{"id":"73cdd18f-79c7-45ec-a9b8-4e873d2b5b28","arxiv_id":"2507.09439","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DyCAST-Net reads causal links and lags from a hybrid TCN plus sparse-attention predictor and reports top F1 and delay accuracy on finance and fMRI causal discovery benchmarks.","lead":"This paper proposes DyCAST-Net, a neural network that combines dilated temporal convolutions with sparsely pruned attention to infer cause-and-effect links and time delays in multivariate time series, and reports higher F1 and delay-estimation scores than TCDF, GCFormer, and CausalFormer on finance and fMRI benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Delay estimator in §3.3.2 is asserted without derivation; in a stacked TCN with skip connections and attention, argmax_k |w_j[k]| × d does not track true input lag, so the headline DEA 0.89 is not yet supported.","rationale":"The reader’s verdict is CONDITIONAL, and I agree that outright rejection is not warranted: the benchmarks are external simulated datasets, which reduces circularity, and the architecture is described in enough detail to test. However, the strongest claim bundles three metrics: F1/Recall rest on permutation-importance edge extraction, while DEA rests on the conv-filter delay readout. Among these, the delay readout is the least secure because it is a single unparameterized formula, contradicted by the paper’s own residual/skip-connection description, and it is never validated. This is internal inconsistency rather than disagreement with the field’s consensus: a stacked dilated TCN with skip connections simply does not have the property that a late-layer filter peak maps linearly to an original time lag. The concrete synthetic-lag experiment, or an independent derivation of the effective receptive field, would settle whether the delay estimator is biased. Until then, the ‘highest DEA’ and ‘more precise causal delays’ parts of the central claim are unverified. This does not move the verdict away from CONDITIONAL; it specifies the condition that must be met: the delay estimator must be derived or empirically validated, and Table 4 must report variance and per-dataset results rather than single point estimates. The reader’s weakest_assumption lists the delay readout as one of two premises, and also flags the ground-truth datasets and confounder assumptions, so our focus agrees partially with that framing while identifying the delay estimator as the most directly checkable load-bearing element.","tokens_in":8951,"tokens_out":4760,"duration_ms":59090,"concrete_test":"Construct synthetic MTS from Y_t = Σ_{l∈L} B_l X_{t−l} + ε with known lags L = {1, 2, 3, 5, 7, 11}, train DyCAST-Net using the Section 4.3 settings, and extract delays using §3.3.2 across 100 seeds. Report the mean and distribution of |d_hat − d| per edge and compare against a maximum-cross-correlation baseline. If the argmax × d heuristic is biased for lags not equal to simple multiples of the dilation factor, or if predicted delays cluster at a few quantized values, the DEA 0.89 claim lacks its only direct support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing component of the SOTA claim is delay estimation: §5 reports DEA 0.89 vs 0.85 for CausalFormer, and the inferred causal graph labels every edge with a delay. Section 3.3.2 defines delay_j = argmax_k |w_j[k]| × d, but this is not derived. In DyCAST-Net, blocks are stacked (Section 3.2.1: depthwise dilated conv, chomp, normalization, residual/skip connection) and time attention is also present. The input to layer l is the output of layer l−1, not the raw channel j; residual/skip connections add earlier, less-dilated features back into later stages. A large absolute weight at position k in a late filter therefore does not correspond to a unique original time offset k·d: the effective receptive field mixes all layers, and nonlinearities destroy any linear peak-to-lag mapping. The paper supplies no derivation, no synthetic validation of the estimator, and no error bars: Table 4 gives one DEA number per model. Since Eq. 5.1 averages absolute delay error with no tolerance, a systematic off-by-one in the heuristic would drastically change the reported advantage. Hyperparameters also conflict (kernel size 6 in Table 2, kernel size 4 in Table 3), making the d factor ambiguous. If the delay readout is unvalidated, the ‘more precise estimation of causal delays’ claim is unsupported even if F1 for edge presence is correct.