{"id":"f94c4ee4-a08d-4e0f-9bf4-ea37d6afabbf","arxiv_id":"2508.15667","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Applying DII to EUA returns, the authors find IBEX35 and coal futures are the most robust causal drivers, with some non-linear links that Granger causality misses.","lead":"This paper applies a non-parametric tool called Differentiable Information Imbalance (DII) to daily returns of EU carbon allowances, alongside standard linear Granger causality. It reports that IBEX35 and coal futures are the most consistent drivers of EUA returns, and argues DII catches non-linear relationships that the linear model misses.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (9) defines IG with an inverted ratio, so Eq. (8) and IG>0 are inconsistent; the reported positive IGs for true causes cannot follow from the stated formula.","rationale":"The reader identified causal sufficiency as the weakest assumption, and that is a valid limitation. However, the most load-bearing concern is a formal internal inconsistency in the definition of the Imbalance Gain, the exact measure used to produce the paper's central empirical claim. Equation (8) states that a causal link exists when adding a variable improves prediction, i.e., min_w DII(with x) < DII(without x). Since a lower DII means better predictive information, this inequality makes the ratio in Eq. (9) greater than 1, forcing IG<0. The paper's assertion that Eq. (8) is equivalent to IG>0 is therefore algebraically false. The same inverted ratio appears in the multivariate IG of Eq. (11). If the code actually implements the inverse ratio (which would yield positive IG for true causes), then the written equations are wrong but the reported findings may still hold; if the code implements the equations as written, the empirical ranking of IBEX35 and CoalFut as positive IG variables is called into question. Either way, the paper as submitted lacks a coherent definition of its central quantity, making the results irreproducible from the text alone. This warrants a conditional acceptance pending correction and verification of the IG formula.","tokens_in":19794,"tokens_out":8087,"duration_ms":88446,"concrete_test":"Using the provided repository, re-run the synthetic experiment of §3.3.1 and compute IG for x1 and x2 exactly as written in Eq. (9) and Eq. (11). If the true causes yield positive IG, the implementation must be using the inverse ratio; if they yield negative IG, the paper's claim that IG>0 indicates causation is false. An algebraic check: substituting Eq. (8) into Eq. (9) gives IG<0, so the equivalence stated in the text cannot hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central causal measure, the Imbalance Gain, is defined in Eq. (9) as IG = 1 - Δ(z0→zτ) / min_w Δ([w x0,z0]→zτ), and the causal condition in Eq. (8) is min_w Δ(...) < Δ(z0→zτ). Since DII decreases when predictive information is added (Δ≈0 means highly predictive), the fraction in Eq. (9) is >1 whenever Eq. (8) holds, so IG<0. The paper nevertheless states that Eq. (8) is equivalent to IG>0. The same inversion appears in the multivariate version, Eq. (11), where the ratio is DII(without xα) / DII(with xα). Consequently, the positive IGs reported in §4 (IBEX35, CoalFut) cannot be obtained from the stated formula if those variables are true causes; either the equations are mis-specified or the code used the inverse ratio. This is a load-bearing internal inconsistency because the entire empirical ranking and the headline overlap claim rest on IG. The public GitHub repository is referenced but not independently verified, so the actual computed quantity is ambiguous.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Differentiable Information Imbalance (DII) as a non-parametric, potentially non-linear method for causal discovery in financial time series, and applies it to EUA (European Union Allowance) returns over January 2013–April 2024. It compares DII-based Imbalance Gain (IG) rankings with multivariate Granger causality from a VAR(1) model, and reports that both methods identify IBEX35 and Coal Futures as important drivers of EUA returns. Two synthetic datasets are used to show that linear Granger causality can miss non-linear causal links and can produce false positives when a non-linear common driver is present.","tokens_in":20058,"tokens_out":3182,"duration_ms":42340,"significance":"If the empirical claims hold, the paper would contribute a non-linear, non-parametric complement to standard Granger-causal analysis of carbon markets, and the overlap with VAR results would strengthen confidence in IBEX35 and coal futures as robust predictors of EUA returns. The paper also provides reproducible code and clearly describes the DII optimization procedure. However, a sign inconsistency in the central IG definition currently undermines the reported positive IG values, so the empirical ranking and the headline overlap claim cannot be accepted without correction. The synthetic illustrations are suggestive but are presented as single-run demonstrations without uncertainty quantification. With a corrected definition and appropriate robustness checks, the contribution could be useful to the q-fin.CP community.","major_comments":[{"comment":"The Imbalance Gain is defined in Eq. (9) as IG = 1 − Δ(z0→zτ) / min_w