{"id":"28a1e2a2-b234-47da-8b98-079e61f00a7e","arxiv_id":"2412.06866","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A new lightweight MLP forecaster using learnable FFT filters and a differencing operation is competitive with, but not consistently better than, current state-of-the-art time series models.","lead":"This paper introduces LMS-AutoTSF, a time series forecasting model that combines learnable frequency-domain trend and seasonal filters with multi-scale processing and a lagged-difference term it calls autocorrelation. It claims state-of-the-art accuracy and lower compute, but its own results show it wins on only four of eight long-term benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'state-of-the-art' claim is contradicted by the paper's own Table 2: on ETTh1, ETTm1, Traffic and Exchange, LMS-AutoTSF's average MSE is worse than a listed baseline, so the central empirical claim is unsupported as reported.","rationale":"I read the paper as making an empirical claim: a lightweight architecture with learnable FFT decomposition and a lagged-difference 'autocorrelation' branch achieves state-of-the-art forecasting accuracy on most benchmarks. That claim is directly testable from the paper's own tables, and it fails there. The reader's weakest assumption focused on Equation 9 and the marginal ablation gain, which is a legitimate mechanism-level concern, but the more decisive problem is that the reported results themselves do not establish SOTA. I agree with the reader's overall REJECT verdict, but I would ground it primarily in the Table 2 contradiction rather than in the autocorrelation mechanism, so my agreement is partial. I give credit where it is due: the code is public, the model is lightweight, and the FLOPs/runtime comparison in Table 5 supports the efficiency claim. However, the conclusion repeats an unqualified SOTA claim that the own experimental section contradicts. The concrete test I propose is a straightforward recomputation and ranking exercise; it would settle whether the headline claim survives, and if it does not, the paper should be reframed as an efficient competitive forecaster rather than a state-of-the-art method.","tokens_in":12896,"tokens_out":4921,"duration_ms":50081,"concrete_test":"Reproduce Table 2 using the released code and the paper's stated protocol (fixed look-back 96, horizons 96/192/336/720, same TSLib baseline replicas). For each dataset, compute the average MSE across the four horizons and rank LMS-AutoTSF among the eight listed models. If LMS-AutoTSF is not top-ranked on at least five of the eight datasets, or if it remains behind FEDformer on ETTh1, ETSFormer on ETTm1, iTransformer on Traffic, and DLinear on Exchange, the state-of-the-art claim is falsified by the paper's own comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, 'LMS-AutoTSF consistently achieves state-of-the-art performance across a wide range of benchmarks,' is load-bearing and fails against the paper's own evidence. In Table 2 (fixed look-back 96, horizons 96/192/336/720), the average MSE of LMS-AutoTSF is 0.441 on ETTh1 while FEDformer is 0.439; 0.377 on ETTm1 while ETSFormer is 0.304; 0.497 on Traffic while iTransformer is 0.444; and 0.353 on Exchange while DLinear is 0.340. These gaps are substantial on Traffic and Exchange, and on the remaining datasets the paper is often competitive rather than superior. The conclusion and contribution 3 state SOTA without qualification, but the reported tables support at most 'competitive on some datasets and efficient.' A secondary internal inconsistency compounds this: the introduction assigns low-pass filtering to seasonality and high-pass filtering to trend, while Section 4 assigns them exactly oppositely, so it is unclear what the learned decomposition is being credited with. The absence of error bars or multiple seeds in the ablation makes the marginal autocorrelation gain in Table 1 (e.g., ETTh1 MSE 0.448 to 0.441) uninterpretable, but the Table 2 contradiction is sufficient on its own to invalidate the headline claim as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LMS-AutoTSF, a time series forecasting architecture that combines multi-scale downsampling, learnable FFT-based low-pass and high-pass filters, two fully connected encoders per scale (one for trend, one for seasonality), and an autocorrelation mechanism implemented as element-wise multiplication by a lagged difference. The authors claim that this lightweight, fully connected design achieves state-of-the-art forecasting performance across long-term and short-term benchmarks, and they support this with experiments on ETT, Weather, Electricity, Traffic, Exchange, PEMS, and M4 datasets, including an ablation study and an efficiency comparison.","tokens_in":13225,"tokens_out":3215,"duration_ms":32584,"significance":"If the state-of-the-art claim were supported, the contribution would be notable: a simple, non-transformer, efficient architecture competing with or beating strong baselines would be a useful addition to the forecasting toolbox. The manuscript has genuine strengths: the source code is publicly available, the experimental coverage is broad (eight multivariate long-term datasets, four PEMS datasets, and M4), and the ablation study isolates the proposed components. However, the central empirical claim is substantially