{"id":"84ec9264-cb07-4305-aeee-388c6f5e2824","arxiv_id":"2505.12917","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"TQNet improves multivariate time series forecasting by using periodically shifted learnable query vectors in attention, achieving top-2 results in 22 of 24 benchmark metrics.","lead":"A new attention mechanism for multivariate time series forecasting uses periodically shifted learnable vectors as queries to capture stable inter-variable correlations, while raw inputs provide keys and values. The resulting TQNet matches or beats prior models on 12 benchmark datasets with a lightweight architecture.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim rests on an incomplete comparison set: Table 2 omits several strong channel-dependence models cited in Related Work, and the 22/24 top-2 count mixes outright wins with second places.","rationale":"I read the paper as an empirical architecture claim, and much of the internal evidence is solid: the architecture is simple and specified, the code is released, the ablation in Table 4 isolates the TQ contribution, Table 9 shows low variance across seeds and learning rates, and the periodic-shift mechanism is a legitimate inductive bias. The reader's weakest_assumption identifies W-sensitivity, which is real and explicitly acknowledged in Section 5, but it is a disclosed limitation shared with CycleNet rather than the most direct threat to the central claim. The most load-bearing condition for 'overall state-of-the-art accuracy' is that the comparison set is representative of currently competitive models. Table 2 includes several strong baselines, but the Related Work section names numerous recent channel-dependence methods that are absent from the benchmark, and the top-2 counting convention makes the headline easier to satisfy than a per-metric win analysis would. A focused rerun with the omitted models and a per-metric ranking would settle whether TQNet is genuinely SOTA or merely competitive with the nine chosen baselines. Because the reader's verdict is already CONDITIONAL and this concern supports that assessment, I recommend no change to the verdict.","tokens_in":29759,"tokens_out":8866,"duration_ms":98786,"concrete_test":"Rerun Table 2 with the same data splits, L=96, horizons, and metric definitions, using the released code and the official implementations of at least SAMformer, SOFTS, TimeMixer, CARD, and UniTST; then report per-metric wins and average rank across the 24 MSE/MAE entries instead of a top-2 count. If any added model is best on at least 3 of the 12 datasets, or if TQNet no longer has the best average rank across the 24 metrics, the 'overall state-of-the-art' claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central claim is not the periodic-W prior (a disclosed limitation, Section 5), but the completeness of the SOTA comparison in Table 2. The abstract and Section 4.2 claim 'overall state-of-the-art accuracy' across 12 datasets, and the evidence is a 22/24 top-2 count against nine baselines: TimeXer, CycleNet, iTransformer, MSGNet, TimesNet, PatchTST, Crossformer, DLinear, and SCINet. Related Work itself cites strong 2024-2025 channel-dependence models that are not benchmarked, including SAMformer, CARD, UniTST, SOFTS, TimeMixer, DUET, and TimeBridge, several of which target exactly the high-dimensional Electricity/Traffic/PEMS settings where TQNet claims its largest wins. A top-2 count also mixes outright wins with second places, and no baseline variance or confidence intervals are reported; for example, Weather MSE 0.242 vs. 0.241 and the ETTh2/ETTm2 rows show that TQNet is not the clear winner on several metrics. If an omitted model wins a substantial share of the 24 metrics, the headline 'state-of-the-art' claim fails. This is a falsifiable comparison-protocol risk, not a disagreement with consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"TQNet is a multivariate time series forecaster built from a single multi-head attention layer whose queries are periodically shifted learnable vectors (one period length W per dataset) and whose keys and values are the raw input channels, followed by a two-layer MLP, an output projection, and optional reversible instance normalization. The authors argue that the shared periodic queries average over W-spaced samples and thereby capture stable cross-channel correlations while preserving local, sample-specific information in the keys and values. The paper reports experiments on 12 datasets with horizons {96,192,336,720} and {12,24,48,96} for PEMS, claims top-2 performance on 22/24 dataset-metric pairs in Table 2, and provides efficiency comparisons. Appendices contain pseudocode, full results, ablation and integration studies, sensitivity to W, robustness over seeds and learning rates, and a theoretical analysis in Appendix A.4. Section 5 candidly lists limitations, including dependence on dataset periodicity and diminishing benefits of multivariate modeling for long look-back windows.","tokens_in":30056,"tokens_out":7904,"duration_ms":86175,"significance":"If the empirical claims are sustained, TQNet would be a valuable contribution: a simple, lightweight architecture with a single attention layer