{"id":"b3952f4d-b117-4ce1-a45e-b2a3ebd401be","arxiv_id":"2501.04339","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DCIts is a convolutional model whose per-sample transition tensor recovers signed, lag-resolved causal coefficients matching the ground-truth generators of eight synthetic multivariate time series.","lead":"This paper introduces DCIts, a convolutional neural network that forecasts multivariate time series while also outputting a per-sample interaction matrix: which series and lags drive each prediction, with signed strengths. The authors show on eight synthetic benchmarks that these learned interactions closely match the true generating equations, making the model a candidate tool for equation discovery from data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No identifiability result: because only the product α = C◦F enters the loss and α is per-sample, the forecasting objective alone does not pin down the recovered coefficients; the 'equation reconstruction' claim therefore needs either a uniqueness proof or an explicit scope limitation.","rationale":"The paper's central claim is empirical: DCIts is tested on benchmarks generated by exactly the class of linear and elementwise-polynomial transition operators that the architecture can represent, and within that scope the coefficient recovery is convincing, with code available and five-run statistics reported. The weakest point is not the empirical result but the leap from 'a solution with low MSE also has interpretable α' to 'the recovered α is the generating equation.' Because the loss depends only on the product C◦F, and α is allowed to be sample-dependent, the forecasting objective does not formally identify a unique coefficient tensor. The reader identified the same cluster of assumptions; I partially agree and sharpen it to the non-identifiability of per-sample α under the forecasting objective, even when the data are generated inside the model family and all relevant variables are observed. A seed-diversity experiment on the public VAR(2) code would settle whether the stability observed over five runs is a property of the method or of the particular optimization path. Since the reader already conditioned acceptance on addressing identifiability and reproducibility, this concern does not move the verdict; it strengthens the need for the stated conditions.","tokens_in":71,"tokens_out":10719,"duration_ms":186867,"concrete_test":"Run the public VAR(2) example from the GitHub repository with 20 random training seeds on the same data and compare the full α tensors and test MSE. If the test MSE values are statistically indistinguishable but the recovered α entries differ by more than the reported tolerances, then the forecasting objective does not identify the coefficients, and the interpretability claim must be weakened from 'reconstruction' to 'one possible reconstruction.' As an analytic companion, construct an alternative coefficient tensor αt + Bt with Σ_{i,j} Bt,n,i,j Qt,i,j = 0 and verify that the training loss is unchanged wherever the alternative can be realized by the network.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the learned transition tensor α = C◦F (Eqs. 10–12) reproduces the generating coefficients, so that forecasting accuracy certifies interpretability. The load-bearing assumption is that the forecasting objective uniquely determines α. This is not established and is not obviously true. The loss only constrains the product C◦F through the linear-in-α prediction; for any sample window Qt, subtracting from αt any tensor Bt whose contracted product with Qt vanishes for every output leaves the prediction unchanged. Since F and C are unregularized outputs of a flexible network and αt is per-sample, the objective admits many per-sample coefficient tensors with the same loss. The small standard deviations over five runs and the close numerical matches on the benchmarks are encouraging empirical evidence, but they are not an identifiability result; the match to the ground-truth matrices may reflect the particular initialization and optimization trajectory rather than a guaranteed property of the method. This matters because Section 3.3.4 and the conclusion interpret α as the underlying equations themselves. The acknowledged failure on the bias term in Dataset 8 is consistent with this fragility: the network found a low-loss solution that does not match the generating process. This is a correctness/scope risk, not an inconsistency within the stated experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"DCIts is an interpretable deep architecture for multivariate time series forecasting that factorizes the learned transition operator into a