{"id":"b10c6087-1485-4ad3-80a9-f6b0d651c608","arxiv_id":"2412.12227","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"EDformer combines moving-average decomposition with an iTransformer-style variate-token encoder and claims state-of-the-art forecasting, but its reported benchmark results do not consistently support that claim.","lead":"EDformer is a transformer model for multivariate time series forecasting that splits each variable's history into trend and seasonal parts, then treats each variable's whole series as a token, following the iTransformer design. The paper claims state-of-the-art accuracy and speed, but its own results show ATFNet beating it on five of eight long-term datasets, and the key iTransformer baseline is missing from the main comparison.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim is contradicted by the paper's own Table V: ATFNet achieves lower average MSE on five of eight datasets, with large margins on ETTh2 and Weather, so the central accuracy claim fails internally.","rationale":"The reader's verdict of REJECT is well supported, but the weakest assumption they identified (baseline fidelity and tuning) is not the single most decisive problem. The most load-bearing concern is internal inconsistency: the paper's own Table V shows EDformer is not state-of-the-art even under the authors' own baseline configurations. This does not depend on re-running or re-tuning any baseline; it is visible from the reported averages. ATFNet outperforms EDformer on five of eight datasets, often by a wide margin, while EDformer's wins are narrow or on datasets where the differences are small. Thus the abstract's claim that EDformer obtains state-of-the-art accuracy is contradicted by evidence in the same paper. This internal contradiction alone justifies rejection, and the reader's additional points about missing code, excluded iTransformer comparison, and explainability metric misuse are consistent with this conclusion. My single concern does not change the reader's verdict, so the recommendation remains REJECT.","tokens_in":18397,"tokens_out":2274,"duration_ms":22614,"concrete_test":"Re-tabulate Table V by simply counting, for each dataset, which model has the lowest average MSE across the four horizons. If ATFNet is strictly better than EDformer on five or more of eight datasets, or if EDformer does not have the lowest MSE on at least half the datasets, the state-of-the-art accuracy claim is rejected by the paper's own numbers. This check requires no new experiments; it is a deterministic comparison of the published table.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that EDformer obtains state-of-the-art forecasting accuracy. This claim is falsified by the paper's own comparison in Table V, independent of any question about external baseline quality. Averaging over horizons 96, 192, 336, 720, ATFNet has lower MSE than EDformer on ETTh1 (0.515 vs 0.594), ETTh2 (0.252 vs 0.538), Weather (0.229 vs 0.304), Exchange (0.436 vs 0.537), and ETTm2 (0.169 vs 0.543). EDformer is better only on ETTm1 (0.422 vs 0.458), Electricity (0.195 vs 0.201), and Traffic (0.467 vs 0.479), with the latter two margins being very small. The win-count row in Table V itself records EDformer with 3 MSE wins and ATFNet with 3 MSE wins, not a state-of-the-art superiority. Because this contradiction appears in the authors' own reported results, the burden is not on reimplementing the baselines; the data as presented already fail to support the headline accuracy claim. The efficiency claim is less directly contradicted, but the 'state-of-the-art accuracy and efficiency' claim is not.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EDformer, an encoder-only transformer for multivariate time series forecasting. It decomposes the input into seasonal and trend components, embeds each variate's entire series as a token using a reverse/inverted embedding, applies self-attention across variates and a feed-forward network per variate, and adds the trend component back at the end. The authors report long-term results on eight benchmarks, short-term results on PEMS and M4, efficiency measurements, ablations, and an explainability analysis. They claim that EDformer achieves state-of-the-art accuracy and efficiency and that its decomposition and reverse embedding are key to its performance.","tokens_in":18676,"tokens_out":3912,"duration_ms":36540,"significance":"The architectural idea of combining series decomposition with an inverted transformer is a reasonable incremental direction, and