{"id":"4227e015-67e7-45e5-b22f-80ad895264b9","arxiv_id":"2411.11340","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A dynamic global plus component loss weighting modestly improves some decomposition-based forecasters, but the paper overstates its results and mischaracterizes PatchTST.","lead":"The paper proposes a hybrid loss for time series forecasting that combines the overall forecast error with errors on seasonal and trend sub-series, and adjusts their weights during training. It claims 0.5-2% average improvement over existing methods without changing model architectures, but the evidence is mixed and one key baseline is mischaracterized.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PatchTST does not natively decompose into seasonal/trend sub-series, so the claimed architecture-agnostic hybrid loss appears undefined for one of the three main baselines.","rationale":"The reader's weakest assumption is well placed and is the most load-bearing issue I can identify. The paper's contribution is explicitly a loss-only change: the abstract promises improvement 'without any modifications to the model architectures,' and Section 4.1 states the baseline's original loss is directly replaced. For that promise to hold, the seasonal and trend component losses must be computable from each baseline model's existing outputs. For PatchTST, this condition is not met in the published architecture: PatchTST operates on patched raw series and has no native seasonal/trend decomposition. Yet Section 2 groups PatchTST with DLinear and FEDformer as methods that 'all employ sliding-window-based time series decomposition,' and Tables 1 and 3 report sub-series losses for PatchTST. The paper never explains how these sub-series outputs are generated for PatchTST. Either the authors added a decomposition module and separate forecasting heads, which falsifies the architecture-agnostic claim, or the numbers are not reproducible from the official model. This is a clean, checkable contradiction rather than a matter of interpretation. Even ignoring this issue, Table 2 shows the average improvement is fragile, with degradations on Electricity and Exchange for FEDformer and PatchTST and with the overall mean dominated by the illness dataset. But the PatchTST mismatch is the decisive weakness because it attacks the core claim directly. A diff against the official repository plus an attempt to compute component losses from the unmodified model would settle the matter, and the outcome would determine whether the headline claim has any valid experimental support.","tokens_in":18560,"tokens_out":5019,"duration_ms":51271,"concrete_test":"Obtain the supplementary code and diff the PatchTST model definition against the official yuqinie98/PatchTST repository. If a seasonal-trend decomposition module and separate sub-series heads appear, the 'no architecture modification' claim is false; if they do not appear, the reported PatchTST Seasonal/Trend losses cannot be computed. Independently, run the official PatchTST on ETTh1 with prediction length 96 and attempt to evaluate Eq. (8) using only native model outputs; if component losses are undefined, the reported PatchTST results cannot be reproduced without changing the model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that every baseline in Tables 1-3 natively produces seasonal and trend forecasts, so that replacing the final MSE loss with Eq. (8) needs no network changes. DLinear and FEDformer roughly satisfy this, but PatchTST does not: the published PatchTST architecture (Nie et al., 2022) applies instance normalization and patching to the raw series and outputs a single prediction, with no seasonal-trend decomposition and no component-level outputs. Section 2 nonetheless states 'These methods all employ sliding-window-based time series decomposition', and Tables 1 and 3 report Seasonal and Trend MSE for 'Patchtst'. Since Section 4.1 says the method 'directly replaces the original loss function', the only way to compute Loss_S and Loss_T for PatchTST is to insert a decomposition layer plus separate sub-series prediction heads. That is an architectural modification, contradicting the abstract's 'without any modifications to the model architectures.' Alternatively, the PatchTST rows are not reproducible from the official model. Because PatchTST is one of only three main baselines and contributes heavily to the reported mean improvement, this single mismatch directly undermines the paper's headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that end-to-end overall losses used by decomposition-based time series forecasting methods introduce a bias that limits accuracy, and it proposes a hybrid loss framework that combines the global loss with seasonal/trend component losses through a dual min-max weighting scheme. The method is applied to DLinear, FEDformer, and PatchTST on eight benchmark datasets, and the authors report an average improvement of 0.5-2% without architectural modifications, with additional ablation studies on weight initialization and loss components.","tokens_in":18752,"tokens_out":3986,"duration_ms":38771,"significance":"If the central claim held, this would be a useful plug-and-play loss modification for