{"id":"522bb51e-79ce-4b22-807d-cf0360736e6a","arxiv_id":"2412.12168","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A seasonal forecasting model that forecasts fixed ascending and descending phases of each day with linear regression and the middle peak phase with a multi-scale convolutional network, reporting gains on CAISO, Electricity, and Traffic.","lead":"MSSD splits each 24-hour seasonal cycle into ascending, peak, and descending thirds, predicts the first and last with linear regression and the middle with a small convolutional network. The authors report large error reductions over several pre-2022 baselines on three public datasets, but provide no code, no full training protocol, and several claims that do not match their tables.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported results are internally inconsistent: Table II and Table IV give incompatible MSEs for the same datasets, input length, and horizons, so the central empirical claim is not assessable from the manuscript.","rationale":"The paper's central claim is empirical: MSSD substantially outperforms strong baselines on three public seasonal datasets. The single most load-bearing obstacle is not the architectural choice of fixed one-third daily slices, but the fact that the two central result tables contradict each other under the stated protocols. For Traffic at I=96/O=96, Table II gives univariate MSE 0.108 while Table IV gives multivariate MSE 0.471; for Electricity the same settings give 0.061 and 0.149. Since the text says multivariate forecasting is done by independently testing each variable, these values should be reconcilable through a stated normalization and averaging rule. The paper provides no such rule, no code, no hyperparameters, no split definition, and no error bars, so the discrepancy cannot be resolved from the manuscript. This directly undermines the abstract's 10% reduction claim and the strong per-table gains. The reader's verdict of REJECT is appropriate, and this analysis reinforces it without moving it to a different category; if the authors release code and correct the tables, the empirical claim may become testable, so a future revised version could be accepted conditional on that evidence.","tokens_in":8828,"tokens_out":5518,"duration_ms":62098,"concrete_test":"Run the model on Traffic and Electricity with input length 96 and prediction length 96 under the univariate protocol of Table II and the multivariate protocol of Table IV, using released code or a precise written protocol. Report the variable subset, normalization, and averaging rule for each table, and verify whether the two MSE values agree within numerical precision. As a minimal check, recompute the Table II Traffic O=96 entry (MSE 0.108) from the same per-variable predictions used for the Table IV Traffic O=96 entry (MSE 0.471); if the numbers cannot be reconciled, the reported improvements are not reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Even before checking the fixed 24-hour and one-third decomposition assumption, the numerical evidence is self-contradictory. Table II reports univariate MSSD on Traffic at input length 96 and prediction length 96 with MSE 0.108 and MAE 0.263; Table IV reports multivariate MSSD on the same dataset and same horizons with MSE 0.471 and MAE 0.294. For Electricity the same comparison is MSE 0.061 in Table II versus 0.149 in Table IV. The paper states that multivariate results were produced by testing MSSD independently for each variable, so under any fixed normalization and averaging rule the univariate and multivariate per-variable means should agree, or the paper must state which variable subset each table uses. A factor-of-four MSE gap means the two tables cannot both describe the same protocol. Because every headline SOTA claim is drawn from these tables, the central empirical claim is unsupported as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MSSD, a decomposition-based forecasting framework for seasonal time series. For each assumed 24-hour cycle (T = 24·i, Eq. 1), the series is split into fixed ascending, peak, and descending thirds (Eq. 2). Ascending and descending components are forecast with linear regression, while the peak component is modeled by SDNet, a multi-scale convolutional network combining Conv1d local compression and dilated causal convolutions. The paper reports univariate and multivariate results on CAISO, Electricity, and Traffic, claiming roughly 10% error reduction over baseline models, state-of-the-art performance, and robustness to noise.","tokens_in":9036,"tokens_out":3101,"duration_ms":32662,"significance":"If the empirical claims were reliable, MSSD would be a useful contribution: its decomposition is interpretable, the architecture is lightweight, and the ablation studies (Tables V and VI) give some evidence that the SDNet module helps relative to self-attention and auto-correlation alternatives. However, the central empirical claim is not assessable from the manuscript as written. The headline results are drawn from tables that contradict each other for the same datasets and horizons, and the experimental protocol lacks the detail needed for reproduction. The decomposition assumption of fixed 24-hour one-third boundaries is plausible for some data but is not validated. Thus the significance is currently unsupported by the evidence presented.","major_comments":[{"comment":"Table II and Table IV report incompatible results for the same datasets, input length, and prediction length. For Electricity