{"id":"0b9847f2-7373-45c1-b43b-071bf0a664e6","arxiv_id":"2411.10152","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Aligning cause-effect pairs by their estimated Granger lag improves channel-dependent forecasting accuracy and transfer learning on synthetic time-series data.","lead":"This paper proposes a pre-training strategy for multivariate time-series forecasting that splits data into cause-effect pairs, estimates causal lags with Granger causality, and shifts the cause series so that lagged effects align in time. On synthetic data, the synchronization improves forecasting accuracy for four neural models and improves transfer to unseen causal structures.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The target-oriented shift aligns cause and effect within the training window, but for H-step-ahead forecasts the shifted cause needed for X_{t+H} lies at t+H, outside the input context; with C=30, H=10 and lags up to 200, the stated mechanism cannot place the needed historical cause in the context.","rationale":"The reader's concern about lag-estimation accuracy is legitimate but secondary: it assumes that if the lags were known exactly, the method would work. The shift/horizon analysis shows that even with oracle lags, the proposed transformation does not place the relevant cause for the forecast target into the input context. This is a correctness risk that undermines the central claim rather than the experimental reporting. The concrete check is cheap: because the synthetic setup has known ground-truth lags, one can inspect the constructed windows and run a controlled lag value. If the check confirms the issue, the paper's interpretation of Tables 1 and 2 cannot be maintained; the improvements would need to be explained by a different mechanism or by an undocumented implementation detail. I therefore recommend a conditional acceptance: the authors must specify the exact shift and window construction, provide code, and show results for lags exceeding the context. If they cannot, the paper should not be accepted as is.","tokens_in":11899,"tokens_out":16345,"duration_ms":165268,"concrete_test":"On a synthetic dataset with a single known pair, set δ=200, C=30, H=10 and forecast origin T. Before training, compute the index of the cause value needed for target X_{T+10}: it is Z_{T-190}. Check whether this value appears in either the synchronized channel (\\tilde Z_{T-30:T}=Z_{T-230:T-200}) or the non-synchronized channel (Z_{T-30:T}); it appears in neither. Re-run Experiment 1 with all pairs having δ=200 (or δ=C+H+1) and otherwise identical. If synchronization still improves MAPE, the implementation must be using information outside the stated context (e.g., longer windows or future covariates), so the paper must disclose that; if it does not improve, the reported gains in Table 1 are not attributable to the proposed mechanism.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Under the definition in §3.2, shifting cause Z by lag δ gives \\tilde Z_s = Z_{s-δ}, so X_s depends on \\tilde Z_s. In §3.3 and Fig. 3, the model predicts X_{t+1:t+H} from context [t-C:t]. For horizon step h, the cause that matters is Z_{t+h-δ}; in the shifted input this is \\tilde Z_{t+h}, which is outside the context. The only channel that contains Z_{t+h-δ} for h≤δ is the non-synchronized cause Z_{t-C:t}, and only if δ-h ≤ C. Thus the synchronization does not bring outside-context lags into the model; it only makes the historical relationship contemporaneous for training. With C=30, H=10 and lags up to 200, any δ>40 means the required Z_{t+H-δ} is absent from both inputs. The paper presents Table 1 as evidence that synchronization captures highly lagged dependencies, but the described pipeline cannot supply that information. The paper does not specify an alternative shift (e.g., δ-H), a longer cause-specific context, or future-known covariates. This is independent of lag-estimation error: even with oracle lags, the shift as described does not achieve the stated goal.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a preprocessing and pre-training strategy for multivariate time-series forecasting in process-industry digital twins. Using linear Granger causality, the method identifies cause-effect pairs and their time lags, then applies a 'target-oriented causal shift' to align each cause series with its effect before training channel-dependent models (TSMixerx, MLPMultivariate, BiTCN, TFT). The shifted cause and a non-shifted copy are fed as inputs alongside the target, and this synchronized format is used both for end-to-end forecasting and for pre-training on multiple synthetic source datasets followed by fine-tuning on a target dataset. Experiments on synthetic data with known injected causal lags report MAPE improvements for all four models in Table 1 and an advantage for synchronized pre-training in Table 2. The central claim is that synchronization brings highly