decomposition-based forecasting, notable for adding a principled dynamic weighting mechanism inspired by DRO. The paper also provides ablations that isolate the contribution of the component loss and the dynamic weights, and it releases code. However, the empirical support is substantially weakened by a mismatch between the method and one of its three main baselines, and by internal numerical inconsistencies in the reported tables.","major_comments":[{"comment":"The statement 'These methods all employ sliding-window-based time series decomposition' is not true for PatchTST. The published PatchTST architecture (Nie et al., 2022) applies instance normalization and patching directly to the raw series and produces one prediction; it has no seasonal-trend decomposition and no component-level outputs. Therefore the component losses Loss_S and Loss_T used in Equations (2) and (8) cannot be computed unless a decomposition layer is inserted into the network. This contradicts Section 4.1, which says the method 'directly replaces the original loss function', and the abstract, which claims improvement 'without any modifications to the model architectures.' Since PatchTST is one of only three main baselines, this mismatch directly undermines the paper's headline claim.","section":"Section 2, Tables 1 and 3"},{"comment":"There is a direct numerical contradiction in the main results. Table 2 reports PatchTST on ETTh2 with hybrid loss as MSE 0.3639 and MAE 0.3929, while Table 3 reports the same configuration as MSE 0.3699 and MAE 0.3989; both are stated to be averages over the same prediction horizons {96, 192, 336, 720}. Since Table 3 is the table that supports the component-level improvement story, this inconsistency makes it impossible to determine the actual effect of the proposed loss for PatchTST on this dataset.","section":"Tables 2 and 3 (PatchTST/ETTh2)"},{"comment":"The claims of 'state-of-the-art performance across a wide range of datasets' and an 'average improvement of 0.5-2%' are not consistent with the full set of reported results. In Table 8, hybrid-loss variants achieve the best result on only 3 of 8 datasets, tying with the LLM baselines, and Table 2 contains clear degradations, for example FEDformer on Electricity and Exchange and PatchTST on Electricity, Exchange, and Weather. In addition, no error bars or statistical significance tests are provided, so improvements that are often below 0.3% in MSE cannot be distinguished from noise.","section":"Abstract and Section 6 (Conclusion)"}],"minor_comments":[{"comment":"In Equation (5), the denominator is written as 'w1 exp(λ1LossG) + wpre_2 exp(λ1LossC)', which is missing the superscript on w1; it should presumably read 'wpre_1'.","section":"Equations (4)-(7)"},{"comment":"The text says 'Our method directly replaces the original loss function', but for PatchTST this is not possible without also changing the network to produce seasonal and trend predictions. Please clarify whether an auxiliary decomposition module was added.","section":"Section 4.1"},{"comment":"For DLinear, the row with w1=0.9 and α=0.9 reports values identical to the row with w1=0.9 and α=0.1 in all four datasets, which suggests a copy-paste error in the ablation table.","section":"Table 5"},{"comment":"The reproducibility statement says the code is in the Supplementary Material, but the arXiv submission does not appear to include a link or attachment; please provide a permanent link for the camera-ready version.","section":"Reproducibility"}],"recommendation":"reject","confidential_remarks":"The PatchTST issue is the decisive one: because PatchTST does not decompose the series, the central claim of architecture-agnostic loss replacement is not substantiated. The numerical inconsistency between Tables 2 and 3 compounds this concern. The core idea is interesting and could be salvageable if the authors restrict the claim to genuinely decomposition-based models and re-run the experiments; as submitted, however, the empirical evidence does not support the paper's main assertions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. The paper is worth a glance because it documents something real: in DLinear, FEDformer, and their PatchTST runs, the trend sub-series error is consistently several times the seasonal error, and that imbalance tracks the overall error. That is a concrete, potentially actionable observation about how end-to-end losses distribute learning across decomposition components. The proposed fix—add per-component losses and dynamically weight them with a dual min-max scheme—is a standard DRO-style weighting, but applying it to this setting is reasonable, and the ablation says dynamic weighting beats fixed weights. So the core idea is plausible.\n\nThe problem is the headline claim: 'without any modifications to the model architectures.' The paper asserts PatchTST uses sliding-window seasonal-trend decomposition. The published PatchTST does not; it applies instance normalization and patching to the raw series and outputs a single prediction. To compute Loss_S and Loss_T for PatchTST, the authors must have inserted a decomposition layer and likely separate prediction heads. That is an architectural change, and it turns the central claim from 'plug-in loss' into 'modified model.' This is not a nitpick—PatchTST is one of three baselines and drives much of the reported average improvement.