at input length 96 and horizon 96, Table II reports univariate MSSD MSE 0.061 and MAE 0.203, while Table IV reports multivariate MSSD MSE 0.149 and MAE 0.284. For Traffic at the same settings, Table II reports MSE 0.108 and MAE 0.263, while Table IV reports MSE 0.471 and MAE 0.294. Since the text states that multivariate analysis was produced by independently testing each variable, under a fixed normalization and averaging rule the per-variable means should agree (or the variable subset should be stated). The factor-of-four MSE gap for Traffic means the two tables cannot both describe the same experimental protocol, so the SOTA claims based on these tables are unsupported as written.","section":"§III.A, Tables II and IV"},{"comment":"The aggregate percentage improvements reported in the text do not match recomputation from Table II. For Electricity, the MAE reductions relative to MICN are (0.341−0.25)/0.341 ≈ 26.7%, (0.354−0.224)/0.354 ≈ 36.7%, and (0.398−0.203)/0.398 ≈ 49.0%, averaging about 37.5%, not the stated 36.6%. For Traffic, the same computation gives approximately −4.8%, +1.2%, and −9.1%, averaging about −4.2%, not the stated −1.67%. The paper's reported averages appear to be computed from numbers other than those shown in Table II, which undermines confidence in the accuracy of the reported empirical results.","section":"§III.A, text vs Table II"},{"comment":"The decomposition assumes every seasonal cycle has an exact 24-hour period (T = 24·i) and that each cycle is partitioned at fixed contiguous one-third boundaries aligned to the start of the day. This assumption enters directly into the definition of the ascending, peak, and descending components and is never validated against any dataset. If actual peaks are phase-shifted or if the series has weekly or multi-day periodic components, the linear-regression components will be misspecified. The conclusion acknowledges the method is only aimed at periodic data, but the paper does not test whether the fixed boundary assumption holds even on the three datasets used.","section":"§II.A, Eqs. (1)–(3)"},{"comment":"The experimental protocol is critically underspecified. The manuscript does not report train/validation/test splits, normalization procedures, hyperparameters (number of heads l, dilation rate d, kernel sizes, number of layers), optimization settings, or error bars / multiple-seed variability. Without these details, no result in Tables II–IV can be reproduced, and the claimed 10% average error reduction cannot be statistically assessed. The paper also does not provide code, which further limits verifiability.","section":"§III.A, experimental setup"}],"minor_comments":[{"comment":"There are numerous typos and grammatical errors, e.g., \"epvolve\" and \"spuch\" in the Introduction, \"Casual Conv\" in Figures 4 and 8 (should be \"Causal Conv\"), and inconsistent capitalization of \"Traffic\" and \"FEDformer\". The paper would benefit from careful proofreading.","section":"Abstract and Introduction"},{"comment":"The text says \"Table 6\" when referring to the robustness results, but the actual table is labeled Table VII; the numbering of tables in the text should be checked throughout.","section":"§III.B, robustness analysis"},{"comment":"The ablation table reports results for horizons 48, 96, 336, and 720 on all three datasets, but the text in §III.B discusses replacing SDNet with self-attention and auto-correlation in a way that does not clearly state which dataset or which horizon is being described. A clearer explanation of the ablation setup is needed.","section":"§III.B, Table V"},{"comment":"The efficiency analysis is only qualitative: the text claims lower time and memory than self-attention and auto-correlation, but no concrete measurements are given. A table or plot with runtime and memory numbers would be needed to support this claim.","section":"§III.B, efficiency analysis"}],"recommendation":"reject","confidential_remarks":"The manuscript has a load-bearing internal inconsistency in its main experimental tables that cannot be resolved by clarification: Tables II and IV report incompatible numbers for the same datasets, input lengths, and horizons, and the text's aggregate error reductions do not match Table II. The missing experimental details (splits, normalization, hyperparameters, code) would require a full re-run of all experiments. In this state, the central claim is not verifiable, and I do not see how it could be fixed within the scope of a revision short of a complete rewriting of the experimental section. I recommend rejection, but I would be open to reconsidering a substantially revised version with a consistent and reproducible empirical evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the decomposition idea is genuinely clean. Splitting each daily cycle into fixed ascending, peak, and descending thirds, fitting linear trends to the first and last, and using a small convolutional net (SDNet, Conv1d plus dilated causal conv) for the peak is simple, interpretable, and a plausible lightweight alternative to attention for this restricted setting. The paper is also honest about scope, saying in the conclusion that it targets periodic series and will not handle Exchange or Weather well.