{"id":"a3fdb85f-58e9-4862-80bb-7f18b5b4e802","arxiv_id":"2412.17323","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"xPatch, a dual-stream CNN/MLP forecaster with exponential moving average decomposition, patches, and channel independence, reports lower MSE/MAE than CARD and PatchTST on most of nine LTSF benchmarks.","lead":"The paper introduces xPatch, a time series forecasting model that separates data into trend and seasonality using exponential moving averages, then processes each with a linear and a convolutional stream. It reports improved accuracy over strong baselines on several standard benchmarks while being cheaper than transformer models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparability is unverified: Appendix K shows most baselines are quoted from other papers while Section 4 implies they were rerun, so the 2-8% SOTA margins need same-pipeline reproduction.","rationale":"I agree with the reader that cross-source baseline comparability is the weakest load-bearing assumption. The contradiction between Section 4's claim that all implementations were derived from official code and Appendix K's detailed sourcing statement makes this more than a generic fairness concern. The reported margins over CARD (2.46% MSE) are small enough to be affected by seed noise, evaluation details, or the extra components xPatch uses, and Appendix J shows the RevIN component alone moves xPatch by 8.67% MSE. Table 15 using TFB baselines is a genuine partial mitigation, and the public code plus thorough ablations are real strengths, but the headline empirical claim still needs a same-pipeline reproduction of the main baselines before it should be treated as robust. The reader's CONDITIONAL verdict therefore remains appropriate without adjustment.","tokens_in":45468,"tokens_out":16252,"duration_ms":159790,"concrete_test":"Use the public TFB harness to rerun xPatch and the seven externally sourced baselines (CARD, TimeMixer, iTransformer, MICN, DLinear, TimesNet, ETSformer) with identical splits, lookbacks, epochs, seeds (at least 5), RevIN setting, and drop-last rule, reporting per-seed means and standard deviations. If any quoted baseline shifts by more than the claimed 2-8% margins, or if xPatch's ranking changes, the SOTA claim is not established; if rankings are stable under this standardized rerun, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical superiority (Section 4, 'Results'). It requires that all baseline numbers are produced under the same protocol as xPatch. Appendix K states that DLinear, TimesNet, and ETSformer results are collected from the TimesNet official repository, and CARD, TimeMixer, iTransformer, and MICN results from their official papers; only RLinear and PatchTST were reproduced by the authors. Yet Section 4 says 'All implementations are derived from the models' official repository code'. These statements conflict for the quoted baselines. Because xPatch additionally uses RevIN, the arctangent loss, and the sigmoid scheduler, and Appendix J reports that RevIN alone changes xPatch results by 8.67% in MSE (larger than the claimed 2.46% margin over CARD), any protocol mismatch in the sourced baselines could erase or invert the reported improvements. Table 15 partially mitigates this by using TFB-repository baselines, but those baselines are still not rerun by the authors under the exact xPatch evaluation pipeline. The SOTA-level claim is therefore conditional on an unverified comparability assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"xPatch is a non-transformer architecture for long-term time series forecasting that decomposes each univariate series via a fixed exponential moving average filter into trend and seasonal components, processes the trend with an MLP stream and the seasonality with a patched depthwise-separable CNN stream, and combines them with a final linear layer. Training uses an arctangent-scaled MAE loss and a sigmoid learning-rate schedule. The paper reports experiments on nine datasets under three protocols (unified lookback, per-model hyperparameter search, and a no-drop-last 'fair' protocol) and compares against eleven baselines, claiming state-of-the-art-level accuracy with lower cost than transformer competitors. Extensive ablations in Appendices E-H and J examine the contribution of each component.","tokens_in":45699,"tokens_out":10327,"duration_ms":88303,"significance":"If the reported results are reproducible and the baselines are genuinely obtained under comparable protocols, xPatch is a meaningful contribution: it shows that patching and channel-independence, previously highlighted in transformer models (PatchTST, CARD), transfer to a strictly convolutional/MLP architecture, and it provides broad ablations with public code. The claimed 2-8% improvements over CARD, TimeMixer, and PatchTST would make xPatch a competitive non-transformer baseline. However, the significance is conditional on baseline comparability and statistical reliability, which are not fully established in this version.","major_comments":[{"comment":"Section 4 states that 'All implementations are derived from the models' official repository code, maintaining the same configurations,' but Appendix K says that DLinear, TimesNet, ETSformer, FEDformer, and Autoformer results are collected from the TimesNet official repository, and CARD, TimeMixer, iTransformer, and MICN results are collected from their respective official papers; only RLinear and PatchTST were reproduced by the authors. These statements conflict for the quoted baselines. Because the headline claim