Focuser (a temperature-sigmoid mask) and a Modeler (an unconstrained coefficient tensor), yielding a per-sample transition tensor α = C ∘ F. The paper claims that α reproduces the ground-truth generating coefficients on synthetic benchmarks—so that forecasting accuracy certifies interpretability—and that the architecture extends to bias and polynomial higher-order terms. Forecasting performance is compared with IMV-LSTM on eight synthetic datasets, with additional experiments on a cubic map and on window-size selection.","tokens_in":20597,"tokens_out":4442,"duration_ms":45371,"significance":"If the reconstruction claim can be guaranteed or appropriately scoped, DCIts is a useful addition to interpretable time-series modeling: it provides per-sample, signed, lag-resolved coefficients; supports order-resolved polynomial extensions; is benchmarked on eight controlled datasets; reports variance over five runs; and ships open-source code. Close numerical matches to ground truth (e.g., the VAR(2) coefficients and the cubic-map terms) are genuinely encouraging. However, the central claim that α recovers the underlying equations currently rests on an unproven identifiability assumption, and the paper's own Dataset 8 bias-shift result contradicts the claim in one setting. The limitations paragraph in Section 4 does not discuss this core assumption, only computational cost and real-world noise. The paper needs either an identifiability proof under explicit assumptions or a formal narrowing of the interpretability claim.","major_comments":[{"comment":"Equation (4) defines (α♢Qt)n,i,j = αn,i,j × Qn,j, but given the stated meaning of αn,i,j as the effect of the ith series at lag j on the nth target, the multiplication should be with Qi,j, not Qn,j. As written, the equation contracts the wrong index and is inconsistent with the tensor shapes in Eq. (3) and Eq. (10). Please correct the typo or, if the indexing is intentional, explain how the additional n-dependence arises.","section":"2.1, Eq. (4)"},{"comment":"The interpretability claim is not identified by the forecasting objective. The loss depends on α only through the prediction (C∘F)♢Qt, so for any sample Qt one can add a per-sample tensor Bt that is annihilated by the contracted product with Qt without changing the loss. Since F and C are per-sample, unregularized outputs of a flexible network, many coefficient tensors attain the same loss. The paper nevertheless concludes in Section 3.3.4 and Section 4 that α reconstructs the generating equations. The Dataset 8 bias result in Section 3.3.4 (b1 = 0.196 ± 0.004 for both regimes, ground truth 0.2 and 0.7) shows concretely that a low-loss solution need not match the generator. Please add either a formal identifiability proof under stated assumptions (e.g., uniqueness of the minimizer within the factorized family) or an explicit limitation stating that α is only an interpretation of the fitted model, and that equality to the data-generating coefficients is an empirical property demonstrated on the synthetic benchmarks, not a guaranteed consequence of the method.","section":"2.1, 2.2, 3.3.4"},{"comment":"The phrase 'statistically significantly different from zero' is never defined. No test statistic, null distribution, threshold, or multiple-comparison procedure is given. This filter is load-bearing because it is used to conclude that all α1,j,l with l > 5 are zero for Dataset 7. Please specify the exact significance test used, how the null distribution was obtained (e.g., across runs, across samples), and how multiple testing over (i,j,l) was handled, or remove the significance wording and report raw coefficient distributions instead.","section":"3.2, Figure 8"},{"comment":"The choice of loss function for the interpretability results appears to be made post hoc. The paper states that results use MAE 'unless stated otherwise,' and Dataset 4 is singled out as a case where 'MAE behaves better with respect to MSE.' If the loss was selected after inspecting interpretability performance on the test set, the reported faithfulness is optimistically biased. Please clarify whether the loss was selected on a validation set, describe the pre-specified selection rule, or report results for both MSE and MAE on all datasets.","section":"3.3, 3.3.2"}],"minor_comments":[{"comment":"The sigmoid formula σT(x) = 1/(1 − e^{−x/T}) appears to have an incorrect sign in the denominator; the standard temperature-sigmoid is 1/(1 + e^{−x/T}). As written, the formula has a singularity at x = 0 and outputs outside (0,1).","section":"2.1, Eq. (9)"},{"comment":"The notation for the p = 0 term is under-specified: the product ⃝0 is defined as the identity matrix I, but the dimensions