overstated relative to the paper's own tables, and the absence of error bars or multiple-seed results makes the reported gains difficult to interpret. The significance of the work is therefore currently limited to a competitive-and-efficient architecture rather than a state-of-the-art one.","major_comments":[{"comment":"The claim in Contribution 3 and the Conclusion that LMS-AutoTSF \"achieves state-of-the-art performance\" and \"consistently achieves the best performance across most datasets\" is directly contradicted by Table 2. For the average MSE over horizons 96/192/336/720 with look-back 96, FEDformer is better on ETTh1 (0.439 vs. 0.441), ETSFormer is substantially better on ETTm1 (0.304 vs. 0.377), iTransformer is substantially better on Traffic (0.444 vs. 0.497), and DLinear is better on Exchange (0.340 vs. 0.353). The differences on Traffic and Exchange are large, and on the remaining datasets LMS-AutoTSF is often competitive rather than superior. Since the headline contribution is explicitly framed as state-of-the-art, this evidence is load-bearing and must be either corrected or the claim must be reframed as \"competitive on many benchmarks while being more efficient.\"","section":"Section 5.2, Table 2"},{"comment":"There is an internal inconsistency in the role of the filters. Section 1 states that \"low-pass filtering is commonly used to capture seasonal variations\" and \"high-pass filtering is employed to model trends.\" In contrast, Section 4, Eq. (5) defines the low-pass filter as capturing the trend, and Eq. (6) defines the high-pass filter as extracting the seasonal component. The two descriptions are exactly reversed. This matters because the paper's first contribution is the \"dynamic decomposition\" into trend and seasonality; if the authors themselves are not consistent about which filter extracts which component, the reader cannot know what the learned decomposition is actually doing. Please correct the introduction or the method description and make the semantics consistent throughout.","section":"Section 1 vs. Section 4 (Eqs. 5-6)"},{"comment":"The ablation study in Table 1 reports only a single run per configuration, with no error bars, standard deviations, or significance tests. The claimed benefit of the integrated autocorrelation is small: on ETTh1 the MSE goes from 0.448 to 0.441, and on ETTm1 from 0.392 to 0.377, while on Weather and Electricity the differences are tiny (0.240 vs. 0.238 and 0.174 vs. 0.175, respectively). These gaps are within the range of typical run-to-run variability for deep forecasting models, so the contribution of autocorrelation is not established. In addition, Eq. (9) multiplies the processed temporal features by the lagged difference and the text asserts this \"enables the model to capture dependencies across time more effectively,\" but no experiment isolates this operation nor analyzes its effect. Please provide multiple seeds with variance and, ideally, a statistical comparison, or soften the claim about autocorrelation's benefits.","section":"Table 1, Eq. (9)"}],"minor_comments":[{"comment":"The caption labels the architecture \"LMS-AutoTS\" while the model name is LMS-AutoTSF; please make the figure consistent with the text.","section":"Figure 1"},{"comment":"The symbol \\Delta T^{(k)} is used for the lagged difference but is never formally defined. Please define it explicitly when it is first introduced.","section":"Section 4, Eq. (9)"},{"comment":"The text refers to a \"supplementary file\" for dataset details, but no supplementary material is provided in the manuscript; either include the file or remove the reference.","section":"Section 5.1"},{"comment":"Reference [12] is cited as \"Autoformer: Searching transformers for visual recognition,\" which is a different paper on visual recognition, not the Autoformer time series forecasting paper by Wu et al. Please correct the citation.","section":"References"},{"comment":"The M4 table reports comparisons only against iTransformer, TimeMixer, and PatchTST, but the text says \"state-of-the-art models\"; this is a limited set, and the conclusion that LMS-AutoTSF is best is not supported by the table, since TimeMixer has lower average sMAPE and MAPE.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a publicly available codebase and a broad set of experiments, but the empirical claims need substantial revision before publication. The state-of-the-art claim is contradicted by the paper's own Table 2, and this is a load-bearing issue for the framing. The filter-semantics inconsistency in Section 1 vs. Section 4 is an editorial problem but also signals a need for clarity about the method. I would not reject outright because the core architecture and efficiency results are potentially useful; however, the authors must either produce a corrected, statistically grounded evaluation or explicitly limit their claims to \"competitive and efficient\" rather than \"state-of-the-art.\""},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper's central claim—consistent state-of-the-art performance across benchmarks—fails against its own Table 2. On ETTh1, ETTm1, Traffic, and Exchange, the average MSE is worse than a listed baseline (FEDformer, ETSFormer, iTransformer, DLinear). That is not a subtle point; the conclusion repeats the SOTA claim without qualification. The stress-test note is right, and I agree with the reader's REJECT verdict, though I want to be fair about what is here.