the paper provides a substantial set of experiments: standard benchmarks, ablations, runtime tables, and an explainability evaluation on Electricity. The paper also names its implementation environment and uses public datasets. However, the central accuracy claim is contradicted by the paper's own Table V, where ATFNet achieves lower average MSE on five of eight datasets. The comparison also omits iTransformer, the direct predecessor of the variate-token design, making the incremental value difficult to assess. As presented, the evidence does not support the headline state-of-the-art claim; this is a load-bearing issue rather than a presentational one.","major_comments":[{"comment":"The state-of-the-art accuracy claim is contradicted by the paper's own averaged results in Table V. Averaging over horizons 96, 192, 336, and 720, ATFNet reports lower average MSE than EDformer on ETTh1 (0.515 vs 0.594), ETTh2 (0.252 vs 0.538), ETTm2 (0.169 vs 0.543), Weather (0.229 vs 0.304), and Exchange (0.436 vs 0.537), and lower average MAE on six of the eight datasets. The # of Total Wins row records only a 3-3 tie in MSE. Thus the claim that EDformer obtains state-of-the-art accuracy is not supported by the data presented.","section":"Abstract; Section V.B; Table V"},{"comment":"The main multivariate long-term comparison omits iTransformer, even though Section IV.B identifies iTransformer as the inspiration for the variate-frame design and iTransformer appears only in the M4 supplementary table (Table VIII). Since EDformer's mechanism is essentially an inverted transformer plus a decomposition block, the absence of iTransformer from the long-term benchmark tables prevents the reader from assessing the incremental contribution of the proposed modifications. Without this baseline, the state-of-the-art claim is not actionable.","section":"Section IV.B; Section V.B; Tables V and VIII"},{"comment":"The ablation results suggest that the baseline implementations may be undertuned. The 'without decomposition' configuration reaches an average MSE of 1.010 on ETTh1 and 0.905 on ETTm1, which is substantially worse than published results for the same underlying inverted-transformer design (e.g., iTransformer reports an average MSE of about 0.5 on ETTh1). Because the paper states that baselines were reimplemented via TSLib (Section V.B), the discrepancy needs to be resolved by releasing exact configurations and hyperparameters; otherwise the reported gains from decomposition and reverse embedding may be artifacts of a weak base rather than genuine improvements.","section":"Section VI; Table XI"},{"comment":"The efficiency claim is not established against the relevant state of the art. Tables IX and X compare EDformer with Autoformer, Informer, Reformer, Pyraformer, FEDformer, and NS-Trans, but omit comparably lightweight and high-performing models such as iTransformer and PatchTST. Showing that EDformer is faster than deliberately heavy architectures does not support the stated claim of state-of-the-art efficiency; the runtime comparison needs to include the models that are actually competitive in accuracy.","section":"Section V.B; Tables IX and X"}],"minor_comments":[{"comment":"Table VI is labeled 'Comparison of Standard deviations ... with prediction horizons (96,192,336,720)'; the numbers appear to be standard deviations across the four horizons, but the text does not specify this, nor does it clarify whether the standard deviation is computed over seeds or horizons. Please state the statistic explicitly.","section":"Table VI"},{"comment":"The text says EDformer has 'competitive performance' in short-term forecasting on PEMS, but Table VII shows that NS-Trans is clearly better on all four PEMS datasets (e.g., PEMS03 MSE 0.106 vs 0.187). Rephrase to reflect the actual results.","section":"Section V.B; Table VII"},{"comment":"Equation (1) uses inconsistent notation: X_t is defined in R^{T x N}, but the history is called X_{t-l} and the future X_{t+h}, and the look-back window L and horizon H are not tied to these expressions. Please unify the notation.","section":"Section III"},{"comment":"Equation (4) writes the input as X in R^{L x D}, while Section III defines the input as R^{T x N}; the dimensions and symbols should be consistent across the problem statement and the architecture description.","section":"Section IV.A"},{"comment":"The sentence following Equation (7) is incomplete: 'where n represents an input sequence of length...'