The exclusion restriction is asserted from within-3-hop graph separation, which is neither necessary nor sufficient for IV validity, so the significant 2SLS coefficients do not by themselves establish the causal claim.","rationale":"The reader's weakest_assumption correctly identifies the load-bearing premise: that within-3-hop separation in FinCaKG establishes the IV exclusion restriction. This is exactly where the causal identification claim fails. The paper reports statistical significance of 2SLS coefficients, but statistical significance in an exactly identified 2SLS model cannot separate causal identification from specification error; it only reflects the estimated second-stage relationship under the maintained exclusion restriction. The Anderson and Cragg-Donald statistics test relevance, not exogeneity, so they do not repair the gap. The graph-based heuristic is also not derived from a formal structural model: FinCaKG edges are extracted causal spans, and the 3-hop cutoff is justified only by a citation to a prior paper's observation about causal chains, not by a theorem connecting graph distance to d-separation or conditional independence. Thus the central claim in the abstract and in Section IV.B.4 is not supported. The classification experiments are plausible, but they do not bear on causal identification. I agree with the reader's rejection, so the verdict should remain unchanged.","tokens_in":8446,"tokens_out":2309,"duration_ms":25317,"concrete_test":"Run a falsification regression on the same 14,099-observation sample used for Table IV: estimate B = α + βA + γZ + controls + ε for each of the two reported chains (economic exposure→EBITDA→governance and crude oil→operating profit→marketable securities). If γ is statistically significant after conditioning on A and the listed controls, the instrument has a direct or confounded association with B, and the exclusion restriction is refuted. Additionally, enumerate directed paths of length >3 from Z to B in FinCaKG; any such path would show that the 3-hop cutoff in Algorithm 1 is too narrow to establish exclusion at the graph level.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim depends on Algorithm 1 and Section III.C, where Z is declared a valid instrument if there is a 3-hop chain Z→A→B and no 3-hop chain Z→...→B. This treats 'no path of length ≤3' as equivalent to the IV exclusion restriction, but it is not. A directed path of length >3 from Z to B, a common ancestor of Z and B, or an unobserved confounder not represented in FinCaKG would all violate exclusion while passing Algorithm 1. Conversely, the presence of a 3-hop path does not by itself imply a structural causal effect in the econometric sense, because the graph edges are extracted from text and are not verified structural equations. The paper's empirical support in Table IV is also misaligned: the Anderson canonical LM statistic and Cragg-Donald Wald F statistic test relevance and weak-instrument strength, not exogeneity. The sentence in Section IV.B.4 claiming 'exogeneity (Cov(Z, ε2) = 0)' is therefore unsupported by any reported test. Since the 2SLS coefficients in Table IV are exactly identified, there is no internal overidentification test to validate the exclusion restriction. The classification improvements are a separate and more defensible contribution, but the headline causal-identification claim is not established by the evidence presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":8691,"tokens_out":4607,"duration_ms":49822,"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":[{"comment":"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.","section":"Section III.C, Algorithm 1"},{"comment":"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.","section":"Section IV.B.4, Table IV"},{"comment":"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.","section":"Section IV.B.3"},{"comment":"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.","section":"Section IV.B.4"}],"minor_comments":[{"comment":"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.","section":"Section III.C, Algorithm 1"},{"comment":"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.","section":"Table I"},{"comment":"The footnote contains the typo 'Attentation' instead of 'Attention'.","section":"Table I footnote"},{"comment":"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.","section":"Section IV.B.1"},{"comment":"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.","section":"Reference [24]"}],"recommendation":"reject","confidential_remarks":"The paper sits at the boundary of machine learning and empirical finance. The classification and graph-mining parts are potentially useful, but the causal-identification claim is not at the standard expected in an economics journal: the graph-separation criterion is not a valid substitute for the IV exclusion restriction, and the reported tests do not address exogeneity. The paper may be more suitable for a computer science or applied-ML venue if the causal claim is reframed as hypothesis generation. I would also flag the heavy reliance on FinCaKG, which comes from the same research group, as both the source of the mined IVs and the basis for their quality scoring."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a useful demonstration of how a causal knowledge graph can generate candidate instruments, but the headline causal claim is not established because the exclusion restriction is assumed from 3-hop graph separation rather than tested.\n\nWhat's actually new: wiring FinCaKG's causal graph to a DFS search for Z→A→B triples where Z is within 3 hops of A but not B, scoring them by hand-set thresholds, and running 2SLS on two chains. That pipeline is not in the cited literature, and the classification results in Table I are a real, if modest, gain: the weighted FinCaKG features beat the baseline by double digits in accuracy and F1. I'd believe that; it's a plausible use of a domain graph for feature selection.