lagged causes into the context window and improves generalization in channel-dependent forecasting.","tokens_in":12084,"tokens_out":7173,"duration_ms":71778,"significance":"If the mechanism worked as stated, the idea of breaking a complex causal structure into synchronized cause-effect pairs for pre-training would be a useful and timely contribution for industrial time-series forecasting, addressing the known weakness of channel-dependent models on misaligned variables. The use of synthetic data with known ground-truth causal lags is a good first validation choice, and including both the synchronized and non-synchronized cause channels is a sensible safeguard against information loss. The paper also tests the strategy across several model families, which strengthens the empirical question under study. However, the central mechanism as written does not deliver the promised out-of-context lag information, and the experimental evidence lacks the uncertainty quantification and baseline controls needed to support the claimed improvements.","major_comments":[{"comment":"The target-oriented shift as defined does not place the cause values needed for H-step-ahead forecasts into the input context. If shifting means \\tilde Z_s = Z_{s-δ_{ij}}, then the synchronized context \\tilde Z_{t-C:t} contains cause values Z_{t-C-δ:t-δ}, while the non-synchronized input Z_{t-C:t} contains values up to Z_t. For target step h, X_{t+h} depends on Z_{t+h-δ}; when δ-h > C, that value appears in neither input channel. With the reported C=30, H=10 and lags up to 200, any δ>40 makes the cause for X_{t+H} unavailable to the model. The statement in §4.1 that synchronization shifts 'causes from outside the context window' is therefore not realized by the described pipeline. The authors should either correct the shift definition (for example, a horizon-specific shift of δ-H), add a cause-specific context that extends further back, or show an alternative construction; the experiments must then be rerun under the corrected definition.","section":"§3.2, §3.3, Fig. 3"},{"comment":"The empirical core of the paper is a synchronized-versus-non-synchronized comparison on ten synthetic datasets, but Table 1 reports a single MAPE per model with no error bars, no standard deviations, and no statistical test across datasets. The paper also never reports how often the data-driven Granger lag estimates match the injected ground-truth lags, so the reader cannot tell whether the improvement reflects correct synchronization or an artifact of the estimation procedure. The manuscript itself acknowledges in §4.1 that the improvements 'are heavily dependent on causal pair extraction and lag estimation'; that step needs to be validated directly, e.g., by lag-recovery rates and a sensitivity analysis around misspecified δ values.","section":"§4.1, Table 1"},{"comment":"Table 2 does not support the unqualified claim in §4.2 that pre-training on many causal structures improves over supervised training. In the non-synchronized rows, pre-training plus fine-tuning (MAPE 2.56) is worse than supervised training on the target only (MAPE 2.42). Only the synchronized configuration shows the claimed improvement (2.21 vs. 2.38). The claim should be restricted to synchronized pre-training, and the comparison needs per-model results or confidence intervals before a generalization statement can be made.","section":"§4.2, Table 2"},{"comment":"The motivation for synchronization is that highly lagged dependencies may lie outside a fixed context window (§2.3), but no baseline with a larger context length is included. Without training the same models with C larger than the maximum relevant lag (e.g., C > 210), the reported NS-versus-S differences cannot distinguish the proposed synchronization mechanism from a generic reformatting of the inputs. This control is necessary to support the paper's central claim that the method specifically recovers out-of-context causal information.","section":"§4.1, §2.3"}],"minor_comments":[{"comment":"The metric is named 'Mean Average Percentage Error' in the text and table caption; the standard name for MAPE is 'Mean Absolute Percentage Error', since the absolute errors are averaged.","section":"§4.1, Table 1"},{"comment":"The section heading contains the typo 'Channel Indpendence' and should read 'Channel Independence'.","section":"§2.3"},{"comment":"The caption says 'context - C denotes the context length, i.e. historic data points'; this should be worded as 'context of length C' to avoid confusion with subtraction.","section":"Fig. 3 caption"},{"comment":"The threshold used for Granger causality is never specified; please report the p-value threshold or other criterion used to select cause-effect pairs in both experiments.","section":"§3.1"},{"comment":"Experimental details for the transfer experiment are incomplete: the target dataset size, number of variables, composition of the 100 source