. Also, the complexity O(n^2 d) should identify n as the number of variates in the inverted design, not the temporal sequence length.","section":"Section IV.C.1"},{"comment":"The interpretability tables report MAE and MSE values but describe higher values as better for comprehensiveness and sufficiency. Since MAE and MSE are error metrics where lower is better, the direction of the 'wins' needs to be justified and explained; otherwise the reader cannot interpret which model benefits most from an explainability method.","section":"Section VII.A; Tables XII-XV"},{"comment":"Reference [12] is listed as 'Autoformer: Searching transformers for visual recognition' and is used in the text as the Autoformer time-series forecasting paper. The actual time-series Autoformer paper should be cited instead.","section":"Reference [12]"},{"comment":"The 'reverse' operation in the embedding is never formally defined. The text says 'reverse embedding' but does not specify the mathematical transformation, making the method difficult to reproduce.","section":"Section IV.B"}],"recommendation":"reject","confidential_remarks":"The core problem is that the paper's own Table V contradicts the central accuracy claim, and the efficiency claim is tested only against a non-representative baseline set. The missing iTransformer comparison is particularly damaging because the architecture is explicitly built on the inverted-transformer idea. These issues are fixable in principle—by expanding the baseline set and rerunning experiments—but as it stands the manuscript does not support its headline contribution, so I cannot recommend acceptance or even minor revision. I would encourage the authors to release code and configurations and to revise the SOTA claims to match the data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the central accuracy claim fails on the paper's own numbers. Table V averages over four horizons and ATFNet beats EDformer on ETTh1, ETTh2, Weather, Exchange, and ETTm2, often by large margins; EDformer's wins on Electricity and Traffic are tiny. The paper says 'state-of-the-art' anyway. That is not a fixable framing issue, it is a contradiction in the evidence.\n\nWhat's actually new: EDformer combines the iTransformer variate-as-token encoder with the moving-average decomposition from Autoformer/FEDformer. The combination is technically new, and the ablation in Table XI does show decomposition and reverse embedding each help on five datasets. The efficiency numbers in Tables IX-X also look plausible: an encoder-only model with a small token count should be fast. Credit is due for the clean ablation design and for citing iTransformer as the inspiration.\n\nSoft spots, in order of severity. First, the missing iTransformer baseline. Since the core encoder is iTransformer's, omitting it from the long-term comparison makes the headline claim unverifiable. Second, no code or hyperparameters. The paper says baselines were reimplemented with TSLib but gives no configurations, and the 'without decomposition' ablation (MSE 1.010 on ETTh1) is far worse than published iTransformer numbers, which suggests the comparison baselines were undertuned. Third, Table VI is mislabeled: it reports standard deviation across horizons, not run stability. Fourth, the explainability section misuses sufficiency: a higher sufficiency score is not better; sufficiency should drop when only key features are kept. The authors mark EDformer's higher sufficiency as wins, which is backwards. Minor: the conclusion section reference is missing.\n\nThis is a paper for someone who wants a quick benchmark of a lightweight decomposed inverted transformer, not for someone who needs validated SOTA claims. It deserves a desk reject, not referee time, because the load-bearing claim is contradicted by the paper's own reported results. If the authors fix the comparison, add code, and correct the metric definitions, it could be a marginal incremental submission. As is, I would not engage.","headline":"The architecture is a clean iTransformer-plus-decomposition combo, but the paper's own Table V contradicts its state-of-the-art claim, so it doesn't hold up.","tokens_in":19201,"tokens_out":1984,"would_cite":false,"duration_ms":19083,"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":"EDformer claims that decomposing a multivariate series into trend and seasonal parts, embedding each variable's whole series as a token, and applying self-attention across variables yields state-of-the-art forecasts with lower compute.","keywords":["time-series forecasting","multivariate forecasting","transformer encoder","seasonal-trend