\n\nWhat is actually new: the specific combination of learnable FFT low/high-pass filters, multi-scale average pooling, and a first-difference 'autocorrelation' term inside a fully connected encoder. Each component is known, but the arrangement is new. The paper ships code, which is good, and evaluates on standard long-term benchmarks plus PEMS and M4. When it is competitive, it is competitive with a small model, and the efficiency numbers (FLOPs, runtime) are credible.\n\nSoft spots, in proportion: (1) The SOTA claim is contradicted by the paper's own numbers, so at best this is 'competitive and efficient on some datasets.' (2) The introduction assigns low-pass to seasonality and high-pass to trend, while Section 4 says exactly the opposite—this makes the learned decomposition story confusing and undermines the technical narrative. (3) The autocorrelation term is element-wise multiplication by the first lagged difference, with no theoretical support, and the ablation shows a tiny gain (e.g., 0.448 to 0.441 MSE on ETTh1) from a single run without error bars or significance testing. (4) No multiple seeds or variance reporting anywhere.\n\nNone of this makes the architecture worthless. The model is a plausible lightweight baseline that could be useful to people building MLP forecasters, and the code availability lowers the barrier to testing it. But as written, the claims are unsupported.\n\nIf this lands on my desk, I would send it to peer review rather than desk-reject—the code, breadth of benchmarks, and lightweight contribution deserve referee time—but I would expect reviewers to demand major revision: drop or heavily qualify the SOTA claim, fix the filter/trend/seasonality inconsistency, and add multi-seed runs with error bars. As is, reject.","headline":"A useful lightweight forecaster buried under unsupported SOTA claims; the paper's own Table 2 contradicts the headline.","tokens_in":13749,"tokens_out":1962,"would_cite":false,"duration_ms":19812,"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":"LMS-AutoTSF claims state-of-the-art forecasting accuracy using learnable frequency filters and lagged-difference autocorrelation.","keywords":["time series forecasting","learnable decomposition","frequency domain filtering","autocorrelation","multi-scale processing","lightweight architecture","multivariate forecasting","FFT"],"falsifier":"Ablate Equation (9) by replacing $\\Delta T^{(k)}$ with a constant vector of ones and retrain on the same eight benchmarks with multiple seeds; if mean squared error does not degrade significantly on most datasets, the claimed benefit of integrated autocorrelation is not supported. The paper's own ablation shows only small gains, e.g., ETTh1 MSE 0.448 to 0.441, with no variance or significance testing.","tokens_in":12679,"feed_emoji":"📈","tokens_out":11331,"duration_ms":89253,"temperature":0.7,"pith_summary":"LMS-AutoTSF is a time series forecasting architecture that aims to show a lightweight fully connected network can match or beat transformer-based models. The paper argues that learnable low-pass and high-pass filters in the frequency domain can isolate trend and seasonal components without predefined decomposition, and that multiplying processed features by the first lagged difference injects autocorrelation into the model. If correct, the approach would mean expensive attention mechanisms are not necessary for state-of-the-art long-term and short-term forecasting on standard benchmarks. The paper reports the best or second-best mean squared error on most datasets, along with far lower FLOP counts than transformer baselines.","feed_headline":"Learnable filters and autocorrelation claim top forecasting accuracy","feed_subtitle":"Uses learnable frequency filters and lagged differences to beat heavier transformer rivals at a fraction of the compute.","key_machinery":"The load-bearing mechanism is the dual-encoder-per-scale processing driven by learnable frequency-domain filters: for each of $K$ downsampled scales, a trainable low-pass filter $\\sigma(-(f - f_{\\text{cutoff}}) s)$ isolates the trend and a trainable high-pass filter $\\sigma((f - f_{\\text{cutoff}}) s)$ isolates the seasonal component, with cutoff and steepness learned per feature. A second mechanism is the autocorrelation injection of Equation (9), where the temporally processed feature $x_{\\text{temp}}$ is multiplied elementwise by the first lagged difference $\\Delta T^{(k)}$ of that scale's trend (or seasonal) input, intended to make the network aware of how much the series changes between consecutive steps. The paper's argument that these two mechanisms together produce better predictions than either alone rests on an ablation table showing monotone MSE improvements from fixed decomposition to learnable decomposition to learnable decomposition plus autocorrelation on ETTh1, ETTm1, Weather, and Electricity.","core_discovery":"The central claim is that the proposed LMS-AutoTSF architecture achieves state-of-the-art performance in multivariate time series forecasting by combining three mechanisms: multi-scale downsampling of the input, learnable FFT-based low-pass and high-pass filters that decompose each scale into trend and seasonal components, and an