datasets, and the fine-tuning protocol (layers, learning rates, epochs) are not specified, which limits reproducibility.","section":"§4.2"},{"comment":"Reference [8] (Fukushima 1975) does not appear to be the appropriate citation for the 'MLPMultivariate' model used in the experiments; please cite the neuralforecast implementation or another directly relevant source.","section":"Reference [8]"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the proposed synchronization mechanism, as written, cannot deliver out-of-context lag information for the forecasting setup used in the experiments. This is fixable by redefining the shift relative to the prediction horizon or by extending the cause context, but it requires redoing the experimental validation. In addition, the absence of uncertainty quantification and lag-recovery checks makes the current empirical support too weak for publication. If the authors address these points, the idea may be worth a second round."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:2411.10152. The core idea is sensible: identify cause-effect pairs with Granger causality, align them by the estimated lag, and train channel-dependent forecasting models on the aligned pairs rather than on the full multivariate mess. That is a clean preprocessing concept, and the authors deserve credit for making it explicit and for adding the non-synchronized cause channel to avoid discarding information. The second thing is that the shift as implemented does not do what the paper says it does.\n\nThe text in §3.2 and Fig. 3 defines the shift as moving the cause forward by δ, i.e., \\tilde Z_s = Z_{s-δ}. In a forecasting setting with context [t-C:t] and horizon H, the cause relevant for predicting X_{t+h} is Z_{t+h-δ}. The shifted channel only provides \\tilde Z_{t-C:t} = Z_{t-C-δ:t-δ}, which ends at t-δ; the needed value for any h is not there. It appears in the non-synchronized channel only when δ ≤ C+h, which with C=30, H=10 means δ ≤ 40. The paper motivates the whole method with lags up to 200. So for most of the regime it claims to address, neither input channel contains the needed cause. This is not a question of lag-estimation error; it fails even with oracle lags. The stated mechanism cannot achieve the stated goal.\n\nWhat the paper does well: the pairwise decomposition is a reasonable way to reduce noise from irrelevant channels; the synthetic data with known ground truth is a good test bed; and the empirical improvement in Table 1, if it survives scrutiny, is worth taking seriously. Including the non-synchronized channel is a thoughtful touch. The transfer experiment in Table 2 is a step beyond simple forecasting.\n\nSoft spots beyond the mechanical one: the experiments report no variance or significance tests, so the 9–32% improvements may be within noise. There is no longer-context baseline, which would be the natural ablated control: if a model with C=60 or C=200 captures the same benefit, then the synchronization is irrelevant. The paper never checks how often the estimated Granger lags match the injected ground truth, despite admitting that performance depends heavily on that step. Table 2 actually shows negative transfer for the non-synchronized case (2.56 vs 2.42), which complicates the pre-training claim; only the synchronized pre-training beats supervised training. And no code or data is provided, so reproducibility is weak.\n\nOn balance, this is a paper with a real idea but a load-bearing flaw in the presentation of the mechanism, plus an experimental section that needs more rigor. It is honest about many limitations, but the central contradiction is not acknowledged. A serious referee should see it, because the shift issue is fixable in principle (e.g., a longer context or a different shift convention), and the empirical question of whether synchronized pairs help is legitimate. But I would not accept it in its current form.\n\nRecommendation: send it to review, but flag the shift/context issue explicitly. It deserves referee time, not because it is close to correct, but because the failure mode is instructive and the question is worth answering properly.","headline":"The shift trick as described cannot put highly lagged causes into the context window, so the paper's central claim is unsupported as written.","tokens_in":12695,"tokens_out":6039,"would_cite":false,"duration_ms":59262,"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":"The paper claims that shifting cause variables forward by their estimated causal lags improves forecasting accuracy across four channel-dependent models and makes pre-training transferable to unseen causal structures.","keywords":["time-series forecasting","causal synchronization","Granger causality","channel dependence","transfer learning","pre-training","digital twins","process industry"],"falsifier":"Run the same synthetic experiment with oracle lags versus lags perturbed by one to ten time