{"id":"f69e73e3-0e57-4f4a-9362-b4b4529ff022","arxiv_id":"2411.17382","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"MFF-FTNet combines CoST-style contrastive learning, TimesNet-style top-k frequency selection, and TCN-style multi-scale convolutions, reporting modest average MSE gains but leaving the forecasting loss undefined.","lead":"This paper proposes MFF-FTNet, a time series forecasting network that combines contrastive learning with frequency-domain and multi-scale time-domain feature extraction. It reports average MSE improvements on five public datasets, but the method omits the forecasting objective and the results do not consistently beat the baselines in its own tables.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No forecasting objective or prediction head is defined: Eq. 16 contains only contrastive losses, so the reported MSE improvements cannot be attributed to a reproducible forecasting procedure.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing gap: the paper never defines a forecasting loss or prediction head. I agree with that assessment. Equation 16 gives the total loss as γ1 L_time + γ2 L_freq, both contrastive objectives, and Section III-A's mapping g(·) is never specified. Without a defined mapping from learned features to future values, the reported test MSE has no algorithmic basis. The central claim, a 7.7% MSE improvement, cannot survive this gap because the comparison in Tables IV and V is between numbers generated by an unspecified procedure and numbers generated by well-defined baselines. This is a coherence failure inside the paper, not a disagreement with community consensus, and it is not resolved by the ablations or parameter studies, which compare variants of the same underspecified pipeline. The absence of code and seeds compounds the problem but is secondary. Since the reader's verdict is already REJECT and my concern supports that verdict, no change to the verdict is needed.","tokens_in":17950,"tokens_out":3269,"duration_ms":32579,"concrete_test":"Implement exactly the model described in Section III using only the components and losses specified in Eqs. 1–16, then attempt to produce a forecast XO for a held-out batch of ETTh1. If no forecasting output can be generated from the architecture as written, the reported MSE in Tables IV–V is not reproducible. Alternatively, obtain the official code and check whether an additional supervised forecasting loss or prediction head is used; if it is, that component is omitted from the paper and is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MFF-FTNet outperforms state-of-the-art forecasters, with a 7.7% MSE reduction. For that claim to hold, the method must specify how learned representations become forecasts XO. The paper does not. Section III-A defines forecasting as learning g(·) with XO = g(X), but g is never specified anywhere. The only training objective is Eq. 16, L_total = γ1 L_time + γ2 L_freq, where L_time (Eq. 15) is a contrastive loss between a backbone representation r_t and a fused representation h_t, and L_freq (Eq. 10) is a contrastive loss on frequency amplitude and phase. Neither term involves future values or any supervised prediction error. No decoder, linear probe, or forecasting layer is described in Section III or in the experimental setup. Consequently, the MSE and MAE numbers in Tables IV and V cannot be produced from the method as written; a reader could not reimplement the reported system. This is an internal incompleteness, not a matter of disagreement with existing benchmarks: the entire experimental evaluation rests on an unstated forecasting mechanism, and the ablation and robustness studies inherit the same gap because they evaluate the same underspecified pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MFF-FTNet, a time-series forecasting architecture combining an adaptive noise augmentation scheme, a Frequency-Aware Contrastive Module (FACM) that operates on FFT spectra with frequency masking and a dual contrastive loss, and a Complementary Time Domain Contrastive Module (CTCM) with multi-scale convolutions and feature fusion. The authors report state-of-the-art results on five datasets, claiming a 7.7% average MSE improvement over baselines in multivariate forecasting and a 2.5% improvement in univariate forecasting, plus ablation, robustness, transfer, and visualization experiments. The central problem is that the manuscript never specifies a forecasting head or any supervised prediction objective; the only training losses defined (Section III-G, Eq. 16) are contrastive losses, so the reported forecasting numbers cannot be reproduced from the described method. Several internal inconsistencies in the FACM equations and in the contrastive loss definitions compound this issue.","tokens_in":18264,"tokens_out":3116,"duration_ms":29454,"significance":"If the framework were fully specified and the empirical claims verified, the idea of combining frequency-selective contrastive learning with multi-scale temporal convolutions could be of interest to the time-series forecasting community. However, the paper as written does not provide a reproducible forecasting method: the prediction mechanism is absent, and the reported experimental results therefore do not establish the claimed improvements. The paper also presents no code, no detailed training procedure for the forecasting output, and the experimental comparison tables contain direct counterexamples to the 'consistently outperforms' claim. The contribution is thus currently not assessable as a forecasting method.","major_comments":[{"comment":"The