decomposition","variate-as-token embedding","explainable AI","forecasting efficiency"],"falsifier":"Re-run the ETTh1 horizon-96 comparison with the strongest baseline at its officially released hyperparameters: if EDformer's reported MSE of 0.404 does not beat the baseline's published MSE under identical splits, the state-of-the-art claim collapses; likewise, the no-decomposition ablation's average MSE of 1.010 on ETTh1 should not be worse than a simple linear forecaster.","tokens_in":18197,"feed_emoji":"📈","tokens_out":9932,"duration_ms":87258,"temperature":0.7,"pith_summary":"The paper introduces EDformer, an encoder-only Transformer for multivariate time series forecasting, and claims it reaches state-of-the-art accuracy and speed on standard long-term and short-term benchmarks. The method first decomposes each input into a seasonal part and a trend part with a moving average, then treats each variable's entire series as one token rather than using time-step tokens. Self-attention runs across variables, while a shared feed-forward network processes each variable's series to learn its temporal structure; the trend is added back before the final projection. A sympathetic reader should take the contribution as this specific combination of decomposition and variate-as-token attention, supported by ablations and explainability analyses.","feed_headline":"EDformer claims state-of-the-art forecasts at lower cost","feed_subtitle":"Decomposition plus variate-as-token attention is claimed to beat bigger transformers on real benchmarks.","key_machinery":"The central mechanism is the variate token formed by the reverse operation: instead of one token per time step, each variable's whole seasonal and trend series is embedded as one token, so the token dimension is the number of variables. Eq. (5) defines the update: $h^0_n = \\mathrm{Embedding}(\\mathrm{Reverse}(X^S_{:,n})) + \\mathrm{Embedding}(\\mathrm{Reverse}(X^T_{:,n}))$, then $H^{(l+1)} = \\mathrm{IntBlock}(H^l)$ and the forecast is $\\hat{Y}_{:,n} = \\mathrm{Projection}(h^L_n)$. The multivariate self-attention block computes attention scores $\\frac{QK^T}{\\sqrt{d_k}}$ across variate tokens, while the shared feed-forward network $\\mathrm{FFN}(H') = \\mathrm{ReLU}(H'W_1 + b_1)W_2 + b_2$ processes each variate's representation separately. This division of labor is what lets EDformer drop temporal attention and positional encoding.","core_discovery":"The paper's central claim is that an encoder-only Transformer with the same basic components can reach state-of-the-art multivariate forecasting if the components are used differently: decompose each input into trend and seasonal parts, embed each variable's entire seasonal series as a single token, run self-attention across variables, pass each variable's representation through a shared feed-forward network, then add the projected trend back. The decomposition is $X_T = \\mathrm{AvgPooling}(\\mathrm{Padding}(X))$, $X_S = X - X_T$; the token update is $h^0_n = \\mathrm{Embedding}(\\mathrm{Reverse}(X^S_{:,n})) + \\mathrm{Embedding}(\\mathrm{Reverse}(X^T_{:,n}))$, followed by stacked blocks of multivariate self-attention and feed-forward layers before projection. Attention therefore captures cross-variate correlations while the feed-forward network stores temporal structure, and no positional encoding is needed because the per-variate MLP preserves sequence order.","pith_inferences":["If the reported speed holds, the decomposition-plus-variate-token recipe could be grafted onto other encoder-only forecasters to cut the cost of temporal attention, not just as a standalone model.","The ablation suggests decomposition alone already helps; a natural test is whether the order of decomposition and reverse embedding matters, or whether a simpler per-variate MLP with trend addition matches EDformer.","The high sufficiency scores on Electricity imply EDformer concentrates predictive signal in a few features; a testable extension is to prune low-attribution channels and measure accuracy retention.","Because the paper provides no code or baseline configurations, an immediate extension is to check how EDformer behaves when compared against baselines at their officially tuned settings."],"forward_implications":["On most long-term benchmarks and horizons, EDformer reports lower average MSE and MAE than the compared transformer baselines, with the largest gains on high-dimensional datasets such as Electricity and Traffic.","EDformer's per-iteration and total execution times are substantially lower than