and an MLP achieving competitive accuracy across 12 standard benchmarks is a useful data point for the channel-dependence versus channel-independence debate. The paper's strengths include released code, a fixed-seed experimental protocol, a robustness table over multiple seeds and learning rates, and extensive ablation/integration studies that probe the contribution of the temporal-query mechanism. However, the headline state-of-the-art claim is not yet established: Table 2 supports it only through a top-2 count against baselines whose numbers are taken from prior papers, several strong channel-dependence models cited in Related Work are not benchmarked, and no variance or significance information is given for the baseline comparisons. The theoretical analysis in Appendix A.4 is presented as a proof but contains informal and partly incorrect steps. These issues are fixable by additional benchmarking and by tempering the claims, so the paper warrants a major revision rather than rejection.","major_comments":[{"comment":"The claim that TQNet achieves 'overall state-of-the-art performance' is supported only by a top-2 count (22/24) in Table 2. This is a weak basis for the headline claim, because top-2 conflates outright wins with second places and because the table omits several channel-dependence models that the paper itself cites in Related Work, including SAMformer, CARD, UniTST, SOFTS, TimeMixer, DUET, and TimeBridge, several of which target the same high-dimensional Electricity/Traffic/PEMS settings where TQNet reports its largest gains. The caption of Table 5 also indicates that baseline numbers are sourced from TimeXer, iTransformer, and CycleNet rather than rerun under a single protocol, and no standard deviations or significance tests are reported; for example, on Weather the MSE difference is 0.242 versus 0.241, and on ETTh2 TimeXer reports 0.368 versus TQNet's 0.378. I request either a complete benchmark under a unified protocol with variance reporting for all methods, or a substantial softening of the 'state-of-the-art' wording to 'competitive with state-of-the-art'.","section":"Section 4.2, Table 2"},{"comment":"The theoretical analysis does not establish the claimed averaging behavior. Eq. (10) states that training maximizes the attention output, but this does not imply that the query correlation matrix equals the key correlation matrix as asserted in Eq. (11); attention weights can be large without Q and K having identical cross-channel correlation structure. Eq. (13) further assumes that the correlation learned by the shared periodic query equals the average of the per-sample key correlations, although correlation is not a linear operation and no derivation is given. Because this appendix is presented as the explanation for why the temporal-query mechanism improves robustness, it should either be replaced by a rigorous statement or explicitly labeled as heuristic intuition.","section":"Appendix A.4, Eqs. (10)-(13)"},{"comment":"The design claim that the default (Q=TQ,K=Raw) configuration 'yields the best performance' is contradicted by the Traffic row of Table 3: (Q=TQ,K=Raw) gives MSE 0.445, while (Q=Raw,K=Raw) gives MSE 0.426. The default setting does improve MAE on Traffic (0.276 versus 0.279), but the text does not state that the conclusion is based on MAE rather than MSE. Since the paper presents these ablations as evidence for the central design choice, the authors should reconcile this exception or report an aggregate statistic, such as mean rank or wins across the seven datasets.","section":"Table 3"}],"minor_comments":[{"comment":"The limitations section candidly acknowledges the dependence on a known period length W and the risk of over-strong multivariate modeling when correlations are weak; these caveats should be reflected in the abstract and conclusion, which currently assert state-of-the-art performance without qualification.","section":"Section 5"},{"comment":"The sensitivity study for W is informative but is conducted only on Electricity; given that W is a central hyperparameter, a small additional table showing W sensitivity on one or two further datasets would materially strengthen the paper.","section":"Section 4.3, Figure 6"},{"comment":"The procedure for selecting W is delegated to CycleNet's appendix and a provided ACF code snippet; the paper should state the ACF criterion used in practice, such as the chosen lag or significance threshold, so that the method is self-contained.","section":"Appendix A.3"},{"comment":"The statement that TQNet has 'near-linear computational overhead' should be explicitly empirical: the attention mechanism is quadratic in the number of channels in general, and Figure 7 shows wall-clock training time on one GPU rather than asymptotic complexity. The y-axis label 'Training Time (seconds/epochs)' should also read 'seconds per epoch'.","section":"Section 4.2, Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the comparison protocol. The paper's own Related Work lists several recent channel-dependence models that are not included in Table 2, and the top-2 count with baselines sourced from earlier papers is not strong enough for the 'state-of-the-art' claim. The core idea is simple, plausible, and supported by useful ablations, so I believe the paper is publishable after a major revision that either adds the missing benchmarks under a unified protocol or carefully rewrites the claims. I do not see evidence of fabrication or ethical issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on TQNet.