steps; if MAPE degrades to or below the non-synchronized baseline under even small perturbation, the claimed benefit is an artifact of perfect lag estimation. Alternatively, generate data where lags change at known break points; if synchronized training underperforms ordinary multivariate training on non-stationary lags, the global-stationarity assumption fails.","tokens_in":11602,"feed_emoji":"🔄","tokens_out":7463,"duration_ms":71822,"temperature":0.7,"pith_summary":"The paper claims that channel-dependent forecasting models learn better from temporally aligned cause-effect pairs than from raw multivariate time series. It proposes a data-preparation pipeline: detect lagged cause-effect relationships with a linear Granger causality test (a regression check of whether one series' past helps predict another), shift each cause variable forward by the estimated lag so it lines up with its effect, and construct training windows that keep both the shifted and the unshifted cause alongside the target effect. On synthetic datasets with known causal lags, this synchronization reduces forecast error, measured by mean absolute percentage error (MAPE), for all four tested models, from 7.24 to 5.07 for the weakest baseline and from 3.29 to 2.99 for the strongest. The paper further shows that pre-training on many synchronized source pairs and fine-tuning on a target dataset yields a MAPE of 2.21, the best configuration tested. If the claim holds, it offers a transfer-learning route for process-industry forecasting where variables differ across plants and cause-effect delays can exceed the model's context window.","feed_headline":"Cause-effect synchronization slices forecast error 9-32%","feed_subtitle":"Pre-training on shifted cause-effect pairs beats raw multivariate training and transfers to unseen causal structures.","key_machinery":"The load-bearing object is the target-oriented causal shift: for each detected cause-effect pair $(Z^{(j)}, X^{(i)})$ with estimated lag $\\delta_{ij}$, the cause series is shifted forward by $\\delta_{ij}$ time steps so that its values align with the effect's present. The shift is produced by a linear VAR Granger causality analysis, a regression-based test of whether past values of one series help predict another, which supplies both the pair selection and the lag. To avoid information loss, the input window stacks the target effect, the shifted cause, and the original unshifted cause, letting the model see both the aligned and the raw series. This construction breaks a complex, possibly disconnected causal graph into many small paired training samples, which the paper argues reduces the noise from causally unrelated or misaligned channels.","core_discovery":"The central claim is that the target-oriented causal shift, moving a cause variable's values forward by its estimated lag $\\delta_{ij}$ before training, converts a misaligned multivariate forecasting problem into a set of aligned cause-effect subproblems, and that this alignment is what lets channel-dependent models exploit cross-variable structure. The paper calls the resulting training scheme a hybrid CI+CD strategy: within each pair, the model sees the effect and both versions of the cause (shifted and unshifted), while across pairs the same model weights are shared, so the network indirectly learns from many cause-effect relationships without being forced to model causally unrelated channels. Experimentally, the claim is supported by consistent MAPE reductions on four architectures and by the transfer result in Table 2, where synchronized pre-training plus fine-tuning reaches 2.21 MAPE versus 2.56 for the non-synchronized counterpart. The authors state that the improvements are heavily dependent on causal pair extraction and lag estimation, and they assume globally stationary causal lags.","pith_inferences":["The paper's own caveat suggests a direct stress test: if estimated lags are even slightly wrong, the shift places irrelevant values in the context; measuring MAPE against lag-estimation error on the same synthetic generator would show how much of the gain depends on perfect lags.","Because the method decomposes the graph into pairs, it ignores the state of other variables in the causal structure; an extension would be to feed the broader graph context or add a reconciliation loss, which the authors list as future work.","The same pairing and shifting idea could be applied to classification or control tasks in digital twins, not just forecasting, since the aligned cause-effect window is task-agnostic.","If paired pre-training transfers to real process data, it may reduce the need for per-plant labeled data, which is the implicit promise of foundation models for the process industry."],"forward_implications":["If synchronization works as claimed, channel-dependent pre-training no longer needs a fixed variable set, because each training sample is a self-contained cause-effect pair; this