{"id":"79c50823-61fd-40bf-8378-efa614c1b43b","arxiv_id":"2509.15394","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"VMDNet applies sample-wise variational mode decomposition with frequency embeddings, per-mode TCN decoders, and a bilevel Stackelberg search for mode count K and bandwidth penalty alpha, achieving best results on strongly periodic electricity demand but not uniformly on weakly periodic data.","lead":"VMDNet is a time-series forecasting framework that decomposes each short input window with variational mode decomposition before prediction, avoiding future information leaking into training data, and tunes the decomposition parameters with a hierarchical search. On a strongly periodic electricity demand dataset it beats seven recent baselines, while on a weakly periodic production dataset it is competitive but not uniformly best.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VMDNet's mode-specific branches assume stable mode identities across short windows; the paper acknowledges drift risk but never measures it, so the core architectural premise is unvalidated.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I find: VMDNet's mode-index-fixed branches require stable mode identities across short windows. This is not merely a theoretical nicety; the entire parallel-decoding design (Section 2.4) and the claimed benefit of 'mode independence' depend on it. The paper's own Section 1 raises the drift problem for prior work, but no experiment in the paper checks whether the selected K and alpha produce stable modes on the tested datasets. The frequency embedding is a partial mitigation, but it does not make the mode-specific convolutions or TCN branches frequency-conditional, so drift can still break alignment. The ablation's large drop when removing parallel decoding (Table 2) shows how critical this alignment is. I considered the abstract/dataset discrepancy as an alternative concern, but that is a presentation-level overclaim and is easily corrected; the mode-stability issue goes to whether the method's core mechanism works as intended. The proposed test—measuring mode-order stability and comparing against a frequency-reordered variant—would directly settle the concern. Since the reader already identified this and issued a conditional verdict, my read does not change the verdict.","tokens_in":8899,"tokens_out":9235,"duration_ms":106598,"concrete_test":"On the Electricity Demand training set, run VMD with the selected (K=4, alpha=5661) on every input window; record the K center frequencies per window. Sort modes by center frequency and compute the fraction of windows whose sorted order differs from the median order, and the overlap of the per-index center-frequency distributions. Then retrain VMDNet with modes explicitly reordered by center frequency per window before the branch assignment, using the same hyperparameters and seeds. If the reordered variant yields materially lower test MSE/MAE, the fixed-index design is misaligned; if it does not, the mode-stability concern is not empirically supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"VMDNet's central architectural choice is to assign each VMD mode index k its own TokenEmbed_k, FreqEmbed_k, and TCN_k (Eqs. 9–12). This is only meaningful if mode k consistently corresponds to the same spectral band across sliding windows. The introduction explicitly warns that short-window VMD 'often yields weakly separated modes whose identities drift across samples' (Section 1), yet no experiment quantifies center-frequency stability or permutation rates. If mode identities swap between windows, branch k receives a mixture of different spectral content; the frequency embedding (FreqEmbed_k) injects the instantaneous center frequency, but the mode-specific convolutional encoder and TCN decoder are not conditioned on frequency and cannot fully compensate. The ablation (Table 2) shows that parallel decoding is responsible for a large gain (MAE +0.044), so a breakdown in mode alignment could cancel the method's main benefit. The reported superiority on Electricity Demand might then reflect the validation-based hyperparameter selection or dataset-specific luck rather than the proposed alignment mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"VMDNet proposes a univariate time-series forecasting framework built on sample-wise Variational Mode Decomposition (VMD). Each lookback window is decomposed independently into K band-limited modes, so no future information enters the decomposition; modes are encoded with mode-specific token and frequency embeddings plus shared temporal/positional embeddings, decoded by parallel dilated TCNs, and fused by an MLP. A Stackelberg-inspired bilevel search over the VMD hyperparameters K and alpha is introduced, using a Forecastability Information Criterion (FIC) and a Mutual Information Criterion (MIC). Experiments are