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DyCAST-Net, a hybrid architecture combining dilated temporal convolutions and dynamic sparse multi-head attention for causal discovery in multivariate time series. It trains one network per target variable, uses channel attention weights and a permutation shuffle test to select candidate causes, and reads off causal delays from the argmax of dilated convolution filter weights. The resulting directed graph is evaluated on Fama-French financial and Smith fMRI simulated benchmarks, where Table 4 reports higher F1, Recall, and Delay Estimation Accuracy (DEA) than TCDF, GCFormer, and CausalFormer. The paper also presents attention heatmaps as interpretable evidence.","tokens_in":9116,"tokens_out":4170,"duration_ms":42567,"significance":"If the empirical claims were fully supported, DyCAST-Net would offer a practically useful, interpretable causal discovery tool with sparse attention and explicit delay labeling. The strengths include a plausible architectural combination of TCN and attention, a statistical shuffle test to filter false positives, and evaluation against established simulated ground truth (Fama-French and Smith fMRI). However, the manuscript currently lacks statistical rigor: the headline comparison table reports point estimates with no variance or significance tests, and the delay-estimation step is an unvalidated heuristic in a network with residual connections and nonlinearities. The central SOTA claim therefore rests on an unsupported component. The ideas are worth pursuing, but the manuscript needs additional validation and revision before the claims can be accepted.","major_comments":[{"comment":"The delay estimator, delay_{j→target} ≈ argmax_k |w_j[k]| × d, is asserted without derivation or validation. In the stacked DyCAST-Block architecture (§3.2.1) with residual/skip connections, normalization, and nonlinearities, the position of the maximum absolute weight in a late filter does not necessarily correspond to a unique original input lag. The paper provides no synthetic validation of this estimator, and Table 4 reports a single DEA number per model with no error bars. Since every edge in the inferred causal graph is labeled with this delay, the claim of 'more precise estimation of causal delays' (Section 5.1) is unsupported, and even the F1 comparison may be affected if the heuristic has a systematic offset.","section":"§3.3.2"},{"comment":"The 5-fold expanding-window protocol yields five point estimates per metric, yet Table 4 reports single numbers with no standard deviation, confidence interval, or significance test. As a result, the differences between DyCAST-Net and CausalFormer (e.g., F1 0.91 vs 0.87; DEA 0.89 vs 0.85) cannot be assessed for statistical reliability. Additionally, Section 4.1 reports post hoc exclusions of one Finance dataset 'lacking causal relationships' and one fMRI dataset 'due to processing limitations' without pre-registered criteria; this risks selection bias and should be justified explicitly.","section":"§4.2, Table 4"},{"comment":"The causal interpretation equates predictive importance (channel attention weights and shuffle-test degradation) with causal influence. This identification silently assumes the absence of hidden confounders and contemporaneous correlations. In real-world MTS, such assumptions rarely hold, and the reported F1 and Recall against simulated ground truth may not transfer to real data. The paper should state this assumption explicitly and either evaluate on a benchmark with known confounders or add a discussion of this limitation as a caveat to the causal claims.","section":"§3.3.1, §3.3.3"},{"comment":"Hyperparameters are inconsistent: Table 2 reports Kernel Size 6, Dilation Rate 4, and Number of Levels 2, whereas Table 3 reports kernel size 4 and dilations 2^0, 2^1, 2^2. Since the delay estimator multiplies the argmax kernel index by a dilation factor, this ambiguity materially affects the DEA computation and the reproducibility of the results. Also, Table 2's 'Significance Level 0.5' is implausible as a significance level (likely 0.05); please clarify how this value is used in the shuffle test.","section":"Tables 2 and 3"},{"comment":"The dataset description is internally inconsistent: Section 4.1 describes 10 Finance datasets and 28 fMRI datasets, but Sections 4.4 and 4.5 refer to 'the Finance dataset consists of 25 time series' and 'fMRI dataset ... sample of 9 time series,' and Figure 3's caption