\n\nThe problem is the evidence. The tables contradict each other. Table II gives univariate Traffic MSE 0.108 at input 96/output 96; Table IV gives multivariate Traffic MSE 0.471 at the same horizons. Electricity is 0.061 versus 0.149. The text says multivariate results were produced by testing MSSD independently for each variable, so under any fixed normalization and averaging the univariate and multivariate per-variable means should agree, or the paper must state a different variable subset. A factor-of-four gap means the two tables cannot both describe the same protocol. The text also claims a 36.6% average MAE reduction on Electricity and a -1.67% reduction on Traffic, both of which are hard to reconcile with the abstract's blanket \"10% reduction\" and with recomputation from Table II. These are not nitpicks; every headline SOTA claim is drawn from these tables.\n\nThe equal-thirds split of a 24-hour cycle is also asserted without validation. It is plausible for hourly electricity and traffic, but the paper never checks whether peaks are actually aligned to the fixed boundaries, and weekly seasonality would break the linear-regression components. That said, this is secondary to the table conflict.\n\nWhat would it take? Release code, a precise experimental protocol, and corrected tables that reconcile the univariate and multivariate numbers. The electricity improvement (MSE from about 0.27 down to 0.09) is so large that artifact-level evidence is mandatory. As written, the central empirical claim is unverifiable and self-contradictory.\n\nA reader wanting a simple baseline for daily-seasonal forecasting might take the decomposition idea and run with it, but should not cite the reported numbers. I would not bring this to reading group and would not send it to a serious referee until the tables and code are fixed. Desk reject, with encouragement to resubmit after major revision.","headline":"Clean decomposition idea, but the reported numbers are internally inconsistent and unverified.","tokens_in":9578,"tokens_out":3361,"would_cite":false,"duration_ms":32333,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Splitting each day into rise, peak, and fall phases produces a lightweight seasonal forecaster that the authors report beats transformer baselines on three public datasets by roughly 10 percent error on average.","keywords":["time-series forecasting","seasonal decomposition","multi-scale convolution","temporal convolutional network","load forecasting","interpretable forecasting","peak modeling","MSSD"],"falsifier":"Run MSSD on a 24-hour-period dataset whose peak occurs near the boundary of the ascending/peak split (e.g., a morning-peak load curve). If MSE rises to baseline levels while a version with learned split points retains the improvement, the fixed one-third boundary is the load-bearing premise; if MSSD still wins, the assumption is not essential.","tokens_in":8568,"feed_emoji":"📈","tokens_out":4902,"duration_ms":43960,"temperature":0.7,"pith_summary":"This paper claims that seasonal time series with a 24-hour period are best forecast by cutting each daily cycle into three fixed phases—ascending, peak, and descending—and modeling each phase separately. The ascending and descending phases are stable enough to fit with plain linear regression, while the peak phase needs a dedicated multi-scale convolutional network (SDNet) that combines local and global features. The authors report that this decomposition, called MSSD, outperforms transformer-based and convolutional baselines on three public datasets (CAISO, Electricity, Traffic) in both short- and long-term forecasting, with about a 10% average error reduction. If right, this matters because it shows a highly interpretable, low-cost model can beat attention-heavy architectures on periodic data.","feed_headline":"Splitting each day into rise, peak, and fall cuts forecast error","feed_subtitle":"Lightweight MSSD reports ~10% error reduction over transformers on electricity, traffic, and CAISO load data.","key_machinery":"The load-bearing object is the fixed three-phase decomposition of each 24-hour cycle (Eq. 2): $T = 24\\cdot i$, with the series split as $x_u = x[0:T/3]$, $x_p = x[T/3:2T/3]$, $x_d = x[2T/3:T]$. This decomposition turns forecasting into three smaller tasks: two linear regressions and one network prediction for the peak fluctuations. The network, SDNet, is built from multi-scale reshaping of the peak segment, local Conv1d compression, and dilated causal convolutions for global context, with a feedforward merge; it replaces self-attention or auto-correlation as the correlation-capturing mechanism.","core_discovery":"MSSD's central claim is that the decomposition $x = x_u + x_p + x_d$, with each 24-hour cycle's first, middle, and last thirds assigned to ascending, peak, and descending components, is a sufficient and beneficial decomposition for seasonal forecasting. The ascending and descending components are modeled by linear regression; the peak component, which carries the most difficult fluctuation patterns, is modeled by SDNet, a multi-branch network using Conv1d, dilated causal convolutions, and multi-scale reshaping to capture local and global correlations. On CAISO, Electricity, and Traffic, MSSD reports lower MSE and MAE than FEDformer, Autoformer, Informer, LogTrans, MICN, LSTM, and LSTNet at most horizons, with especially large gains on Electricity (e.g., MSE 0.087 vs 0.267 at horizon 24). The authors interpret these results as evidence that explicit phase decomposition plus targeted network design for the peak is superior to generic attention or convolution on seasonal data.","pith_inferences":["The fixed one-third split assumes the daily peak lies in the middle eight hours of the day (e.g., roughly 08:00–16:00); real-world load curves with evening