autocorrelation term computed as the element-wise product of the temporally processed features with the first lagged difference of the encoder input. The trend and seasonal encoders each use fully connected layers for temporal and channel interactions, and their per-scale forecasts are concatenated and projected to the final prediction. The paper reports that this design outperforms or matches transformer-based models such as iTransformer, PatchTST, and TimeMixer on long-term benchmarks, achieves competitive results on PEMS short-term traffic forecasting, and does so with substantially lower execution time and FLOP counts.","pith_inferences":["If the autocorrelation mechanism is genuinely useful, replacing the fixed one-step difference with learnable lag operators selected by the input's autocorrelation function could improve accuracy further; the paper does not explore this.","The efficiency comparison is reported on a single GPU; an independent inference benchmark on CPU or edge hardware would test whether the FLOP advantage translates to real-world deployments.","The ablation's small gains suggest the lagged-difference product may matter most on nonstationary series; stratifying results by trend strength (e.g., Exchange vs. Weather) could reveal where the mechanism carries the performance."],"forward_implications":["If the central claim holds, fully connected MLP architectures can achieve competitive or better long-term forecasting accuracy than transformer models, suggesting attention is not required for these benchmarks.","The reported FLOP and execution-time reductions would make the method suitable for resource-constrained forecasting applications, such as on-device traffic or energy prediction.","The learnable frequency-domain decomposition would allow the model to adapt to datasets with different trend-seasonality structures without manual decomposition choices.","The simple lagged-difference autocorrelation provides a low-cost alternative to attention for capturing temporal dependencies, one that could be dropped into other MLP forecasters."],"supporting_citations":[{"why":"TimeMixer, the multiscale MLP baseline that LMS-AutoTSF must beat to support its state-of-the-art claim on long-term benchmarks.","marker":"[18]"},{"why":"iTransformer, a transformer baseline whose cross-channel attention is contrasted with the paper's fully connected channel processing.","marker":"[6]"},{"why":"PatchTST, a patching transformer baseline used in the long-term and short-term comparisons.","marker":"[9]"},{"why":"DLinear, the linear decomposition baseline that motivates the paper's learnable frequency-domain decomposition.","marker":"[10]"},{"why":"FEDformer, a frequency-enhanced transformer baseline, used to justify the frequency-domain filtering approach.","marker":"[15]"},{"why":"The frequency-domain MLP method that supplies the FFT-based filtering idea for the learnable decomposition.","marker":"[19]"}],"fun_headline_variants":["Learnable frequency filters and autocorrelation beat transformer rivals","Multi-scale filtering plus autocorrelation for efficient high-accuracy forecasting","LMS-AutoTSF: lightweight forecasting with learnable trend-season filters","Autocorrelation and dual-scale encoders outpace heavier transformers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that multiplying the temporally processed features by the first lagged difference of the encoder input is what lets the model capture temporal dependencies; if that elementwise multiplication contributes nothing, the architecture reduces to a multi-scale MLP with learnable frequency filters, and the paper's key innovation claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Learnable frequency filters and autocorrelation beat transformer rivals","Multi-scale filtering plus autocorrelation for efficient high-accuracy forecasting","LMS-AutoTSF: lightweight forecasting with learnable trend-season filters","Autocorrelation and dual-scale encoders outpace heavier transformers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1493,"prompt_tokens":960,"completion_tokens":533,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":461}},"tokens_in":576,"tokens_out":533,"duration_ms":5700,"temperature":1.0,"reasoning_tokens":461,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:45:33.893345+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ablate Equation (9) by replacing $\\Delta T^{(k)}$ with a constant vector of ones and retrain on the same eight benchmarks with multiple seeds; if mean squared error does not degrade significantly on most datasets, the claimed benefit of integrated autocorrelation is not supported. The paper's own ablation shows only small gains, e.g., ETTh1 MSE 0.448 to 0.441, with no variance or significance testing.","supporting_citations":[{"cited_title":"itrans- former: Inverted transformers are effective for time series forecasting","cited_arxiv_id":null,"evidence_quote":"iTransformer, a transformer baseline whose cross-channel attention is contrasted with the paper's fully connected channel processing."},{"cited_title":"Fedformer: Frequency en- hanced decomposed transformer for long-term series forecasting","cited_arxiv_id":null,"evidence_quote":"FEDformer, a frequency-enhanced transformer baseline, used to justify the frequency-domain filtering approach."}],"review_version":1}