is what makes cross-dataset pre-training possible.","Causes that occurred before the context window become visible after shifting, so models can exploit highly lagged dependencies (hours or days in process plants) without extending the context.","The hybrid CI+CD construction lets a shared-weight model learn cross-variable dependence within pairs and generalizable cause-effect patterns across pairs, which is the basis of the reported transfer gains.","Synchronized pre-training plus fine-tuning outperforms both non-synchronized pre-training and target-only training on the synthetic benchmark, providing a concrete training recipe to test on real industrial data."],"supporting_citations":[{"why":"Supplies the linear VAR Granger causality procedure used to identify cause-effect pairs and estimate their lags.","marker":"[2]"},{"why":"Defines Granger causality, the conceptual basis for calling one variable a cause of another in the lag detection step.","marker":"[10]"},{"why":"Provides the synthetic data generator used to inject known linear causal structures with lags up to 200 steps, giving ground truth for evaluation.","marker":"[24]"},{"why":"Introduces the lead-lag synchronization idea that the target-oriented causal shift adapts to bring causes outside the context window into it.","marker":"[30]"},{"why":"Supplies the neural forecasting framework used to implement and train the four compared channel-dependent models.","marker":"[21]"},{"why":"One of the four evaluated models (TSMixerx) that shows reduced MAPE under synchronization.","marker":"[4]"},{"why":"One of the four evaluated models (MLPMultivariate) that shows reduced MAPE under synchronization.","marker":"[8]"},{"why":"One of the four evaluated models (BiTCN) that shows reduced MAPE under synchronization.","marker":"[25]"},{"why":"One of the four evaluated models (TFT) that shows reduced MAPE under synchronization.","marker":"[15]"}],"fun_headline_variants":["Lag-shift cause-effect pairs cut multi-dim forecast error by 9-32%","Synchronized cause-effect pre-training beats raw multivariate training","Causal shift alignment yields 9-32% lower MAPE in forecasting","Pre-train on lag-aligned cause-effect pairs for up to 32% error cut","Shift causes by estimated lag to reduce forecast error by 32%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands or falls on the estimated causal lags $\\delta_{ij}$ being accurate and stable over time; if a lag is wrong or drifts, shifting the cause aligns it with irrelevant history, and the synchronized training loses its advantage.","fun_headline_variants_meta":{"raw":{"variants":["Lag-shift cause-effect pairs cut multi-dim forecast error by 9-32%","Synchronized cause-effect pre-training beats raw multivariate training","Causal shift alignment yields 9-32% lower MAPE in forecasting","Pre-train on lag-aligned cause-effect pairs for up to 32% error cut","Shift causes by estimated lag to reduce forecast error by 32%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000326,"raw_usage":{"total_tokens":1813,"prompt_tokens":923,"completion_tokens":890,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":790}},"tokens_in":539,"tokens_out":890,"duration_ms":9303,"temperature":1.0,"reasoning_tokens":790,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:54:30.458921+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same synthetic experiment with oracle lags versus lags perturbed by one to ten time steps; if MAPE degrades to or below the non-synchronized baseline under even small perturbation, the claimed benefit is an artifact of perfect lag estimation. Alternatively, generate data where lags change at known break points; if synchronized training underperforms ordinary multivariate training on non-stationary lags, the global-stationarity assumption fails.","supporting_citations":[{"cited_title":"Automated Knowledge Graph Learning in Industrial Processes","cited_arxiv_id":"2407.02106","evidence_quote":"Supplies the linear VAR Granger causality procedure used to identify cause-effect pairs and estimate their lags."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the synthetic data generator used to inject known linear causal structures with lags up to 200 steps, giving ground truth for evaluation."},{"cited_title":"NeuralForecast: User friendly state-of-the-art neural forecasting models","cited_arxiv_id":null,"evidence_quote":"Supplies the neural forecasting framework used to implement and train the four compared channel-dependent models."},{"cited_title":"SELF-ORGANIZING MULTILAYERED NEURA L NETWORK","cited_arxiv_id":null,"evidence_quote":"One of the four evaluated models (MLPMultivariate) that shows reduced MAPE under synchronization."},{"cited_title":"Parameter Efficient Deep Probabilistic Forecasting","cited_arxiv_id":"2112.02905","evidence_quote":"One of the four evaluated models (BiTCN) that shows reduced MAPE under synchronization."}],"review_version":1}