or morning peaks would violate this, so an adaptive phase-boundary search is a natural extension the paper does not explore.","The robustness analysis (Table VII) shows error rising steadily with injected noise, with MSE roughly doubling at 20% perturbation; one could test whether the linear-regression components or SDNet is the weaker link under noise.","The paper's scope is explicitly limited to periodic data; applying MSSD to exchange or weather series, which it concedes it does not handle, would test whether the decomposition generalizes beyond strict 24-hour seasonality.","The ablation replacing SDNet with self-attention suggests the convolutional peak module, not the decomposition alone, drives much of the reported gain; a version with the fixed decomposition but a stronger attention block could isolate this further."],"forward_implications":["If the fixed-third decomposition holds, the ascending and descending parts of a seasonal cycle can be modeled without any learned nonlinearity, drastically cutting parameter count and computational cost.","The peak segment becomes the only component needing a neural network, so the model's accuracy hinges on how well SDNet (or an equivalent) captures peak fluctuation patterns.","On the reported datasets, longer input lengths help MSSD more consistently than transformer baselines, suggesting the model exploits long-term periodic structure rather than short-term recurrence.","The reported gains on Electricity (MSE from 0.267 to 0.087 at horizon 24) imply that for strongly periodic load and traffic data, phase-aware decomposition can yield substantial reductions in squared error."],"supporting_citations":[{"why":"Supplies the premise that the standard period of seasonal cycle time series is 24 hours, which grounds the fixed T = 24·i decomposition.","marker":"[14]"},{"why":"Provides the basis for using one-dimensional convolution-based decomposition of time series, which the module extends.","marker":"[13]"},{"why":"MICN is the strongest convolutional baseline and the target of the ablation where its convolution module is replaced by SDNet.","marker":"[12]"},{"why":"FEDformer is the best-performing transformer baseline that MSSD must beat on all three datasets.","marker":"[15]"},{"why":"Autoformer is a decomposition transformer baseline whose auto-correlation mechanism is compared against SDNet.","marker":"[10]"},{"why":"Informer serves as an efficient-transformer baseline for long-sequence forecasting comparisons.","marker":"[9]"},{"why":"LogTrans is a transformer baseline included in the univariate and long-term comparisons.","marker":"[16]"},{"why":"LSTM is an RNN baseline included in the multivariate comparisons.","marker":"[5]"},{"why":"LSTNet is an RNN baseline with convolutional and recurrent components included in the multivariate comparisons.","marker":"[8]"}],"fun_headline_variants":["Rise-peak-fall split cuts forecast error 10%","Peak-focused daily split trims error by 10%","Three-phase day decomposition cuts forecast error","Daily rise-peak-fall split improves forecast accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every seasonal cycle is exactly 24 hours long and the daily peak always falls in the middle third of the day, so the fixed boundaries in Eq. (2) never require adjustment.","fun_headline_variants_meta":{"raw":{"variants":["Rise-peak-fall split cuts forecast error 10%","Peak-focused daily split trims error by 10%","Three-phase day decomposition cuts forecast error","Daily rise-peak-fall split improves forecast accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000886,"raw_usage":{"total_tokens":3816,"prompt_tokens":926,"completion_tokens":2890,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":2825}},"tokens_in":542,"tokens_out":2890,"duration_ms":20646,"temperature":1.0,"reasoning_tokens":2825,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:29:34.629582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MSSD on a 24-hour-period dataset whose peak occurs near the boundary of the ascending/peak split (e.g., a morning-peak load curve). If MSE rises to baseline levels while a version with learned split points retains the improvement, the fixed one-third boundary is the load-bearing premise; if MSSD still wins, the assumption is not essential.","supporting_citations":[{"cited_title":"Analysis and modeling of seasonal time series,","cited_arxiv_id":null,"evidence_quote":"Supplies the premise that the standard period of seasonal cycle time series is 24 hours, which grounds the fixed T = 24·i decomposition."},{"cited_title":"Micn: Multi-scale local and global con- text modeling for long-term series forecasting,","cited_arxiv_id":null,"evidence_quote":"MICN is the strongest convolutional baseline and the target of the ablation where its convolution module is replaced by SDNet."},{"cited_title":"Informer: Beyond efficient transformer for long sequence time-series forecasting,","cited_arxiv_id":null,"evidence_quote":"Informer serves as an efficient-transformer baseline for long-sequence forecasting comparisons."},{"cited_title":"Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting,","cited_arxiv_id":null,"evidence_quote":"LogTrans is a transformer baseline included in the univariate and long-term comparisons."},{"cited_title":"Modeling long-and short-term temporal patterns with deep neural networks,","cited_arxiv_id":null,"evidence_quote":"LSTNet is an RNN baseline with convolutional and recurrent components included in the multivariate comparisons."}],"review_version":1}