\n\nThe soft spot is the causal part. Algorithm 1 and Section III.C treat 'no path of length ≤3 from Z to B' as the exclusion restriction. That's not valid. A longer path, a backdoor path through a common cause, or an unobserved confounder would all violate exclusion while passing the algorithm. And the graph edges come from text, not from verified structural equations, so even a 3-hop path doesn't guarantee a causal effect. The empirical section then makes it worse: Anderson canonical LM and Cragg-Donald F test relevance and weak-instrument strength, not exogeneity. The sentence in IV.B.4 claiming Cov(Z, ε2)=0 is unsupported by any test, and with exactly identified equations there's no overidentification check. So the two significant 2SLS coefficients are suggestive but not evidence for the causal claim as stated.\n\nMinor issues: the quality thresholds (3 points, weights ≥5.0) are arbitrary; the two case-study chains are selected post hoc from the same graph that generated them; no code or data are shipped; and FinCaKG is from the same group, with no external validation. These would matter for a revision.\n\nThe classification part stands on its own, and the IV-mining pipeline could be a useful heuristic for generating candidates that a human then checks with proper econometric tests. As it stands, the causal claim is overclaimed.\n\nWho's it for: researchers working on text-derived causal graphs and automated IV discovery, and applied people who want a structured way to brainstorm instruments. A serious referee should see it, because the pipeline is novel and the flaws are fixable; the authors need to add a real exogeneity argument or reframe the output as candidate generation. I'd say send to peer review with the expectation of major revision.","headline":"The classification half is solid, but the headline causal claim rides on an untested graph-separation assumption that does not establish the exclusion restriction.","tokens_in":9224,"tokens_out":1989,"would_cite":false,"duration_ms":67083,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["causal inference","instrumental variables","causal knowledge graph","finance","two-stage least squares","interpretability","feature selection","text mining"],"falsifier":"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.","tokens_in":8236,"feed_emoji":"📈","tokens_out":6217,"duration_ms":51575,"temperature":0.7,"pith_summary":"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.","feed_headline":"Two text-mined financial instruments pass 2SLS significance tests","feed_subtitle":"ETE-FinCa turns a financial causal knowledge graph into instruments; two chains survive 2SLS with p<0.01.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Constructs the FinCaKG financial causal knowledge graph that ETE-FinCa mines for IV candidates.","marker":"[10]"},{"why":"Supplies the assumption that causal chains lose meaning after three hops, justifying the distance rule.","marker":"[24]"},{"why":"Provides the structural causal model and exclusion-restriction framework that IV reasoning relies on.","marker":"[4]"},{"why":"Establishes the potential-outcome framework that motivates using instruments for causal effects.","marker":"[2]"},{"why":"Represents an automatic IV-generation baseline whose lack of interpretability the paper contrasts with its own approach.","marker":"[7]"}],"fun_headline_variants":["Two financial instruments pass 2SLS significance","Knowledge graph yields two significant instruments","Two causal instruments from FinCaKG pass 2SLS","ETE-FinCa identifies two significant causal instruments","Two instruments from FinCaKG survive 2SLS with p<0.01"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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 $A$.","fun_headline_variants_meta":{"raw":{"variants":["Two financial instruments pass 2SLS significance","Knowledge graph yields two significant instruments","Two causal instruments from FinCaKG pass 2SLS","ETE-FinCa identifies two significant causal instruments","Two instruments from FinCaKG survive 2SLS with p<0.01"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00121,"raw_usage":{"total_tokens":4966,"prompt_tokens":913,"completion_tokens":4053,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":3974}},"tokens_in":529,"tokens_out":4053,"duration_ms":25295,"temperature":1.0,"reasoning_tokens":3974,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:59:39.853556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"FinKG: A core financial knowledge graph for financial analysis,","cited_arxiv_id":null,"evidence_quote":"Constructs the FinCaKG financial causal knowledge graph that ETE-FinCa mines for IV candidates."},{"cited_title":"Exploring causal chain identification: Compre- hensive insights from text and knowledge graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the assumption that causal chains lose meaning after three hops, justifying the distance rule."},{"cited_title":"Causality,","cited_arxiv_id":null,"evidence_quote":"Provides the structural causal model and exclusion-restriction framework that IV reasoning relies on."},{"cited_title":"Estimating causal effects of treatments in randomized and nonrandomized studies,","cited_arxiv_id":null,"evidence_quote":"Establishes the potential-outcome framework that motivates using instruments for causal effects."},{"cited_title":"Auto IV: Counterfactual prediction via automatic instru- mental variable decomposition,","cited_arxiv_id":null,"evidence_quote":"Represents an automatic IV-generation baseline whose lack of interpretability the paper contrasts with its own approach."}],"review_version":1}