\n\nThe genuinely new piece is the Temporal Query: learnable vectors, shifted by the dataset's period, used as attention queries while keys and values stay on the raw input. It's a small idea, clearly explained, and the paper does the right legwork to show it matters. Ablations show the Q=TQ/K=Raw configuration beats both pure-local and pure-global versions, and the integration study shows the trick transfers to iTransformer, PatchTST, and DLinear. The code is out, 12 datasets, multiple horizons, and the efficiency figures are honest about the architecture's cost.\n\nThe soft spots are all around the strength of the claim. The abstract says 'state-of-the-art accuracy,' but Table 2 only pits TQNet against nine baselines, several of which are older. The Related Work section itself names SAMformer, CARD, UniTST, SOFTS, DUET, and TimeBridge — strong channel-dependence models that are never benchmarked. The 22/24 top-2 count is a rank aggregate, not a win count; on ETTh2 and Weather MSE TQNet is second, and the gaps are often small. Without baseline variance or confidence intervals, 'overall SOTA' is more of a slogan than a measurement. The theoretical analysis in Appendix A.4 is also informal — it reads as a heuristic about averaging correlations, not as a proof. And the W sensitivity is a real limitation, though the paper discloses it and shows graceful degradation.\n\nNone of that sinks the paper. The central mechanism is sound, the empirical story is consistent, and the limitations section is refreshingly candid. I think a serious referee could reasonably ask for (a) a broader baseline table or a softer claim, and (b) either a rigorous version of the A.4 argument or an explicit statement that it's intuition only. The core idea is portable and likely to be useful beyond this specific architecture.\n\nMy recommendation: send it out. It deserves referee time. The comparison protocol needs tightening before the strongest claim is accepted, but the contribution is real.\n\nFor a reading group, I'd say yes — it's a good example of a simple, well-tested modification to attention that people will want to know about.","headline":"Solid, simple idea with a portable query trick; the SOTA claim overreaches relative to the baseline set.","tokens_in":30528,"tokens_out":4450,"would_cite":true,"duration_ms":41579,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Periodically shifted learnable queries let a single attention layer match much larger multivariate forecasting models.","keywords":["Temporal Query","multivariate time series forecasting","channel dependence","attention mechanism","periodic shifting","learnable queries","inter-variable correlation","lightweight model"],"falsifier":"Take a dataset with no clear periodicity, or deliberately set $W$ wrong (e.g., $W=167$ on hourly Electricity), and run TQNet against the same baselines; the paper's Figure 6 already shows error rising in that case, so a systematic sweep over aperiodic series would reveal whether the claimed gains depend entirely on the period prior. If TQNet still matches or beats channel-independent baselines without a correct $W$, the global-correlation-averaging explanation would need revision; if it collapses, the prior is load-bearing.","tokens_in":29570,"feed_emoji":"📈","tokens_out":5310,"duration_ms":53377,"temperature":0.7,"pith_summary":"This paper claims that the bottleneck in multivariate time series forecasting is not model size but the instability of sample-level inter-variable correlations, and that a single, cheap fix — periodically shifted shared queries in attention — recovers stable global correlations. The proposed TQNet uses one attention layer and a shallow MLP, yet reaches top-two accuracy in 22 of 24 benchmark error metrics across 12 real-world datasets. If the claim holds, robust correlation modeling, rather than depth or channel-mixing capacity, is what drives accuracy in multichannel forecasting.","feed_headline":"Periodic shared queries beat larger forecasters on 12 datasets","feed_subtitle":"A one-layer attention model captures stable cross-variable correlations at near-linear cost.","key_machinery":"The Temporal Query (TQ) is a learnable parameter matrix $\\theta_{\\text{TQ}} \\in \\mathbb{R}^{C\\times W}$, initialized to zero. For a sample at time $t$, the segment $\\theta_{\\text{TQ}}^{t,L}$ of length $L$ is extracted starting at index $t \\bmod W$, so samples spaced $W$ apart reuse identical queries. In attention, these queries meet keys and values taken from the raw input $X_t$, and the softmax scores $QK^\\top/\\sqrt{L}$ mix the global prior with local evidence. The load-bearing identity is Equation (13): the correlation learned by a query equals the average of the sample-level correlations at that period phase, which is what makes the model robust to non-stationary disturbances.","core_discovery":"The paper's central discovery