says 'RMFI.' These inconsistencies make it impossible to know which datasets were actually used, how the ground-truth causal graphs were defined, and how the reported metrics aggregate across datasets.","section":"§4.1, §4.4, §4.5"}],"minor_comments":[{"comment":"The caption 'Causal relationships in RMFI data' should read 'fMRI data'; 'RMFI' is a typo.","section":"Figure 3"},{"comment":"The abstract claims validation on 'marketing datasets,' but the experimental section uses only Finance and fMRI data; no marketing dataset appears in the evaluation.","section":"Abstract"},{"comment":"The sentence 'Table 3 presents a comparative analysis' refers to the performance comparison table, which is actually numbered Table 4; please update the cross-reference.","section":"§5"},{"comment":"The conclusion mentions 'MGCFORMER' as a compared method, but the correct name is 'GCFormer' throughout the rest of the paper.","section":"Conclusion"},{"comment":"The phrase 'This can be aggregated if multiple layers contribute' is vague; no aggregation method is specified, so it is unclear how the final delay is computed when multiple layers are used.","section":"§3.3.2"},{"comment":"The 'Remark on Interpretation' says edge values represent the estimated delay or alternatively the confidence/strength derived from attention weights; this ambiguity should be resolved, as the evaluation in Table 4 presumably uses only the delay interpretation.","section":"§3.3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from a careful revision addressing the statistical reporting and the delay estimator. I also note the absence of code or a reproducibility statement; given the hyperparameter inconsistencies (Tables 2 and 3), it would be hard for a third party to reproduce the results. This is my main concern for the editor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: DyCAST-Net is a hybrid TCN-plus-sparse-attention architecture for multivariate time series causal discovery, evaluated on Fama-French and Smith fMRI benchmarks. The specific combination is new, and the authors get credit for benchmarking against external simulated ground truth. But the paper's most distinctive claim—more precise causal delay estimation—rests on an unvalidated heuristic that does not survive scrutiny, and the empirical evidence as presented is too thin to support \"consistently outperforms.\"\n\nWhat's actually new: the architecture is a recombination of known pieces (dilated causal convolutions, multi-head attention, hard threshold sparsity, RMSNorm/LayerScale, channel weighting, permutation importance). That's not a conceptual advance, but it's a legitimate engineering contribution. The benchmark results on Fama-French and Smith fMRI are new empirical data points, and the paper makes a credible attempt to compare against three relevant baselines (TCDF, GCFormer, CausalFormer). The shuffle test for channel selection is standard but used honestly.\n\nThe soft spots are significant. First, the delay estimator in Section 3.3.2 is asserted without derivation: argmax_k |w_j[k]| × d does not map to original time lag in a stacked TCN with residual/skip connections and attention, because later layers see earlier layers' outputs, not raw inputs. Without a derivation or synthetic validation, the DEA numbers in Table 4 are not interpretable. Second, Table 4 reports single point estimates with no variance, confidence intervals, or significance tests across the 5-fold protocol—so even the F1 gains cannot be distinguished from noise. Third, the abstract promises marketing datasets and advertising case studies, but the body only reports finance and fMRI; that mismatch, plus two datasets excluded post hoc (one for \"lacking causal relationships,\" one for \"processing limitations\"), undercuts the generality claims. Fourth, hyperparameters conflict (kernel size 6 in Table 2 vs 4 in Table 3; dilation rate 4 vs 2^l) and no code or data is released, which hurts reproducibility. None of these is a demonstrated fatal error in the core idea, but together they mean the paper is nowhere near \"consistently outperforms.\"\n\nWho is this for? Researchers working on temporal causal discovery with deep learning might find the architecture a useful starting point. But it needs a serious round of revision: derive or validate the delay estimator, add error bars, fix the inconsistencies, and release artifacts.