of superiority over CARD (2.46% MSE, Table 13) depends on numbers not produced under the authors' pipeline, and because Appendix J shows that RevIN alone changes xPatch MSE by 8.67%, a small protocol mismatch (lookback length, training epochs, data splits, or normalization) could erase or invert the reported margins. The authors should either rerun all baselines under an identical protocol, or restrict all 'surpasses' claims to rerun baselines and clearly label quoted results as non-comparable.","section":"§4 and Appendix K"},{"comment":"All main tables report averages over three random seeds but no standard deviations, per-seed values, or significance tests. Many of the reported advantages are within a few percent (e.g., Table 13 ETTh1: xPatch MSE 0.428 vs TimeMixer 0.447; Table 14 ETTm1: 0.341 vs 0.350), and in some settings xPatch is worse than the same baselines (e.g., Table 13 Traffic and Electricity MSE vs CARD, Exchange vs MICN, Solar vs TimeMixer). Without variance information it is impossible to assess whether the claimed improvements exceed seed noise. Please add error bars or per-seed results for the central comparisons.","section":"Tables 13-15"},{"comment":"The 'fair' protocol without the drop-last trick (Table 15) is the most convincing setup, but it does not include CARD, the main competitor in the headline claim, and its baselines are quoted from the TFB repository rather than rerun in the authors' environment. Thus the comparison with CARD is only shown under the two less-controlled protocols, which is exactly where protocol mismatch is most likely. For the SOTA claim to stand, the fair protocol should include reruns of at least CARD, TimeMixer, and PatchTST under the authors' exact data split, lookback search, and evaluation code.","section":"Table 15 and Appendix K"},{"comment":"The statement that xPatch 'surpasses CARD by 2.46% in MSE' is based on an average of per-dataset relative differences, computed as 100*(CARD - xPatch)/CARD per dataset and then averaged. This aggregate is dominated by ILI (24.7%) and ETTh2 (13.3%), while xPatch is worse than CARD on four of the ten datasets (Traffic, Electricity, Exchange, Solar). Reporting this single number without per-dataset breakdowns or a significance measure overstates the consistency of the advantage. Please either report the distribution of per-dataset differences or use a more informative summary (e.g., median relative improvement, win/tie/loss counts against each competitor).","section":"§4 (Results) and Table 13"}],"minor_comments":[{"comment":"The claim that the EMA decomposition module was 'optimized ... to O(1) time complexity' is incorrect: the dot product in Eq. (28) is O(L) for a window of length L. What is achieved is a vectorized computation with precomputed weights, not constant time.","section":"Appendix D"},{"comment":"The SMA definition writes 'st = (xt+...+xt+k-1)/k' but then defines XT = AvgPool(Padding(X)) without showing the padding operation in the equation; please clarify which side is padded, with what values, and how st relates to the pooled output.","section":"§3.1, Eq. (1)"},{"comment":"The MAE value 0.287 for DLinear* at horizon 192 appears implausibly low compared to the value 0.389 for the original DLinear and is inconsistent with the other horizons in the same block; please check whether this is a typo.","section":"Table 7, ETTm1, T=192, DLinear* row"},{"comment":"The caption states the comparison is 'between the baselines and the models with RevIN instance normalization,' while the text says the comparison is 'between the original state-of-the-art models and versions trained without using RevIN'; please make the direction and the meaning of the asterisk unambiguous.","section":"Appendix J, Table 12 caption"},{"comment":"The notation 'S (max 25) ↓' is unclear; please define what the downward arrow and 'S' refer to (presumably the number of stationary chunks) and explain the table's reading order.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"The authors should be asked to clarify the provenance of every number in Tables 13-15 and to provide the exact commands and configurations used for the rerun baselines. If rerunning all baselines is infeasible, the claims should be scaled back to comparisons against rerun baselines, with quoted baselines clearly marked as not directly comparable. The current Section 4 overstates what is verified. I also note that the 'fair' protocol omits the main competitor, which weakens the SOTA claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid empirical paper, but the headline “better than CARD/PatchTST” claim is not as robust as the writing implies. The architecture is a known dual-stream linear/nonlinear recipe with an EMA-based decomposition module swapped in for the usual SMA, plus patching and channel independence borrowed from PatchTST. What is actually new is narrow: the specific combination, the arctangent loss, and the sigmoid LR schedule. The paper earns credit for being honest about the ablations — EMA on other models, loss and LR comparisons, RevIN effects, and a fair-protocol table without the drop-last trick — and the code is public. The EMA decomposition itself is a genuinely nice tweak: it handles spiky traffic data more sensibly than SMA, and the ADF stationarity analysis in Appendix C is a good check.