reported on two Polish electricity datasets (demand and production) at horizons 96, 192, and 336, with seven baselines and an ablation study. The central claims are that decomposition is leakage-free, that mode-specific parallel decoding improves forecasting, and that the bilevel search selects effective VMD parameters.","tokens_in":9198,"tokens_out":4420,"duration_ms":52560,"significance":"If the claims hold, VMDNet makes a useful contribution by addressing information leakage in decomposition-based forecasting and by giving each mode an explicit spectral identity. The paper ships code and data, reports results over five seeds, and provides ablations of the main components, which supports reproducibility. The leakage-free property is structurally guaranteed by Eqs. (4)-(7), and the idea of frequency-aware, mode-aligned decoding is well motivated. However, the central architectural assumption---that mode identities are stable enough across sliding windows to justify per-mode encoders and decoders---is not tested. The empirical evidence is also narrower than the abstract claims: only two datasets are used, and the weak-periodicity results do not show consistent superiority. The ablation confounds architectural capacity with the mode-separation mechanism. These issues need to be resolved before the central claims can be considered established.","major_comments":[{"comment":"The architecture assigns a dedicated TokenEmbed_k, FreqEmbed_k, and TCN_k to each VMD mode index. This is meaningful only if mode k consistently corresponds to the same spectral band across windows. The paper itself warns that short-window VMD 'often yields weakly separated modes whose identities drift across samples' (Section 1), but no experiment measures center-frequency stability, mode permutation rates, or alignment quality. If identities swap, branch k receives mixed spectral content and the claimed alignment benefit is not guaranteed. Please add a quantitative analysis of omega^(k,b) distributions across windows (e.g., pairwise permutation alignment or spectral centroid drift), and test a frequency-conditioned shared decoder as an alternative to per-mode branches. This is load-bearing because the ablation attributes a large gain to parallel decoding.","section":"Section 1; Eqs. (9)-(12); Table 2"},{"comment":"The ablation 'w/o parallel decoding' replaces K TCN branches with a single TCN operating on the mean of mode embeddings. This changes both the architecture and the parameter count, so the observed MAE drop (+0.044) could be due to reduced capacity rather than to the value of per-mode decoding. Similarly, 'w/o VMD' uses a single TCN on raw input with no capacity-matched control. To support the claim that mode-specific parallel decoding is responsible for the gain, please include a control with matched parameter count (e.g., a single TCN with K-times width, or a shared TCN applied to concatenated modes without mixing). Without this, the central design conclusion is confounded.","section":"Table 2; Section 3.5"},{"comment":"The abstract states that experiments are conducted on 'three widely used electricity demand datasets' and that VMDNet 'consistently outperforms state-of-the-art baselines.' The full text and Section 3.1 describe only two datasets, and Table 1 shows that on Electricity Production VMDNet is not consistently best: DLinear achieves a lower average MSE (0.596 vs. 0.665) and wins the MSE at horizons 192 and 336. Please correct the abstract and calibrate the claims to the actual evidence, e.g., 'achieves state-of-the-art when periodicity is strong and remains competitive under weak periodicity.'","section":"Abstract vs. Section 3.1 and Table 1"},{"comment":"The paper motivates FIC and MIC as forecasting-aware criteria for selecting K and alpha, but it is not shown that minimizing these criteria improves downstream forecasting. The final selection in Section 2.5 uses validation-set performance after multiple restarts, not FIC/MIC alone. The 'w/o bilevel searching' ablation replaces the proposed scheme with PSO minimizing envelope entropy, but does not isolate the contribution of the FIC/MIC objectives. Please report the correlation between FIC/MIC values and validation/test forecast error, and compare the final selected (K, alpha) against random search or a simple grid using the same validation re-ranking. This is needed to support the claim that the Stackelberg-inspired search is the relevant mechanism, rather than standard validation-based tuning.","section":"Section 2.5, Eqs. (13)-(16); Table 2"}],"minor_comments":[{"comment":"Column headers appear swapped: the reported values (0.156 MSE, 0.261 MAE for the full model) match Table 1's 'MSE' and 'MAE' order, not 'Test MAE' followed by 'Test MSE' as printed. Please correct.","section":"Table 2"},{"comment":"Eq. (13) uses AR order r in the FIC penalty, while Section 3.3 reports 'AR order