\n\nThere are also proportionality issues. The average gain is around 0.7%, not a clean 0.5–2%; several datasets (FEDformer on Electricity and Exchange, PatchTST on Weather) get worse. No error bars or multiple seeds are given, and the authors chose λ1 and λ2 on the validation set. The 'state-of-the-art' claim in the conclusion is contradicted by their own Table 8, where hybrid models beat all others on only three of eight datasets. Code is promised in the supplementary material but not visible in the preprint.\n\nWho is this for? People who use DLinear/FEDformer and want a small extra margin might try the loss. Researchers studying loss design for decomposition-based forecasting will find the imbalance observation useful. But as submitted, the evidence does not support the paper's abstract, and the PatchTST mismatch needs to be fixed or removed.\n\nI would not desk reject this—the experiment is cheap to redo and the idea is testable. Send it to review with a clear request for code, seeds, and an honest treatment of what changed for PatchTST.","headline":"Useful observation about sub-series loss imbalance, but the central 'no architecture changes' claim collapses over PatchTST, and the evidence is too thin to support the advertised gains.","tokens_in":19312,"tokens_out":2530,"would_cite":false,"duration_ms":24037,"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":"The paper claims that the standard end-to-end loss used to train decomposition-based time series forecasters is biased, letting the trend sub-series be learned far worse than the seasonal one, and that a hybrid loss dynamically weighting…","keywords":["time series forecasting","hybrid loss","seasonal-trend decomposition","distributionally robust optimization","dual min-max","component loss","deep learning forecasting"],"falsifier":"Run the official PatchTST code with its default settings and inspect whether the model decomposes the input into seasonal and trend sub-series before applying the hybrid loss; if no such decomposition exists (or one must be added), the claim that the loss works with no architectural modification is falsified. A second check: apply the hybrid loss to a dataset with no clear trend and verify whether the reported gains disappear, as the mechanism predicts the trend-weighting term is what drives the improvement.","tokens_in":18307,"feed_emoji":"📈","tokens_out":6365,"duration_ms":55155,"temperature":0.7,"pith_summary":"The paper argues that the single end-to-end loss used to train decomposition-based time series forecasters is not a neutral objective: it lets the trend sub-series be learned far worse than the seasonal sub-series, and that imbalance is a major source of forecast error. To test this, the authors measure component-level MSEs for DLinear, FEDformer, and PatchTST and find the trend error is often two to five times the seasonal error. They then propose a hybrid loss that adds seasonal and trend component losses to the global loss, with a dual min-max weighting rule that dynamically up-weights the larger error. On eight datasets the framework improves average performance by 0.5–2% without changing the underlying model architectures, with the largest gain on the illness dataset. The claim matters because it suggests a substantial accuracy gain is available purely from how the model is trained.","feed_headline":"Hybrid loss lifts time-series forecast accuracy by 0.5–2%","feed_subtitle":"A dual min-max blend of global and seasonal/trend losses cuts forecast error without changing model architectures.","key_machinery":"The central object is the hybrid loss $L = w_1 L_G + w_2(\\alpha L_S + \\beta L_T)$ with a dual min-max formulation. The outer objective $\\min_\\theta \\max_{w_1+w_2=1} w_1 L_G + w_2 L_C$ balances the global forecast error against the summed component error, and the inner objective $\\min_\\theta \\max_{\\alpha+\\beta=1} \\alpha L_S + \\beta L_T$ balances seasonal against trend error. Weights are updated by the mirror-descent-style rule $w_i^\\text{cur} \\propto w_i^\\text{pre} \\exp(\\lambda\\, L_i)$, so the loss that is currently larger receives more weight. This mechanism is what forces the model to attend to the weaker sub-series while still minimizing the global error.","core_discovery":"The central claim is that overall loss introduces a bias that hinders the learning of significant sub-series, particularly the trend, and that restoring balance through component-level supervision improves forecasts. Concretely, the paper reports that under the original MSE loss the seasonal and trend sub-series are learned unevenly across datasets, and the worse sub-series accounts for roughly 80% of the overall error. The proposed remedy is a hybrid loss $L = w_1 L_G + w_2(\\alpha L_S + \\beta L_T)$, where $L_G$ is the global MSE and $L_S$, $L_T$ are MSEs on the seasonal and trend sub-series. The weights come from a dual min-max problem: an outer max over the global versus component loss, and an inner max over seasonal versus trend, solved by mirror-descent-style exponential updates. The paper's evidence is that this