\n\nMy recommendation: this deserves peer review, not desk reject, because the underlying idea is plausible and the benchmark setup is sound. But if I were the editor, I'd send it back with major revision required.","headline":"A plausible hybrid TCN/sparse-attention architecture with external benchmark ground truth, but the delay-estimation claim is unsupported and the empirical evidence lacks variance estimates.","tokens_in":9818,"tokens_out":3362,"would_cite":false,"duration_ms":35066,"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":"DyCAST-Net claims that temporal causal structure in multivariate time series can be recovered from a forecasting network's trained parameters: causal channels come from attention weights plus a shuffle test, and causal delays come from…","keywords":["Causal Discovery","Multivariate Time Series","Dilated Convolutions","Sparse Attention","Financial Analytics","Marketing Impact Analysis","Delay Estimation","Temporal Causal Graph"],"falsifier":"Generate multivariate series from a linear structural model with a known hidden common cause $Z$ driving both $X$ and $Y$ but no direct edge $X\\to Y$; if DyCAST-Net's shuffle test still certifies $X\\to Y$ because permuting $X$ degrades the forecast through $Z$'s influence, the causal-identification step is falsified. A complementary test: include ground-truth edges with zero delay, since the architecture's causal masking and convolutions only use past values, and check whether those contemporaneous edges are unrecoverable and recall falls measurably.","tokens_in":8579,"feed_emoji":"🧠","tokens_out":6609,"duration_ms":74645,"temperature":0.7,"pith_summary":"DyCAST-Net is a hybrid neural architecture that tries to establish a direct pipeline from raw multivariate time series to an interpretable directed causal graph, complete with estimated time delays. The paper argues that after training the network for next-step prediction, the model's own parameters carry causal information: channel-level attention weights plus a shuffle test identify which input series matter, and the peaks of dilated convolution filters fix how far back in time each cause acts. On simulated Fama-French financial data and simulated fMRI brain-activity data, the authors report that DyCAST-Net outperforms TCDF, GCFormer, and CausalFormer in F1 score, recall, and delay-estimation accuracy. The payoff, if the claim holds, is causality discovery that does not require controlled interventions and whose results are human-auditable through attention heatmaps and delay-labeled edges.","feed_headline":"One network reads causal links and delays straight from its weights","feed_subtitle":"DyCAST-Net beats TCDF, CausalFormer, and GCFormer at causal discovery on finance and simulated fMRI benchmarks.","key_machinery":"The load-bearing mechanism is the combination of channel-level attention weighting, dilated depthwise convolutions with causal masking, and the post-hoc reading of parameters. An adaptive threshold zeroes out small attention weights, producing dynamic sparsity that filters spurious connections; RMSNorm and LayerScale stabilize training; and the shuffle test validates each candidate cause by permuting that channel's time points and measuring whether the forecast error worsens. The delay estimator uses the index of the maximum absolute filter weight $k^* = \\arg\\max_k |w_j[k]|$ times the dilation factor $d$, optionally aggregated across layers, to place each causal edge at a specific lag.","core_discovery":"On its own terms, the paper's central claim is that temporal causal discovery can be solved by interpreting a trained forecasting model rather than by searching over causal graphs. DyCAST-Net combines dilated temporal convolutions with dynamic sparse multi-head attention; once trained to minimize next-step mean squared error, a channel is declared a cause of the target when its learned attention weight exceeds a threshold and shuffling its values degrades prediction error. The causal delay is then read off as $\\arg\\max_k |w_j[k]| \\times d$ from the convolution filters, and directed edges labeled with those delays form the temporal causal graph. The reported numbers on the two simulated benchmarks are F1 0.91 (finance) and 0.89 (fMRI), recall 0.90 and 0.89, and delay estimation accuracy 0.89, the latter compared with 0.85 for CausalFormer.","pith_inferences":["Because causation is identified with predictive importance in a single forecasting model, the discovered graph should be read as mixing direct and mediated influence: permuting a mediator also degrades