p=2'. Please harmonize the notation.","section":"Eq. (13) and Section 3.3"},{"comment":"The schematic is difficult to read in the typeset version; the mode-branch structure and the fusion step should be more clearly labeled.","section":"Figure 1"},{"comment":"The variational objective is typeset with several garbled symbols; please ensure the Hilbert transform and demodulation terms are rendered cleanly.","section":"Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest in the full-text abstract about the two-dataset scope and the weaker weak-periodicity results, which helps. The main concern is not the leakage-free logic (which is sound) but whether the per-mode alignment premise is actually validated; the ablation capacity confound and the FIC/MIC validation gap are also fixable with additional experiments. I would not reject, but the current version overclaims in the abstract and leaves the central architectural assumption undemonstrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is clean: decompose each input window separately with VMD so no future information enters the decomposition, then give each mode its own embedding and TCN branch. The leakage-free claim is defensible, and the architecture is a sensible response to the known mode-drift problem in short-window decomposition.\n\nWhat is actually new: sample-wise VMD per sliding window is a real fix for the causality violation that plagues most decomposition-based forecasting. The parallel per-mode TCN branches are a natural fit for band-limited modes, and the ablation shows they matter—removing them costs 0.044 MAE. The bilevel Stackelberg-style search over K and alpha is a bit gimmicky as a framing, but it is a reasonable way to tune two coupled hyperparameters, and the comparison against PSO helps. Code is available, which is a plus.\n\nSoft spots, in order of size. First, the metadata abstract claims three datasets and consistent superiority, while the paper itself uses two datasets and the weak-periodicity results are mixed—DLinear beats VMDNet on the long horizon. That overstatement needs to be fixed. Second, the mode-stability stress test is legitimate: short-window VMD can produce drifting mode identities, and the paper never measures how stable the center frequencies are across windows. The frequency embedding is designed to compensate, but the ablation shows removing it costs little, which makes me wonder whether the per-mode branches are actually robust to identity swaps. I would ask for a simple experiment—track center frequencies across consecutive windows or compute a mode-permutation rate. Third, the authors tune hyperparameters on validation and apply minor adjustments per horizon, which introduces a mild selection-bias risk, though the five-seed reporting reduces the alarm.\n\nOverall, this is honest progress in a specialized subfield. It deserves a serious referee, but the authors should be pushed on the abstract and on measuring mode stability before publication. For someone doing decomposition-based forecasting, this is a useful paper to know.","headline":"A genuinely leakage-free VMD forecasting method with solid empirical support on one dataset, but an overstated metadata abstract and an unmeasured mode-stability assumption.","tokens_in":9651,"tokens_out":3872,"would_cite":false,"duration_ms":45506,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62M10","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"VMDNet claims a leakage-free, sample-wise variational mode decomposition network that outperforms seven state-of-the-art baselines on strongly periodic electricity demand forecasting.","keywords":["variational mode decomposition","time series forecasting","electricity demand","information leakage","temporal convolutional network","bilevel optimization","frequency-aware embedding","periodicity"],"falsifier":"On the Electricity Demand dataset, compute the standard deviation and label-switch rate of each mode's center frequency across all sliding windows. If modes frequently swap order or frequencies drift by more than the bandwidth spacing, VMDNet's fixed mode-aligned decoders are misaligned, and a permutation-invariant variant should match or beat it.","tokens_in":8798,"feed_emoji":"⚡","tokens_out":3395,"duration_ms":32969,"temperature":0.7,"pith_summary":"The paper introduces VMDNet, a forecasting model that first decomposes each input window separately with Variational Mode Decomposition, so no future information can enter the decomposition; then encodes each band-limited mode with a frequency-aware embedding and decodes it with its own temporal convolutional network; and finally selects VMD's two key hyperparameters through a Stackelberg-inspired bilevel search. The central claim is that when periodicity is strong, this causal, mode-disentangled design beats seven recent baselines on all tested horizons on the Electricity Demand