is that replacing the queries of self-attention with a learnable matrix $\\theta_{\\text{TQ}}\\in\\mathbb{R}^{C\\times W}$, cyclically selected by $t \\bmod W$, makes attention compute inter-variable correlations that are averaged over all periods aligned at that phase, while keys and values from the raw input preserve local, sample-specific detail. This design matches the dataset-level correlation matrix more closely than either pure self-attention (raw queries and keys) or a purely global query-key pair, as shown by the ablation in Table 3. The averaging identity in Appendix A.4 states $\\operatorname{Corr}(Q_i) \\approx \\frac{1}{N}\\sum_{n=0}^{N-1}\\operatorname{Corr}(K_{i+nW})$, which is the mechanism that suppresses noise, outliers, and missing values. The paper then demonstrates that this single mechanism, inside an otherwise minimal network, yields state-of-the-art forecasts on 12 datasets and improves existing models when plugged into them.","pith_inferences":["The periodic-shared query is effectively a learnable seasonal smoothing of the channel-correlation matrix; one could extend it to multiple periods (e.g., daily plus weekly) with an ensemble of TQ vectors and expect further gains on multi-periodic datasets.","If $W$ is unknown, an adaptive variant that learns or infers the period from autocorrelation or gradient descent would remove the main practical obstacle; the paper currently leaves $W$ as a user-supplied prior.","Setting $W=1$ degrades TQ to a per-channel identifier, which suggests the method implicitly learns channel embeddings; this connection to channel-identity models could be explored explicitly."],"forward_implications":["TQNet reaches top-two accuracy in 22 of 24 error metrics across the twelve datasets, so a one-layer attention plus MLP can be a competitive default multivariate forecaster.","Plugging the TQ query into iTransformer, PatchTST, and DLinear improves all three, so the technique transfers across channel-mixing and channel-independent architectures.","Training time stays close to DLinear's even with 862 channels, making the approach practical in high-dimensional settings.","The benefit of multivariate modeling shrinks as the look-back window grows, so the technique is most valuable when historical data is short."],"supporting_citations":[{"why":"Supplies the periodicity prior and the $W$ selection guidance that TQNet adopts for its periodic shifting interval.","marker":"(Lin et al., 2024b)"},{"why":"Provides the channel-mixing attention baseline and the instance normalization scheme TQNet builds on, plus baseline numbers used for comparison.","marker":"(Liu et al., 2024c)"},{"why":"Is the strongest attention-based competitor and the source of reproduced baseline results, including the multivariate-to-univariate setting.","marker":"(Wang et al., 2024c)"},{"why":"Defines the linear-model efficiency target and serves as a channel-independent baseline and an integration host for the TQ technique.","marker":"(Zeng et al., 2023)"},{"why":"Is the channel-independent baseline that TQNet improves upon when augmented with a TQ-based attention module.","marker":"(Nie et al., 2023)"},{"why":"Is cited as the computational method, autocorrelation function, for choosing the periodic length $W$.","marker":"(Madsen, 2007)"}],"fun_headline_variants":["TQNet: one attention layer beats deep forecasters on 12 sets","One attention layer with periodic queries wins on 12 datasets","Cyclic attention queries nail cross-variable patterns fast","Shifted queries make attention model dataset-wide patterns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The technique presumes each dataset has one stable dominant period $W$, and that the user sets $W$ to that period; nothing in the model discovers or adapts $W$ automatically.","fun_headline_variants_meta":{"raw":{"variants":["TQNet: one attention layer beats deep forecasters on 12 sets","One attention layer with periodic queries wins on 12 datasets","Cyclic attention queries nail cross-variable patterns fast","Shifted queries make attention model dataset-wide patterns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3156,"prompt_tokens":939,"completion_tokens":2217,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":2149}},"tokens_in":555,"tokens_out":2217,"duration_ms":17454,"temperature":1.0,"reasoning_tokens":2149,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:23:05.987880+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset with no clear periodicity, or deliberately set $W$ wrong (e.g., $W=167$ on hourly Electricity), and run TQNet against the same baselines; the paper's Figure 6 already shows error rising in that case, so a systematic sweep over aperiodic series would reveal whether the claimed gains depend entirely on the period prior. If TQNet still matches or beats channel-independent baselines without a correct $W$, the global-correlation-averaging explanation would need revision; if it collapses, the prior is load-bearing.","supporting_citations":[{"cited_title":"Time series analysis","cited_arxiv_id":null,"evidence_quote":"Is cited as the computational method, autocorrelation function, for choosing the periodic length $W$."}],"review_version":1}