paper never defines a forecasting loss or a prediction head. Equation (16) defines L_total = γ1 L_time + γ2 L_freq, where L_time (Eq. 15) and L_freq (Eq. 10) are contrastive objectives, and neither term involves the future values XO or any supervised prediction error. Section III-A states that the goal is to learn g(·) with XO = g(X), but g is never specified, and no decoder, linear probe, or forecasting layer is described anywhere in Section III or Section IV. Consequently, the MSE and MAE numbers in Tables IV and V cannot be produced from the method as written, and a reader cannot reimplement the reported system.","section":"Section III-G / Eq. (16)"},{"comment":"The FACM equations are dimensionally inconsistent. The input is r ∈ R^{T×K}, so its FFT should be a complex tensor of size T×K (or c×K after truncation), and averaging over the K dimension yields a vector of length T, not D. Yet Eq. (5) declares F ∈ R^{c×D}, and Algorithm 1 in Step 2 creates weights ωp ∈ R^{D×K/2} and biases βp ∈ R^{c×K/2}, mixing the latent dimension K with the input dimension D. This makes Equations (4)–(7) ill-defined as written; the authors must clarify the actual shapes used in their implementation.","section":"Section III-E, Eqs. (4)–(6) and Algorithm 1"},{"comment":"The contrastive loss formulas do not match the accompanying text. The text states that 'different samples within the same batch are considered negative pairs,' but the denominators in Eqs. (8) and (9) sum over frequency indices k (k ≠ j) within the same pair of augmented views, not over other batch samples. Thus the equations define a loss that treats different frequency components as negatives, which is a different objective from the described batch-level contrastive learning. This discrepancy makes it impossible to determine what loss was actually minimized during training.","section":"Section III-E3, Eqs. (8)–(9)"},{"comment":"The claim that MFF-FTNet 'consistently outperforms' baselines is contradicted by the reported numbers. In Table IV, on WTH with horizon 720, MFF-FTNet has MSE 0.534 while TCN achieves 0.498. In Table V, on ETTh2 with horizon 720, MFF-FTNet has MSE 0.212 versus TCN's 0.200, and on WTH horizons 168, 336, and 720, TCN and CoST surpass MFF-FTNet. Table VII also shows that on WTH→ETTh1 with horizon 24, CoST achieves 0.375 versus MFF-FTNet's 0.382. The aggregated percentage improvements in the text (7.7%, 2.5%) therefore do not reflect a consistent advantage, and the central claim needs to be substantially qualified.","section":"Section IV-D, Tables IV and V"}],"minor_comments":[{"comment":"The mean and standard deviation are written as functions of xt for t = 1,...,T, but xt is previously defined as a slice of X at time t with all D features; the notation should make clear whether the mean and variance are computed per feature dimension or as scalars.","section":"Section III-C, Eq. (1)"},{"comment":"The text states that the Weather dataset contains 11 features, while Table II lists 12 features. Please reconcile this discrepancy.","section":"Section IV-A and Table II"},{"comment":"The paper claims an 8.9% MSE reduction on the ETT datasets compared to CoST, but it is not clear whether this figure is the average over the four ETT datasets or computed in some other way; the averaging procedure should be stated explicitly.","section":"Section IV-D, text after Table V"},{"comment":"The figure is referenced in the abstract and introduction but appears to be a qualitative performance comparison without a caption or axis labels in the manuscript text; it should be described and referenced properly.","section":"Figure 1"},{"comment":"The ablation table reports a single MSE/MAE value for each variant without specifying the prediction horizon or whether the numbers are averaged over the horizons shown in the main tables; the experimental protocol for these ablations should be documented.","section":"Section IV-E, Table VI"},{"comment":"The sensitivity analysis plots show performance trends but no numerical values or error bars; specifying the reported metric and the evaluation protocol would strengthen the claims about the effect of L′ and D′.","section":"Section IV-G, Figures 7 and 8"}],"recommendation":"reject","confidential_remarks":"The paper's central problem is not a disagreement with current consensus but an internal incompleteness: no forecasting objective exists in the manuscript, so the empirical evaluation is not reproducible from the described method. Adding a prediction head and supervised loss would change the entire experimental pipeline, making this more than a local fix. The dimensional inconsistency in the FACM and the mismatch between the text and formulas of the contrastive loss further indicate that the manuscript is currently not ready for review. I would recommend rejection, with the possibility of a fresh submission if the authors provide a fully specified forecasting procedure and rerun the experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the architecture is a new combination of established pieces: CoST backbone, TimesNet-style top-k FFT selection, multi-scale convolutions, and dual contrastive losses. That combination is not in the cited literature, and the adaptive noise augmentation is a reasonable idea. Second, the central flaw is that the paper never defines a forecasting objective or prediction head. The only training loss is Eq. 16, a weighted sum of two contrastive losses. Neither term involves future values or supervised prediction error. So the MSE and MAE numbers in Tables IV and V cannot be produced from the method as written. A reader could not reimplement the reported system.