Δ([w x0,z0]→zτ). Since smaller Δ means better predictability, the causal condition in Eq. (8) is min_w Δ([w x0,z0]→zτ) < Δ(z0→zτ). Under Eq. (9) this inequality makes the ratio greater than 1, so IG < 0, not IG > 0 as claimed. The multivariate version Eq. (11) has the same inversion: the numerator is the DII without xα and the denominator is the DII with xα, so a genuine causal variable yields a negative IG. Consequently, the positive IGs reported in Figures 4–6 for true causes cannot be produced by the formula as written. Either the equations are mis-specified or the code uses the inverse ratio. This is load-bearing because the entire empirical ranking and the paper's central overlap claim rest on IG. Please correct the definition and verify which quantity was actually computed.","section":"§3.2.2, Eqs. (8)–(11)"},{"comment":"The empirical IG values are presented as evidence of causal effects, but no significance threshold, confidence interval, or null distribution is provided. The authors acknowledge in §6 that DII lacks a simple significance test, yet the conclusion states that both methods 'detect significant causal effects' from IBEX35 and Coal Futures. Given that the DII optimization involves stochastic mini-batches and multiple hyperparameters, a bootstrap or permutation test is needed to distinguish non-zero IG from noise. Without such a test, the empirical overlap between DII and Granger rankings cannot be regarded as statistically supported.","section":"§4 and Appendix A"},{"comment":"The synthetic demonstrations are single-run experiments with no repeated seeds, error bars, or sensitivity analysis. The DII training uses random mini-batches and Adam optimization, so the reported weights and IGs could vary across runs. Claims such as 'the DII approach appears robust against this drawback' (Section 3.3.2) are not supported without repeated-seed statistics. Additionally, the key hyperparameters — the 0.1 prefactor in Eq. (14), the 5% neighbourhood fraction, the mini-batch size, the temporal exclusion window, and the number of epochs — are fixed without sensitivity analysis. At minimum, show stability of the synthetic conclusions across seeds and across reasonable hyperparameter perturbations.","section":"§3.3 and Appendix A"}],"minor_comments":[{"comment":"The text 'heavAC100 per tonne' appears to be a typo; presumably '€100 per tonne' is intended.","section":"Introduction, Section 1"},{"comment":"The columns in Table 2 appear misaligned or garbled: for example, the EUA row shows Max = -0.0143 and 25% = 0.0010, which are inconsistent with the other rows and with the text. Please reformat and verify all entries.","section":"Table 2"},{"comment":"The text says 'In the first panel, the IG highlights...' but the left panel of Figure 6 is labelled as F-statistics and the central panel as IG. The wording should refer to the correct panel.","section":"Section 4, Figure 6"},{"comment":"Reference [1] is listed as 'Phys. Rev. Lett., pages –, Jun 2025' with no volume/article number; please complete the citation. Reference [20] appears to be an unpublished manuscript; if it is not accepted or posted, please mark it clearly.","section":"References"},{"comment":"The neighbourhood parameter k is first described as 5% of the points and later as 5 points in each mini-batch of size 100. This is consistent (5% of 100), but the wording could be clarified to avoid confusion.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The DII methodology originates largely from papers co-authored by members of this group, and the synthetic benchmarks are constructed by the authors rather than an independent testbed. A neutral third-party implementation or comparison with an existing causal-discovery baseline would substantially increase confidence. The sign inversion in the IG definition is a serious technical issue, but it appears fixable; I would not reject solely on that basis if the authors can confirm the computed quantity and correct the equations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new part is the application: taking the published DII/II framework to EUA returns on a 2013–2024 dataset and lining it up against multivariate Granger F-statistics. The synthetic examples (VAR false negatives and false positives) are clean and make a legitimate pedagogical point. The paper also deserves credit for being explicit that causal sufficiency is likely violated and that IG has no built-in significance test; the authors do not claim more statistical machinery than they build.\n\nNow the problem. Eq. (8) defines a causal improvement as min_w Δ([w x0,z0]→zτ) < Δ(z0→zτ). Because Δ≈0 means high predictive power, that condition is exactly \"adding x lowers the imbalance.\" But Eq. (9) sets IG = 1 − Δ(baseline)/min_combined. Under the causal condition the ratio is >1, so IG is negative, not positive. The multivariate version Eq. (11) has the same inversion: IG = 1 − min_without/min_with, again negative when the variable helps. The paper states Eq. (8) is equivalent to IG>0, and reports positive IGs for IBEX35 and CoalFut. Those positives cannot come from the stated formula if those variables are true predictors. Either the equations are mis-specified or the code computed the inverse ratio. The manuscript alone cannot disambiguate, and the GitHub repo is referenced without a commit hash or verification.