of C0, F0, and the broadcasting rule for (C0∘F0)♢I are not stated. Please clarify how the bias term is represented for each target series n.","section":"2.4, Figure 3 and Eq. (15)"},{"comment":"The text alternates between 'DCIts' and 'DCITs'; please standardize the acronym throughout, including in table and figure captions.","section":"3.1, Table 1 and Figure 5"},{"comment":"The claim that the model 'matches or often surpasses existing interpretability methods' is supported only by a comparison with IMV-LSTM. Adding at least one more recent interpretable baseline or softening the claim would make the statement proportionate.","section":"Abstract and Introduction"},{"comment":"The statement 'we did extensive analysis on the influence of temperature' is not accompanied by any results. Either report a temperature-sensitivity experiment (e.g., in supplementary material) or remove the claim.","section":"2.1"},{"comment":"The Appendix text says 'bhigh and bhigh switch, based on a persistence duration' — the duplicated 'bhigh' appears to be a typo; it should read 'blow and bhigh switch.'","section":"Appendix A, Dataset 8"}],"recommendation":"major_revision","confidential_remarks":"The benchmarks are drawn from the authors' own earlier work, and the only deep baseline is IMV-LSTM, which limits the novelty claim. The identifiability issue is the main substantive concern: without a uniqueness result or an explicit scope limitation, the equation-reconstruction claim is not established. The Dataset 8 bias failure is a concrete counterexample and should be either fixed or honestly discussed as a known limitation. I would encourage the editor to treat the revision as requiring a substantive methodological addition (identifiability/scoping) rather than prose changes alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Domjan and Horvatic have built something real. The DCIts architecture—a Focuser that masks irrelevant lags and series, a linear Modeler that assigns signed coefficients, and a Hadamard product forming a per-sample transition tensor—is a clean, sensible design. The best evidence is the coefficient recovery: on the synthetic benchmarks, the learned α values match ground-truth matrices to 1e-3 or better, with tiny standard deviations across runs. That is not circular: the model is trained only on forecasting error, the ground-truth coefficients are external, and the data are generated from known equations. The higher-order extension and the bias term are straightforward but useful, and the GitHub release makes the work reproducible.\n\nThe soft spots are real but manageable. First, Eq. (4) has an index error: it multiplies by Q_{n,j} rather than Q_{i,j}. This is almost certainly a typo, but as written it obscures the model definition and should be fixed. Second, the 'statistically significant' language in Section 3.3 is never defined—no test, threshold, or procedure is given. Third, the loss function (MSE vs MAE) appears to be chosen after looking at results for some datasets; the paper should at least state this was a model selection decision, not a post hoc justification. Fourth, the benchmarks are the authors' own prior work; that is not a flaw by itself, but it means the evidence is narrower than a fully independent evaluation.\n\nThe load-bearing issue is identifiability. The paper claims the learned α reproduce the generating equations. But the forecasting loss constrains only the product C◦F, and C and F are unregularized outputs of a flexible network. Nothing in the architecture guarantees that the optimization's solution equals the true coefficients; the close matches on the benchmarks are encouraging but not a theorem. The authors do acknowledge the Dataset 8 failure—the model misses the bias shift and substitutes a spurious α instead—which is exactly the kind of failure this concern predicts. The paper should either provide an identifiability argument or explicitly scope the interpretability claim to cases where the data-generating process is within the model class and the optimization converges to the relevant solution.