the compared models, so it can be retrained more often and used in settings with limited compute.","Ablations show that both decomposition and reverse embedding contribute to accuracy, and their combination is better than either alone.","On Electricity, EDformer scores highest on comprehensiveness and sufficiency across five explainability methods, indicating that its predictions track the features that attribution methods flag as important.","The efficiency result implies that the model can be updated more frequently on streaming data, making it feasible for near-real-time forecasting workflows."],"supporting_citations":[{"why":"supplies the variate-as-token reverse embedding idea that EDformer adapts to decomposed seasonal and trend signals.","marker":"[14]"},{"why":"supplies the moving-average decomposition into trend and seasonal components and serves as a primary baseline.","marker":"[12]"},{"why":"is a long-sequence transformer baseline whose quadratic-attention inefficiency motivates EDformer's lightweight design.","marker":"[11]"},{"why":"provides a frequency-domain decomposed transformer baseline that EDformer must beat on long-term benchmarks.","marker":"[16]"},{"why":"provides the benchmark time-series library, dataset splits, and the comprehensiveness/sufficiency explainability metrics used in the evaluation.","marker":"[28]"},{"why":"is the non-stationary transformer baseline whose normalization strategy EDformer contrasts with its own per-variate layer norm.","marker":"[29]"},{"why":"is a recent time-frequency ensembling baseline that EDformer compares against on long-term datasets.","marker":"[30]"},{"why":"is a multi-scale local-global context baseline that EDformer compares against on long-term datasets.","marker":"[31]"}],"fun_headline_variants":["EDformer: decompose first, then attention across variables","Variate-as-token attention beats bigger transformers","EDformer reuses Transformer parts to top charts","Decompose, embed, attend across variables: EDformer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison baselines were faithfully reimplemented and equally tuned; if they were undertuned, the reported accuracy gains are artifacts rather than real improvements.","fun_headline_variants_meta":{"raw":{"variants":["EDformer: decompose first, then attention across variables","Variate-as-token attention beats bigger transformers","EDformer reuses Transformer parts to top charts","Decompose, embed, attend across variables: EDformer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000443,"raw_usage":{"total_tokens":2242,"prompt_tokens":946,"completion_tokens":1296,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1233}},"tokens_in":562,"tokens_out":1296,"duration_ms":9130,"temperature":1.0,"reasoning_tokens":1233,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:42:36.889899+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the ETTh1 horizon-96 comparison with the strongest baseline at its officially released hyperparameters: if EDformer's reported MSE of 0.404 does not beat the baseline's published MSE under identical splits, the state-of-the-art claim collapses; likewise, the no-decomposition ablation's average MSE of 1.010 on ETTh1 should not be worse than a simple linear forecaster.","supporting_citations":[{"cited_title":"Autoformer: Searching transformers for visual recognition","cited_arxiv_id":null,"evidence_quote":"supplies the moving-average decomposition into trend and seasonal components and serves as a primary baseline."},{"cited_title":"Informer: Beyond efficient transformer for long sequence time- series forecasting","cited_arxiv_id":null,"evidence_quote":"is a long-sequence transformer baseline whose quadratic-attention inefficiency motivates EDformer's lightweight design."},{"cited_title":"Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting","cited_arxiv_id":null,"evidence_quote":"provides a frequency-domain decomposed transformer baseline that EDformer must beat on long-term benchmarks."},{"cited_title":"Sa-timeseries: Self-attention time series models, 2024","cited_arxiv_id":null,"evidence_quote":"provides the benchmark time-series library, dataset splits, and the comprehensiveness/sufficiency explainability metrics used in the evaluation."},{"cited_title":"Non-stationary transformers: Exploring the sta- tionarity in time series forecasting","cited_arxiv_id":null,"evidence_quote":"is the non-stationary transformer baseline whose normalization strategy EDformer contrasts with its own per-variate layer norm."}],"review_version":1}