\n\nThe other soft spots are secondary: single synthetic runs with no error bars, hyperparameters chosen without sensitivity analysis, and no confidence intervals for the empirical rankings. These would matter even if the sign were fixed. But the sign issue is not cosmetic—it undermines the empirical conclusion as written.\n\nWho is this for? Energy-finance empiricists interested in EUA drivers and method developers using DII. If the authors fix the sign and re-run (or confirm the code used the opposite convention), the applied result is worth revisiting. As is, I would not cite the empirical ranking.\n\nRecommendation: send to peer review anyway, with a request that the authors reconcile Eqs. (8)–(11) with the reported values and add basic uncertainty quantification. The topic is relevant and the fix appears straightforward; a serious referee could get the paper into usable shape. I would not desk-reject it.","headline":"The empirical headline doesn't follow from the paper's own equations—the sign in the Imbalance Gain is inverted, so the positive IGs in Section 4 cannot be produced by Eq. (9)/(11).","tokens_in":20579,"tokens_out":3922,"would_cite":false,"duration_ms":44901,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A rank-based metric finds the same EUA drivers as Granger tests, plus the nonlinear ones Granger misses.","keywords":["EU ETS","carbon allowances","causal discovery","Differentiable Information Imbalance","Granger causality","non-linear causality","financial returns","EUA"],"falsifier":"Rebuild the DII and Granger analysis on a dataset that adds a plausible measured confounder, such as European industrial production or natural gas prices; if the IBEX35/coal-futures causal weights vanish, then the paper's claimed direct causal effects are confounded.","tokens_in":1307,"feed_emoji":"📈","tokens_out":1309,"duration_ms":60627,"temperature":0.7,"pith_summary":"Carbon prices in the EU emissions trading system are driven by an entangled set of energy, equity, and macro variables, and most causal studies rely on linear Granger or vector autoregression tests. This paper argues that a non-parametric, rank-based tool—the Differentiable Information Imbalance—can detect the same linear drivers while also catching non-linear couplings that VAR models miss. Applied to daily financial returns from January 2013 to April 2024, both methods single out the IBEX35 Spanish stock index and ICE Rotterdam coal futures as the clearest causal drivers of EUA returns. Two synthetic systems show that the disagreements between the methods are explainable: linear tests miss quadratic causal links and can invent links when a non-linear common driver exists. If the paper is right, carbon price modelling should treat DII as complementary evidence to linear causality, not a replacement for it.","feed_headline":"One metric finds EUA drivers Granger tests miss","feed_subtitle":"Both linear and rank-based tests flag IBEX35 and coal futures; the rank-based metric adds nonlinear links.","key_machinery":"The central object is the Differentiable Information Imbalance (DII), a rank-based, gradient-optimised measure of how well distances in a weighted predictor space reproduce neighbourhood structure in a target space. Its causal variant, the Imbalance Gain (IG), quantifies how much adding a candidate variable to the predictor set improves prediction of the target one day ahead; a positive IG means the variable carries causal information. The DII carries the argument by being non-parametric, requiring no distributional or linearity assumptions, and by giving each variable a single optimised weight comparable to VAR coefficients, which enables a one-to-one comparison with Granger causality.","core_discovery":"The paper's central empirical claim is that IBEX35 and coal futures are robust causal drivers of EUA returns, detected by both a linear VAR/Granger approach and the non-linear, model-free DII approach. The authors further claim that, where the two methods disagree, the differences stem from known limitations of linear causality: on synthetic data, multivariate Granger tests fail to detect a quadratic causal coupling (a false negative) and detect a spurious causal link transmitted through a non-linear common driver (a false positive), while the DII-based Imbalance Gain detects the true structure in both cases. The paper therefore proposes the Imbalance Gain as a non-linear analogue of the Gra","pith_inferences":["If DII's non-linear advantage generalises, existing Granger/VAR-based spillover studies of carbon and energy markets may undercount crisis-time transmission, because those links can be quadratic or threshold-shaped.","The same pipeline could be transferred directly to other compliance carbon markets, such as the UK ETS or California allowances, where the same non-linearity problem in price formation applies.","A testable extension is to compute IG at lags greater than one day: coal futures may lead EUA returns at a horizon that the paper's τ = 1 analysis does not examine.","Because IG currently lacks a significance threshold, a practical selection rule could be built by permuting predictor labels; the paper itself proposes resampling techniques as future work."],"forward_implications":["EUA