the forecast, so edges may not isolate direct paths.","The causal reading silently assumes no hidden confounders and no contemporaneous causation, so a natural extension would benchmark DyCAST-Net against conditional-independence methods on synthetic data with a known latent common cause.","The delay estimate's resolution is tied to the dilation schedule and kernel size, so datasets with delays that are not integer multiples of the dilation base may force rounding; testing on continuous-delay ground truth would reveal this limit.","Attention heatmaps could be turned into a quantitative test by checking whether their structure on held-out domains matches known anatomical or economic pathway structure."],"forward_implications":["If the claim is right, one can obtain a temporal causal graph simply by training a forecaster and inspecting its weights, avoiding explicit graph search or intervention experiments.","Delay-labeled edges make the output directly usable for lag-sensitive decisions, such as how many periods before a macroeconomic shock affects a market factor.","Dynamic sparsity plus the shuffle test gives a false-positive filter, which the paper argues matters in noisy environments and high-dimensional settings.","Because the architecture is domain-agnostic, the same pipeline should transfer to other multivariate time series problems, including marketing mix and sensor data.","On the two benchmarks used, the reported numbers (F1 0.91 on finance and 0.89 on fMRI, DEA 0.89) define the comparison point that future causal-discovery models must beat."],"supporting_citations":[{"why":"Supplies the TCDF baseline and the attention-convolution approach that DyCAST-Net extends and must outperform.","marker":"[nauta2019causal]"},{"why":"Supplies the CausalFormer baseline, which is the strongest comparator and the reference for the DEA comparison.","marker":"[Kong2024CausalFormerAI]"},{"why":"Provides the Fama-French Three-Factor model used to generate the finance benchmark's ground-truth causal graphs.","marker":"[Kleinberg2013]"},{"why":"Provides the simulated BOLD networks that serve as ground truth for the fMRI benchmark.","marker":"[Smith2011]"},{"why":"Supplies the multi-head attention mechanism that the dynamic sparse attention and interpretability pipeline builds on.","marker":"[vaswani2017attention]"},{"why":"Supplies the GCFormer baseline, a Granger-causality attention model used in the comparative study.","marker":"[Xing2023GCFormerGC]"}],"fun_headline_variants":["Read causal links and delays straight from a network's learned weights","No graph search needed: sparse attention weights reveal temporal causes","DyCAST-Net beats baselines by reading delays off its own filters","Sparse attention exposes hidden causal patterns and lagged delays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that 'permuting this channel makes the forecast worse' means 'this channel causes the target,' which silently assumes there are no hidden confounders driving both and no contemporaneous same-time effects, and that the simulated Fama-French and Smith fMRI graphs used as ground truth are complete and correct.","fun_headline_variants_meta":{"raw":{"variants":["Read causal links and delays straight from a network's learned weights","No graph search needed: sparse attention weights reveal temporal causes","DyCAST-Net beats baselines by reading delays off its own filters","Sparse attention exposes hidden causal patterns and lagged delays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001037,"raw_usage":{"total_tokens":4386,"prompt_tokens":985,"completion_tokens":3401,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":3330}},"tokens_in":601,"tokens_out":3401,"duration_ms":28335,"temperature":1.0,"reasoning_tokens":3330,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:58:19.860963+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate multivariate series from a linear structural model with a known hidden common cause $Z$ driving both $X$ and $Y$ but no direct edge $X\\to Y$; if DyCAST-Net's shuffle test still certifies $X\\to Y$ because permuting $X$ degrades the forecast through $Z$'s influence, the causal-identification step is falsified. A complementary test: include ground-truth edges with zero delay, since the architecture's causal masking and convolutions only use past values, and check whether those contemporaneous edges are unrecoverable and recall falls measurably.","supporting_citations":[],"review_version":1}