\n\nFor whom: applied researchers who want per-sample, signed, lag-resolved interaction structure from multivariate time series, and who are willing to treat the recovery as empirical rather than guaranteed. The paper deserves a serious referee, and with the notation fixed, the significance test described, and the identifiability limitation stated, it would be a solid contribution. I'd send it to review.","headline":"A genuinely interpretable forecasting architecture with strong empirical coefficient recovery on synthetic benchmarks; the equation-reconstruction claim needs a scope limitation, but the paper deserves serious review.","tokens_in":21080,"tokens_out":3161,"would_cite":true,"duration_ms":31374,"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":"DCIts claims a deep forecaster can output the exact coefficients of the equations generating a time series.","keywords":["multivariate time series forecasting","interpretable deep learning","model interpretability","mechanistic interpretability","transition tensor","convolutional neural network","higher-order terms","causal discovery"],"falsifier":"Run the same protocol on a dataset in which an unobserved variable drives both an observed cause and its effect, so the observed lags do not close the system; if DCIts still reports a stable nonzero $\\alpha$ for the spurious link, the identity between $\\alpha$ and the causal coefficients fails. A second, already-present check is Dataset 8's high-bias regime, where the paper reports a bias near 0.196 and a lag-one coefficient near 0.49 instead of the generating bias 0.7 and zero coefficients; a reader can rerun that experiment and look for the same miss.","tokens_in":20114,"feed_emoji":"📈","tokens_out":8404,"duration_ms":73873,"temperature":0.7,"pith_summary":"DCIts is a deep convolutional forecaster for multivariate time series whose output includes, for every prediction window, a per-sample transition tensor $\\alpha$ that says which source series, at which lags, with what sign and magnitude, drive each target series. The paper argues that on synthetic benchmarks with known generating equations, this tensor equals the ground-truth coefficient matrices to several decimal places, so DCIts is not just forecasting but reconstructing the equations that produced the data. It reports forecasting error equal to or better than the IMV-LSTM baseline while providing local interpretability for every sample. The point of the paper is that interpretability can be built into a deep forecaster as a factorization of its computations, with prediction accuracy acting as a faithfulness constraint.","feed_headline":"Deep forecaster reveals the equations driving the data","feed_subtitle":"DCIts returns signed, lag-by-lag coefficients for every prediction and matches known ground truth on synthetic benchmarks.","key_machinery":"The load-bearing object is the transition tensor $\\alpha \\in \\mathbb{R}^{N \\times N \\times L}$, whose entry $\\alpha_{n,i,j}$ is the coefficient for how series $i$ at lag $j$ contributes to series $n$'s next value. DCIts computes $\\alpha$ as the Hadamard (elementwise) product of two tensors produced by the same convolutional backbone: the Focuser $F = \\sigma_T(H)$, a temperature-dependent sigmoid mask selecting relevant source-series and lag pairs, and the Modeler $C = W_H H + b_H$, an unconstrained linear coefficient map that can be negative and thus captures anti-correlation. The prediction is $X_{n,t+1} = \\sum_{i,j} ((C \\circ F) \\diamondsuit Q_t)_{n,i,j}$, where $\\diamondsuit$ denotes elementwise multiplication with broadcasting over the window $Q_t$. The extended version adds parallel Focuser and Modeler branches for the bias term ($p=0$) and for elementwise polynomial powers of the window ($p \\ge 1$), giving order-resolved coefficients $\\alpha^{(p)} = C^{(p)} \\circ F^{(p)}$.","core_discovery":"The paper's central claim is that the transition operator of a multivariate time series can be learned explicitly as $\\alpha = C \\circ F$, the Hadamard product of a Focuser mask $F$ and a Modeler coefficient tensor $C$, and that this $\\alpha$ recovers the actual coefficients of the data-generating process. On a VAR(2) example the model returns entries such as $0.40 \\pm 10^{-5}$ against ground truth $0.40$; on the autoregressive Dataset 2 it returns diagonal $\\beta$ values near one against an identity ground truth; on the cross-correlated Dataset 4 it returns $\\alpha_{1,2,2} = 0.396 \\pm 0.004$ against the generating $2/5$; on Dataset 7 it recovers negative interactions such as $\\alpha_{2,1,2} = -0.94 \\pm 0.04$ against $-1$; and on the cubic process it recovers both linear and cubic coefficients with the correct sign and order while setting all quadratic terms to zero. Extended DCIts also detects regime-dependent dynamics in Dataset 8, where the learned coefficients split into bimodal distributions matching the switching ground truth, though the model misses the bias shift in the high regime.","pith_inferences":["If the identity between $\\alpha$ and the generating coefficients holds beyond polynomial synthetic systems, DCIts would give empirical researchers a signed, lag-resolved local causality map that current feature-importance methods do not provide.","The Dataset 8 bias miss suggests the factorization is not unique when a