weighting, integrated as a drop-in replacement for the loss, reduces both component and overall errors across most datasets, with dynamic weighting necessary because fixed-weight and component-only variants perform worse.","pith_inferences":["If the diagnosis generalizes, the same dual-loss idea could be applied to other decompositions (frequency bands, multi-scale components) where a single global metric hides an imbalanced component.","The 'no architecture modification' claim is narrower than it appears: computing component losses requires the model to expose seasonal and trend outputs, which the published PatchTST architecture does not do without an added decomposition branch.","The learned weight trajectories could double as a diagnostic: a persistently high weight on trend flags datasets where the global loss is hiding a systematic slope error.","A direct stress test of the mechanism would be to apply the loss to a model whose decomposition is intentionally removed; the framework should lose its advantage if the component losses are the active ingredient."],"forward_implications":["Replacing the loss function is enough: DLinear, FEDformer, and PatchTST improve by 0.5–2% on average with no change to their architectures.","The trend sub-series improves most, so the hybrid loss specifically fixes forecasts that miss the overall slope of the series.","Dynamic weights are essential: using only component losses or fixed weights consistently underperforms the adaptive scheme.","The framework is a plug-and-play training module, so it can be layered onto future decomposition-based backbones as they appear.","The benefit is not universal: on Electricity and Exchange, whose series have abrupt changes, the hybrid loss can slightly worsen results, indicating the mechanism suits data with a clear trend-seasonal structure."],"supporting_citations":[{"why":"Supplies the DLinear baseline and the seasonal-trend decomposition outputs the component losses are computed on.","marker":"(Zeng et al., 2023)"},{"why":"Supplies the FEDformer baseline, a transformer-based decomposition model tested with the hybrid loss.","marker":"(Zhou et al., 2022)"},{"why":"Supplies the PatchTST baseline; the paper's central claim depends on it performing sliding-window decomposition.","marker":"(Nie et al., 2022)"},{"why":"Provides the variance-based regularization and convergence argument cited to justify the mirror-descent weighting.","marker":"(Duchi & Namkoong, 2019)"},{"why":"Introduces the stochastic gradient DRO method the dual min-max formulation is modeled on.","marker":"(Namkoong & Duchi, 2016)"},{"why":"Grounds the distributionally robust optimization perspective used to motivate the min-max objective.","marker":"(Wiesemann et al., 2014)"},{"why":"Defines the sliding-window seasonal-trend decomposition that produces the sub-series.","marker":"(Faltermeier et al., 2010)"},{"why":"Provides the seasonal-trend decomposition framework underlying the sub-series losses.","marker":"(Cleveland et al., 1990)"}],"fun_headline_variants":["Hybrid loss cuts forecast error by up to 2%","Hybrid loss fixes component bias in forecasting","Dual min-max loss balances global and component errors","Component-aware loss lifts time-series forecast accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every baseline model exposes a seasonal and a trend sub-series during training, so the component losses can be added without altering the model; if that decomposition is not actually present, the method changes the architecture rather than only the loss.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid loss cuts forecast error by up to 2%","Hybrid loss fixes component bias in forecasting","Dual min-max loss balances global and component errors","Component-aware loss lifts time-series forecast accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000986,"raw_usage":{"total_tokens":4206,"prompt_tokens":990,"completion_tokens":3216,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":3155}},"tokens_in":606,"tokens_out":3216,"duration_ms":22317,"temperature":1.0,"reasoning_tokens":3155,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:38:06.056138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the official PatchTST code with its default settings and inspect whether the model decomposes the input into seasonal and trend sub-series before applying the hybrid loss; if no such decomposition exists (or one must be added), the claim that the loss works with no architectural modification is falsified. A second check: apply the hybrid loss to a dataset with no clear trend and verify whether the reported gains disappear, as the mechanism predicts the trend-weighting term is what drives the improvement.","supporting_citations":[{"cited_title":"For nation- illness dataset, the input length is 104 and prediction lengths are {24, 36, 48, 60 }, respectively","cited_arxiv_id":null,"evidence_quote":"Supplies the DLinear baseline and the seasonal-trend decomposition outputs the component losses are computed on."},{"cited_title":"Sliding empirical mode decomposition","cited_arxiv_id":null,"evidence_quote":"Defines the sliding-window seasonal-trend decomposition that produces the sub-series."}],"review_version":1}