\n\nThe paper does some things well. The motivation is clearly stated, the modules are described with enough detail to understand the intended mechanism, and the ablation study shows each component contributes, even if the gains are small. The robustness experiments on noisy and missing data are a nice addition, and the paper engages with the relevant literature (CoST, TimesNet, TS2Vec). For a representation-learning recipe, the design is coherent.\n\nBut the soft spots are significant. Beyond the missing forecasting head, there are dimensional inconsistencies: Eq. 5 and Algorithm 1 mix D and K dimensions, and the contrastive losses in Eqs. 8-9 treat frequencies as negatives while the text says negatives come from other batch samples. The claim of consistent outperformance is also contradicted by the paper's own tables: TCN wins on WTH-720, CoST wins on ETTm2-24, and several transfer-learning rows go to CoST. No code, seeds, or error bars are provided, and hyperparameters are tuned per dataset, so the 7.7% average improvement is not supported as a robust finding. The ablation differences are modest (e.g., 0.616 vs. 0.638 MSE), so calling them substantial is an overstatement.\n\nWho is this for? A reader working on contrastive representation learning for time series might find the module combination worth a look, but only as a starting point. As a forecasting paper, it is incomplete. I would not send this to serious peer review in its current form. The right move is a major revision that specifies the forecasting head and supervised loss, fixes the dimension and contrastive-loss mismatches, and redoes the experiments with a reproducible protocol. If the authors cannot supply those, desk reject.","headline":"A plausible combination of known components, but the paper never specifies how representations become forecasts, so the headline MSE gains are not reproducible from the text.","tokens_in":18760,"tokens_out":2583,"would_cite":false,"duration_ms":26562,"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":"MFF-FTNet claims a 7.7% MSE win by fusing frequency and temporal contrastive features.","keywords":["contrastive learning","time series forecasting","frequency domain","multi-scale feature fusion","noise robustness","self-supervised representation learning","long-term forecasting"],"falsifier":"Implement the training loop exactly as described with only the contrastive losses and check whether the trained model can output a forecast for any horizon; if the stated equations are the whole procedure, there is no mechanism to produce the prediction window, so the MSE numbers in the tables must come from an additional, undocumented component. Reproducing the tables only after adding an explicit forecasting head would settle the question.","tokens_in":17754,"feed_emoji":"📉","tokens_out":7211,"duration_ms":63186,"temperature":0.7,"pith_summary":"This paper proposes MFF-FTNet, a self-supervised framework that learns time series representations by contrasting views in both the frequency and time domains. The authors aim to show that combining a frequency-aware contrastive module, which keeps only the strongest spectral components, with a multi-scale time-domain contrastive module, which merges features from eight different kernel sizes, produces representations that are more robust to noise and missing data and more accurate for forecasting. On five real-world datasets they report that this approach consistently outperforms five baselines, with an average 7.7% reduction in MSE for multivariate forecasting and 2.5% for univariate forecasting. The intended contribution is a general recipe for aligning spectral and multi-scale temporal features in one contrastive objective.","feed_headline":"Frequency-plus-time contrastive fusion cuts forecast error 7.7%","feed_subtitle":"Self-supervised model claims fusing spectral and multi-scale temporal views beats five forecast baselines.","key_machinery":"The load-bearing machinery is the pair of contrastive modules. The Frequency-Aware Contrastive Module transforms the latent representation with FFT, keeps the top-k amplitude frequencies (masking ratio $\\alpha_m$), applies learnable weights and bias, returns via iFFT, and then computes separate contrastive losses on the amplitude and phase vectors ($L_{amp}$ and $L_{phase}$, combined as $L_{freq} = \\lambda L_{amp} + (1-\\lambda) L_{phase}$). The Complementary Time Domain Contrastive Module runs eight parallel 1D convolutions with kernel sizes [1,2,4,8,16,32,64,128], stacks the outputs, fuses them with 2D convolution and average pooling, and computes a time contrastive loss $L_{time}$ between the backbone output $r$ and the fused representation $h$. The total training loss is $L_{total} = \\gamma_1 L_{time} + \\gamma_2 L_{freq}$; these two losses are what the learning objective actually contains, and the reported forecasting performance is attributed to their combination.","core_discovery":"The central discovery, on the paper's own terms, is that splitting contrastive learning by frequency amplitude and phase, and simultaneously aligning the backbone representation with a fused multi-scale temporal representation, yields forecasting representations that beat existing contrastive and transformer baselines on every tested horizon. The authors attribute the gain to the frequency masking operation, which discards low-amplitude (mostly noisy) frequency bins, and to the multi-scale fusion, which lets the model see both short spikes and long trends at once. The