\n\nNow the soft spots, in proportion to how much they matter. The central empirical claim of SOTA-level superiority depends on baseline numbers that were not produced under the authors' own pipeline. Appendix K says DLinear, TimesNet, ETSformer, FEDformer, and Autoformer results come from the TimesNet repository, and CARD, TimeMixer, iTransformer, and MICN come from their official papers; only RLinear and PatchTST were actually reproduced by the authors. Yet Section 4 says all implementations are derived from official code. That is a real inconsistency. The margins over CARD are 2.46% MSE under unified settings and 5.29% under hyperparameter search, but on Traffic, Electricity, and Exchange, xPatch is actually worse than CARD. The paper reports averages over three seeds but gives no error bars, so we cannot tell whether the 2% gains are noise. The RevIN ablation shows xPatch improves 8.67% with RevIN — larger than the entire claimed margin over CARD — which means any protocol mismatch in how baselines were trained or normalized could flip the ranking. The TFB fair-protocol table (Table 15) partially mitigates this, since baselines there come from a standardized repository, but those baselines are still quoted, not rerun under the exact xPatch evaluation loop.\n\nThat said, the core message survives the baseline-comparability critique: a non-transformer with patching and channel independence can be competitive with transformer-based models on standard benchmarks at lower inference cost. That is a useful and mostly credible result for the applied forecasting community, even if the exact ordering against CARD is provisional. The paper deserves a serious referee. I would send it out, with the explicit request that the authors provide seed-wise variance and either rerun the key baselines (at least CARD, PatchTST, TimeMixer) in their own pipeline or clearly mark which numbers are quoted and justify comparability. The architecture and ablations are worth keeping; the SOTA wording needs to be tempered to “competitive on many datasets, not uniformly better.”","headline":"A credible non-transformer forecaster whose SOTA claim rests on baseline numbers the authors quoted rather than reran; worth refereeing, but the margins are thinner and more conditional than the abstract suggests.","tokens_in":46259,"tokens_out":1519,"would_cite":true,"duration_ms":19070,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a non-transformer, dual-stream CNN/MLP model called xPatch can match or beat transformer-based long-term forecasting models by combining exponential moving average decomposition with patching and channel…","keywords":["time series forecasting","exponential moving average decomposition","patching","channel independence","dual-stream architecture","CNN","arctangent loss","learning rate schedule"],"falsifier":"Run xPatch and each main baseline (at least CARD, PatchTST, TimeMixer, and iTransformer) in a single codebase on the same GPU, using the same data splits, lookback lengths, prediction horizons, seed protocol, and number of epochs, and check whether the reported margins of roughly 2-8% in MSE and MAE persist; if the margins shrink below noise or reverse, the empirical superiority claim fails.","tokens_in":45236,"feed_emoji":"📈","tokens_out":5779,"duration_ms":50789,"temperature":0.7,"pith_summary":"The paper tries to establish that a non-transformer model can match or beat transformer-based long-term forecasting models by borrowing their best ideas, namely patching and channel independence, and combining them with exponential moving average decomposition. It introduces xPatch, a dual-stream network in which an MLP stream handles the trend and a CNN stream handles seasonality, and reports that under a unified evaluation protocol xPatch achieves the best average MSE on 60% of datasets and best average MAE on 70%, beating the transformer baseline CARD by about 2.5% in both metrics. The authors also claim that their arctangent loss and sigmoid learning-rate schedule improve several existing forecasters when swapped into them. If correct, this matters because it suggests that recent gains of transformer forecasters come substantially from patching and channel independence rather than from attention itself, and that cheaper CNN/MLP architectures can reach similar accuracy.","feed_headline":"CNN+MLP forecaster beats transformer rivals on most benchmarks","feed_subtitle":"Exponential decomposition plus patching gives xPatch better average error on 60-90% of datasets at a fraction of transformer cost.","key_machinery":"The central object is the exponential seasonal-trend decomposition: with smoothing factor $\\alpha$, the trend is $s_t = \\alpha x_t + (1-\\alpha) s_{t-1}$ and the seasonal component is the residual $x_t - s_t$, computed in $O(1)$ time as a dot product with geometric weights. This replaces the simple moving average decomposition used by several prior models; the trend is sent through a linear MLP stream, while the seasonality is sent through a CNN stream that patches the series into windows of length $P=16$ with stride $S=8$ and applies depthwise separable convolutions with patch-sized kernels. The arctangent loss, with scaling coefficient $\\rho(i) = -\\arctan(i) + \\frac{\\pi}{4} + 1$, and a two-sigmoid learning-rate schedule are auxiliary mechanisms claimed to stabilize training and improve accuracy.","core_discovery":"xPatch claims to be the top-performing forecaster among ten compared models on standard long-term forecasting benchmarks, averaged over four prediction horizons, under both a fixed unified protocol and a hyperparameter-search protocol. The empirical anchor is the averaged tables: in unified settings it wins best MSE on 60% of datasets and best MAE on 70%; against CARD it improves by 2.46% in MSE and 2.34% in MAE, and against PatchTST