constant shift can be explained either by the bias branch or by a spurious lag coefficient; a diagnostic extension would be to compare $\\alpha$ across multiple optima of the forecasting loss.","The order-resolved higher-order branches could double as a misspecification test: on non-polynomial dynamics the higher-order coefficients should scatter across samples instead of stabilizing, signaling that the interpretable polynomial form is not the right generator."],"forward_implications":["From the learned $\\alpha$ a user can read off the generating equations directly, as demonstrated by the VAR(2) reconstruction where all matrix entries match their ground-truth values to roughly $10^{-4}$ or better.","Per-sample $\\alpha$ lets the model separate different regimes in one dataset: in Dataset 8 the coefficients $\\alpha_{2,1,5}$ and $\\alpha_{2,4,2}$ form bimodal distributions centered at the two switching ground truths.","The significance pattern of $\\alpha$ also identifies the maximum relevant lag, so the window size can be chosen as the smallest lag with significant coefficients, as shown for Datasets 2, 4, and 7.","Higher-order branches recover the true polynomial order of the generator: on the cubic process the model finds the linear and cubic terms, finds no quadratic terms, and classifies the third series as purely linear.","Interpretability does not cost accuracy: DCIts reports equal or lower mean squared error than the IMV-LSTM baseline on all eight datasets in the paper."],"supporting_citations":[{"why":"Supplies the eight benchmark datasets with known generating equations that the $\\alpha$ and $\\beta$ reconstructions are verified against.","marker":"[2]"},{"why":"Provides the interpretable baseline whose forecasting error and stability DCIts is compared with on all eight datasets.","marker":"[4]"},{"why":"Defines the six-criterion evaluation scheme used to position DCIts against the baseline.","marker":"[3]"},{"why":"A two-step architecture that DCIts extends by not decoupling autocorrelation from cross-correlation.","marker":"[25]"},{"why":"Another two-step architecture DCIts contrasts with to justify its Focuser and Modeler design.","marker":"[26]"},{"why":"An ensemble feature-importance causality approach that reports existence of lagged influence without sign or magnitude, motivating the signed transition tensor.","marker":"[24]"}],"fun_headline_variants":["Deep forecaster writes the equations it uses","AI reveals signed lag-by-lag coefficients","Neural net shows its work on time series","Interpretable deep model recovers true dynamics","From black box to clear equations in forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The data are assumed to be generated exactly by a finite-order polynomial (or linear) function of the observed lags of the observed series, with no unobserved confounders, and the forecasting loss is assumed to have a unique minimum at which the fitted coefficients equal the causal ones.","fun_headline_variants_meta":{"raw":{"variants":["Deep forecaster writes the equations it uses","AI reveals signed lag-by-lag coefficients","Neural net shows its work on time series","Interpretable deep model recovers true dynamics","From black box to clear equations in forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1465,"prompt_tokens":1019,"completion_tokens":446,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":379}},"tokens_in":635,"tokens_out":446,"duration_ms":4893,"temperature":1.0,"reasoning_tokens":379,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:35:08.461294+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same protocol on a dataset in which an unobserved variable drives both an observed cause and its effect, so the observed lags do not close the system; if DCIts still reports a stable nonzero $\\alpha$ for the spurious link, the identity between $\\alpha$ and the causal coefficients fails. A second, already-present check is Dataset 8's high-bias regime, where the paper reports a bias near 0.196 and a lag-one coefficient near 0.49 instead of the generating bias 0.7 and zero coefficients; a reader can rerun that experiment and look for the same miss.","supporting_citations":[{"cited_title":"Bari´ c, P","cited_arxiv_id":null,"evidence_quote":"Supplies the eight benchmark datasets with known generating equations that the $\\alpha$ and $\\beta$ reconstructions are verified against."},{"cited_title":"A Performance-Explainability Framework to Benchmark Machine Learning Methods: Application to Multivariate Time Series Classifiers","cited_arxiv_id":"2005.14501","evidence_quote":"Defines the six-criterion evaluation scheme used to position DCIts against the baseline."}],"review_version":1}