dataset, and remains competitive when periodicity is weak. If true, it shows that explicitly separating oscillatory components before learning is a practical path to both accuracy and interpretability in short-term demand forecasting.","feed_headline":"Sample-wise VMD beats 7 baselines on electricity demand","feed_subtitle":"Decomposing each window separately keeps future data out, and per-mode TCNs capture strong periodic patterns.","key_machinery":"Variational Mode Decomposition (VMD) applied sample-wise: each input window x^(b) is decomposed into K band-limited modes u_k^(b) with center frequencies omega_k^(b) by solving a constrained variational problem via ADMM, using only past observations. The architecture then uses mode-specific token and frequency embeddings plus shared time and positional embeddings, K parallel causal TCN decoders with dilated convolutions, and a two-layer MLP fusion. A Stackelberg bilevel search treats K as leader and alpha as follower, using a Forecastability Information Criterion and a Mutual Information Criterion to pick the pair that suits the downstream predictor.","core_discovery":"On the paper's own terms, the discovery is that a leakage-free forecasting pipeline can be built by applying K-mode VMD independently to each sliding window of length P, embedding each mode's center frequency along with its amplitude, and running one lightweight dilated TCN per mode before fusing branch outputs. Because each window's decomposition uses only data available at the forecast time, the approach is causal by construction. Experimentally, VMDNet achieves the best MSE and MAE across horizons 96, 192, and 336 on the strongly periodic Electricity Demand dataset, and the best average MAE on the weakly periodic Electricity Production dataset, with ablations showing that removing VMD, pa","pith_inferences":["If mode center frequencies are stable across windows, the same architecture could be applied to multivariate and cross-domain forecasting by sharing mode branches across related series; the paper leaves this extension explicitly to future work.","A testable extension is to make the FIC/MIC selection differentiable or to warm-start VMD with center frequencies from the previous window, which could reduce the per-window computational cost while preserving leakage-freedom.","The strongest open risk is mode identity drift: since the paper itself notes that short-window VMD often yields weakly separated modes whose identities drift, VMDNet's fixed mode ordering could misalign across windows; measuring per-window frequency variance on the benchmark data would settle how often this happens."],"forward_implications":["VMDNet's sample-wise VMD guarantees that no future information leaks into the decomposition, so the reported gains on the Electricity Demand dataset are not an artifact of look-ahead bias.","Strongly periodic forecasting can be done with small causal TCN branches rather than large transformer backbones, since VMD already isolates the periodic structure.","The bilevel selection of K and alpha provides a forecasting-aware alternative to reconstruction-error or entropy-based VMD tuning.","The method transfers to weakly periodic data without collapsing, though linear baselines may win on long horizons."],"fun_headline_variants":["Causality-safe VMD forecasting beats 7 baselines","Leak-free VMD + per-mode TCNs sharpen demand forecasts","Sample-wise VMD kills leakage, boosts electricity prediction","Bilevel-tuned VMD trumps baselines on demand data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that applying K-mode VMD separately to each 336-point window yields stable, consistently ordered band-limited modes whose center frequencies do not drift or swap across windows, so that the same mode-specific embeddings and TCN branches stay aligned.","fun_headline_variants_meta":{"raw":{"variants":["Causality-safe VMD forecasting beats 7 baselines","Leak-free VMD + per-mode TCNs sharpen demand forecasts","Sample-wise VMD kills leakage, boosts electricity prediction","Bilevel-tuned VMD trumps baselines on demand data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000419,"raw_usage":{"total_tokens":1971,"prompt_tokens":700,"completion_tokens":1271,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":1211}},"tokens_in":444,"tokens_out":1271,"duration_ms":10109,"temperature":1.0,"reasoning_tokens":1211,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:12:37.979740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the Electricity Demand dataset, compute the standard deviation and label-switch rate of each mode's center frequency across all sliding windows. If modes frequently swap order or frequencies drift by more than the bandwidth spacing, VMDNet's fixed mode-aligned decoders are misaligned, and a permutation-invariant variant should match or beat it.","supporting_citations":[],"review_version":1}