return modelling should treat IBEX35 and coal futures as the primary market variables to monitor, since both a linear and a non-linear method independently flag them.","DII-based Imbalance Gain can flag non-linear causal links that Granger F-tests systematically miss, reducing false negatives relative to VAR-only analysis.","Where linear methods create a spurious causal link through a non-linear common driver, DII/IG can avoid that false positive.","Agreement between DII and VAR on a set of drivers can be used as a validation step before committing to a Granger causality conclusion.","Until resampling-based significance tests are developed, DII causal rankings should be used alongside linear statistical tests rather than alone."],"supporting_citations":[{"why":"Defines the Information Imbalance, the rank-based distance measure that DII generalises.","marker":"[28]"},{"why":"Introduces the Differentiable Information Imbalance with automatic feature weighting via gradient descent.","marker":"[49]"},{"why":"Shows how the Information Imbalance detects causal couplings in dynamical systems and supplies the Imbalance Gain formulation.","marker":"[21]"},{"why":"Extends DII to high-dimensional time-series causal discovery, the framework the paper applies.","marker":"[1]"},{"why":"Defines Granger causality, the linear benchmark against which DII results are compared.","marker":"[29]"},{"why":"Prior application of the Information Imbalance to EUA price determinants, which the paper extends to returns.","marker":"[42]"},{"why":"Documents the IBEX35–EUA relationship that the empirical results corroborate.","marker":"[33]"},{"why":"Shows coal futures and EUA volatility links, supporting the coal-futures causal finding.","marker":"[15]"}],"fun_headline_variants":["DII spots nonlinear EUA drivers that Granger misses","Rank-based test finds EUA causes linear methods overlook","Nonlinear causality test reveals extra EUA return drivers","New metric beats Granger on synthetic nonlinear links","Information imbalance catches what Granger causality can't"],"cache_read_input_tokens":22272,"weakest_assumption_plain":"The analysis assumes causal sufficiency—that no unobserved variable drives both EUA returns and the predictors in the dataset; the authors openly note this is likely false because factors like risk aversion and policy expectations are unmeasured.","fun_headline_variants_meta":{"raw":{"variants":["DII spots nonlinear EUA drivers that Granger misses","Rank-based test finds EUA causes linear methods overlook","Nonlinear causality test reveals extra EUA return drivers","New metric beats Granger on synthetic nonlinear links","Information imbalance catches what Granger causality can't"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1034,"prompt_tokens":652,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":396,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":396,"tokens_out":382,"duration_ms":4829,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:44:30.857842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rebuild the DII and Granger analysis on a dataset that adds a plausible measured confounder, such as European industrial production or natural gas prices; if the IBEX35/coal-futures causal weights vanish, then the paper's claimed direct causal effects are confounded.","supporting_citations":[{"cited_title":"Ranking the information content of distance measures.PNAS nexus, 1(2):pgac039, 2022","cited_arxiv_id":null,"evidence_quote":"Defines the Information Imbalance, the rank-based distance measure that DII generalises."},{"cited_title":"Automatic feature selection and weighting in molecular systems using differentiable information imbalance","cited_arxiv_id":null,"evidence_quote":"Introduces the Differentiable Information Imbalance with automatic feature weighting via gradient descent."},{"cited_title":"Robust inference of causality in high-dimensional dynamical processes from the information imbalance of distance ranks","cited_arxiv_id":null,"evidence_quote":"Shows how the Information Imbalance detects causal couplings in dynamical systems and supplies the Imbalance Gain formulation."},{"cited_title":"Linear scaling causal discovery from high-dimensional time series by dynamical community detection","cited_arxiv_id":null,"evidence_quote":"Extends DII to high-dimensional time-series causal discovery, the framework the paper applies."},{"cited_title":"Investigating the price determinants of the european emission trading system: a non-parametric approach.Quanti- tative Finance, 24(10):1529–1544, 2024","cited_arxiv_id":null,"evidence_quote":"Prior application of the Information Imbalance to EUA price determinants, which the paper extends to returns."},{"cited_title":"What happens to the relationship between eu allowances prices and stock market indices in europe?Energy Economics, 81:13–24, 2019","cited_arxiv_id":null,"evidence_quote":"Documents the IBEX35–EUA relationship that the empirical results corroborate."},{"cited_title":"Volatility spillovers and causality of carbon emissions, oil and coal spot and futures for the eu and usa.Sustainability, 9(10), 2017","cited_arxiv_id":null,"evidence_quote":"Shows coal futures and EUA volatility links, supporting the coal-futures causal finding."}],"review_version":1}