by 4.76% in MSE and 6.20% in MAE. The improvement is attributed to decomposing each univariate series with exponential moving average into trend and seasonal parts, routing them through separate linear and non-linear streams, and training with a robust arctangent loss and a sigmoid learning-rate schedule.","pith_inferences":["If the baseline-comparability caveat is resolved, the result strengthens the case that permutation-invariant attention is a structural disadvantage for long continuous sequences, since a CNN/MLP model using the same patching and channel-independence recipe matches or beats transformers.","A natural testable extension is to make the smoothing factor $\\alpha$ learnable per channel or per dataset; the paper fixes $\\alpha = 0.3$ for dual-stream models and $\\alpha = 0.1$ for transformer decomposition blocks, suggesting the optimal smoothing level is data-dependent.","The arctangent loss can be viewed as a smooth interpolation between MAE and CARD's signal-decay loss, so it may also benefit other regression tasks with long prediction horizons, which the paper does not explore.","The paper's fair-protocol table shows narrower margins against additional baselines, so a public benchmark that re-runs every method in one environment would settle how much of the reported advantage is architectural rather than protocol-driven."],"forward_implications":["If the averaged-table results hold, patching and channel independence transfer cleanly from transformer backbones to CNN backbones, so attention is not required to obtain their forecasting benefits.","The EMA decomposition module can replace the simple moving average decomposition in existing models; the paper reports consistent gains on Autoformer, FEDformer, DLinear, and PatchTST.","The arctangent loss and sigmoid learning-rate schedule improve several existing forecasters when used in place of their default training objectives, making these components separable contributions.","xPatch's measured per-step training and inference times are about 3.1 ms and 1.3 ms, several times lower than CARD and TimeMixer, implying comparable accuracy at lower compute if the baseline comparison holds."],"supporting_citations":[{"why":"Supplies the patch embedding and channel-independence design that xPatch adapts, and serves as the main transformer baseline to beat.","marker":"Nie et al. 2023"},{"why":"Introduces the signal-decay loss that the arctangent loss generalizes, and provides the CARD transformer baseline used for the headline comparisons.","marker":"Wang et al. 2024b"},{"why":"Defines the unified forecasting protocol with fixed lookback and prediction lengths, and supplies several baseline results used in the comparison tables.","marker":"Wu et al. 2023"},{"why":"Establishes the simple linear baseline that challenged transformers and the seasonal-trend preprocessing paradigm that xPatch extends.","marker":"Zeng et al. 2023"},{"why":"Provides the simple moving average decomposition with padding that the paper argues biases trend extraction, and serves as a baseline for EMA-replacement experiments.","marker":"Wu et al. 2021"},{"why":"Represents the prior transformer approach using exponential smoothing, giving context and a baseline for the proposed EMA decomposition.","marker":"Woo et al. 2022"},{"why":"Provides the TimeMixer MLP baseline and the hyperparameter-search protocol used in the second set of experiments.","marker":"Wang et al. 2024a"},{"why":"Supplies the exponential smoothing formulation on which the EMA seasonal-trend decomposition is based.","marker":"Gardner Jr 1985"}],"fun_headline_variants":["xPatch: Exponential decomposition tops transformer forecasters","Dual-stream time-series model beats transformers on benchmarks","Exponential seasonal-trend decomposition improves forecasting accuracy","xPatch: Simpler streams outperform transformer attention in forecasting","Non-transformer model wins on most time-series forecasting benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported improvements assume that baseline numbers taken from other papers and code repositories were produced under the same experimental conditions as the authors' own runs, including the same history length, training epochs, randomness, data splits, and hardware.","fun_headline_variants_meta":{"raw":{"variants":["xPatch: Exponential decomposition tops transformer forecasters","Dual-stream time-series model beats transformers on benchmarks","Exponential seasonal-trend decomposition improves forecasting accuracy","xPatch: Simpler streams outperform transformer attention in forecasting","Non-transformer model wins on most time-series forecasting benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1263,"prompt_tokens":871,"completion_tokens":392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":316}},"tokens_in":487,"tokens_out":392,"duration_ms":3951,"temperature":1.0,"reasoning_tokens":316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:35:16.886480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run xPatch and each main baseline (at least CARD, PatchTST, TimeMixer, and iTransformer) in a single codebase on the same GPU, using the same data splits, lookback lengths, prediction horizons, seed protocol, and number of epochs, and check whether the reported margins of roughly 2-8% in MSE and MAE persist; if the margins shrink below noise or reverse, the empirical superiority claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the exponential smoothing formulation on which the EMA seasonal-trend decomposition is based."}],"review_version":1}