reported numbers show the full model outperforming the strongest baseline by 7.7% average MSE on multivariate ETT datasets and 2.5% on univariate tasks across all five datasets. As stated, the improvement follows from the complementary domain alignment rather than from a new prediction head.","pith_inferences":["Because the paper specifies no forecasting head and no supervised loss, a reader cannot tell from the text alone how the final predictions are produced; the most natural reading is that an undocumented linear projection or fine-tuning step maps the representation to future values.","A direct extension would be to make that step explicit by training a linear forecaster on top of the frozen contrastive representation; the paper's claims imply this linear probe should reproduce most of the reported accuracy.","The two-module design is modular: the same frequency-aware and time-domain contrastive modules could be attached to other backbones for tasks like anomaly detection or classification, though the paper only evaluates forecasting.","The reported noise-robustness suggests that top-k frequency masking is effectively a learned denoiser, which might be used as a preprocessing layer in other models rather than only inside a contrastive objective."],"forward_implications":["If the claimed 7.7% multivariate MSE gain holds, frequency-domain contrastive learning combined with multi-scale temporal fusion is a practical alternative to transformer-based forecasters on electricity and weather data.","The robustness experiments imply that the learned representation is insensitive to up to 40% missing training data, which would make the approach attractive for real-world sensor streams with dropouts.","The transfer-learning results imply that representations pretrained on one dataset (e.g., weather) can be fine-tuned effectively on another (e.g., ETT), reducing the need for large labeled forecasting datasets.","The ablation ordering suggests that the multi-scale time-domain module contributes more than the frequency module on the tested datasets, which would guide future designs toward wider kernel banks."],"supporting_citations":[{"why":"Supplies the backbone encoder and the seasonal-trend contrastive loss that the frequency-aware module adapts into amplitude and phase losses.","marker":"[21]"},{"why":"Contrastive time series representation baseline whose multi-scale approach is extended by the time-domain module.","marker":"[15]"},{"why":"Provides the top-k frequency selection and masking idea for keeping prominent spectral components.","marker":"[31]"},{"why":"Provides the adaptive weight and bias computation used to weight the selected frequency components.","marker":"[33]"},{"why":"Temporal-neighborhood contrastive baseline that also combines time and frequency features and is outperformed in the tables.","marker":"[34]"},{"why":"Transformer long-sequence forecasting baseline that the proposed model compares against in multivariate and univariate tasks.","marker":"[18]"},{"why":"Convolutional sequence model used as a baseline that the proposed model is measured against.","marker":"[35]"}],"fun_headline_variants":["Frequency and time contrastive fusion cuts forecast error 7.7%","Spectral-temporal contrastive fusion improves multivariate forecasts by 7.7%","Noise-aware multi-scale fusion across frequency and time beats baselines","Fusing spectral and temporal views trims MSE by 7.7%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported forecasting accuracy rests on the unstated assumption that a prediction step exists beyond the two contrastive losses, because the paper defines the total loss as $\\gamma_1 L_{time} + \\gamma_2 L_{freq}$ and never specifies the mapping $g(\\cdot)$ that turns the learned representation into future values.","fun_headline_variants_meta":{"raw":{"variants":["Frequency and time contrastive fusion cuts forecast error 7.7%","Spectral-temporal contrastive fusion improves multivariate forecasts by 7.7%","Noise-aware multi-scale fusion across frequency and time beats baselines","Fusing spectral and temporal views trims MSE by 7.7%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000588,"raw_usage":{"total_tokens":3002,"prompt_tokens":926,"completion_tokens":2076,"prompt_tokens_details":{"cached_tokens":896},"prompt_cache_hit_tokens":896,"prompt_cache_miss_tokens":30,"completion_tokens_details":{"reasoning_tokens":1996}},"tokens_in":30,"tokens_out":2076,"duration_ms":56640,"temperature":1.0,"reasoning_tokens":1996,"cache_read_input_tokens":896,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:09:50.295271+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement the training loop exactly as described with only the contrastive losses and check whether the trained model can output a forecast for any horizon; if the stated equations are the whole procedure, there is no mechanism to produce the prediction window, so the MSE numbers in the tables must come from an additional, undocumented component. Reproducing the tables only after adding an explicit forecasting head would settle the question.","supporting_citations":[{"cited_title":"Ts2vec: Towards universal representation of time series,","cited_arxiv_id":null,"evidence_quote":"Contrastive time series representation baseline whose multi-scale approach is extended by the time-domain module."},{"cited_title":"Timesnet: Temporal 2d-variation modeling for general time series analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the top-k frequency selection and masking idea for keeping prominent spectral components."}],"review_version":1}