{"id":"7bf430ea-512c-4194-8914-80a5e7b5bea7","arxiv_id":"2505.05738","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FOCUS forecasts multivariate time series by learning segment prototypes offline and attending to these prototypes online, achieving linear complexity and claimed state-of-the-art accuracy on seven benchmark datasets.","lead":"FOCUS is a forecasting model that clusters historical time-series segments into reusable prototypes offline, then uses those prototypes to replace quadratic self-attention with linear-time attention online. The paper reports top accuracy on 26 of 28 benchmark settings at far lower FLOPs and memory than eight comparison models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'state-of-the-art' claim depends on a benchmark that omits several standard strong baselines and tunes FOCUS's p,k by grid search while using untuned baseline configurations; the top-1 win rate may not survive a fair comparison.","rationale":"The reader's verdict was CONDITIONAL, with the weakest assumption identified as the representativeness/stationarity of offline prototypes and the correctness of Theorem 1. I agree that Theorem 1 is not a valid proof of the approximation claim: it invokes Johnson-Lindenstrauss with mismatched dimensions, applies a union bound that does not yield o(1) failure probability, and never bounds the actual ProtoAttn output because softmax is omitted. However, the theorem is not what the abstract or the headline contribution depends on for the empirical claim; the method's linear complexity follows directly from the fixed prototype count k, and the accuracy claim rests on Table III. The most load-bearing weakness for the central claim is therefore the benchmark construction: a comparison set that excludes several standard strong baselines and an asymmetric tuning protocol. This concern is concrete and testable: rerun with the omitted baselines and a symmetric hyperparameter budget. If FOCUS still wins the majority of settings, the empirical claim is much better supported; if not, the 'state-of-the-art' wording is unsupported. This does not contradict the reader's conditional verdict, so I leave the verdict unchanged, but the emphasis differs from the reader's stated weakest assumption.","tokens_in":21533,"tokens_out":4869,"duration_ms":52153,"concrete_test":"Run official implementations of iTransformer, TSMixer, TiDE, and ModernTCN on the same seven datasets at horizons 96 and 336, using the same train/validation/test splits and a comparable hyperparameter search budget for both FOCUS and each baseline; then count top-1 placements across the 28 MSE/MAE metric cells. If FOCUS wins fewer than half of the 28 cells, or if any omitted baseline wins more cells than FOCUS, the 'state-of-the-art' claim should be weakened. Reporting 5-seed means and standard deviations for FOCUS and the best baseline would also establish whether the margin is within noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is comparative: FOCUS 'achieves state-of-the-art accuracy' and 'ranking top-1 accuracy among the 8 models for comparison on 26 out of the 28 settings' (Abstract, Sec. I, Table III). For that claim to hold, the comparison set must be representative and the comparison fair. Both conditions are insecure in Sec. VIII-A. The paper states that 'we use the original configurations for all baseline models,' while for FOCUS, 'the segment length p and the number of prototypes k, were obtained through the grid-search method.' This asymmetry means FOCUS can benefit from per-dataset validation-based tuning that the baselines are not given, which can inflate apparent win rates even when the underlying method is not better. Moreover, the baseline set omits recent standard forecasting models that the paper itself cites in Sec. IX: iTransformer [55], TSMixer [29], TiDE [20], and ModernTCN [58]. These are widely used strong baselines in long-term multivariate forecasting, and several are specifically designed to be efficient and accurate linear/MLP alternatives to PatchTST. Without them, 'top-1 among the 8 models' is a claim only about the included subset. The theoretical Theorem 1 is indeed not correctly proved and does not bound the actual softmax attention output, and that should be fixed or removed. But the empirical fairness issue is the most load-bearing for the paper's headline 'accurate and efficient state-of-the-art' claim: if the missing baselines outperform FOCUS on a substantial share of settings, the central claim fails even though the method itself may be sound and the complexity analysis correct.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FOCUS, a two-phase multivariate time series forecasting method. In an offline phase, it segments the training series and clusters the segments into k prototypes using Euclidean distance plus Pearson correlation. In the online phase, input segments are assigned to the nearest prototypes, and a module called ProtoAttn computes attention between embedded prototypes and input segments, aiming to model long-range dependencies in linear time. A dual-branch architecture extracts temporal and entity features and fuses them through a readout/gating mechanism. The paper claims state-of-the-art accuracy (top-1 among 8 compared models on 26 of 28 settings) with reduced FLOPs and memory, and it reports ablations, parameter studies, outlier-robustness experiments, and case studies on seven benchmark datasets.","tokens_in":21789,"tokens_out":8551,"duration_ms":93580,"significance":"If the empirical claims hold, FOCUS would provide a practical linear-complexity alternative to quadratic self-attention for long-horizon multivariate forecasting, with the additional appeal of interpretable offline-discovered prototypes. The paper includes public code, reports experiments on seven standard datasets, and provides extensive ablations and sensitivity analyses, which are strengths. However, the theoretical justification in Sec. VI-B is not valid as written, and the headline accuracy claim rests on a benchmark that omits several standard recent baselines and tunes FOCUS's hyperparameters while using untuned baseline configurations. The central methodological idea is defensible, but the strongest claims are not yet fully supported.","major_comments":[{"comment":"The theorem does not prove what the paper claims. The section is titled 'Approximation Analysis of ProtoAttn', but the statement and proof only concern a low-rank approximation \\tilde P = AC of the input matrix P; the actual ProtoAttn output in Eq. (18), namely A softmax(C_Q K^T / sqrt(d_k)) V, involves a softmax nonlinearity and value projections, and none of these objects appear in the proof. The dimensions are also inconsistent: P is in R^{l x p}, w is drawn from a column of W_Q W_K^T in R^{d x d}, so P w^T is undefined unless p = d, and \\tilde P = AC with C in R^{k x d} has dimension l x d, not l x p. Equation (24) is not a valid Johnson-Lindenstrauss statement as written: for R in R^{k x d}, x in R^k, y in R^k, the products x R^T R y^T and x y^T have incompatible shapes, and the bound k = 5 log r / (epsilon^2 - epsilon^3) is not derived. Because the abstract and Sec. VI-B use this theorem to justify that linear-complexity proto-attention approximates self-attention, this is a load-bearing theoretical gap. The authors should either provide a correct, properly dimensioned argument that bounds the attention output, or remove the theorem and support the approximation claim empirically.","section":"Sec. VI-B, Theorem 1 and Eqs. (20)-(31)"},{"comment":"The headline claims of 'state-of-the-art accuracy' and 'top-1 among the 8 models on 26 out of 28 settings' are not yet supported by a fair comparison. The baseline set omits iTransformer [55], TSMixer [29], TiDE [20], and ModernTCN [58], all of which are cited in Sec. IX and are standard strong models in long-term multivariate forecasting; a 26/28 win rate over the included subset does not establish state-of-the-art performance. In addition, Sec. VIII-A states that 'we use the original configurations for all baseline models', while for FOCUS the segment length p and prototype count k were obtained by grid search, so FOCUS receives per-dataset validation-based tuning that the baselines do not. To support the accuracy claim, the authors should add the missing baselines or explicitly restrict the claim to the compared subset, and should show that the win rate is robust under a fairer protocol, such as fixed p and k for FOCUS or validation-based tuning for all models.","section":"Sec. VIII-A, Table III"},{"comment":"The paper's own generalization study admits that test-set segments can be unseen relative to the offline prototypes, yet the evidence is only qualitative: Fig. 9 shows a single test instance, and no quantitative measure of prototype coverage or approximation error on test data is reported. Since the method's accuracy relative to full self-attention depends on the premise that a small fixed set of offline prototypes remains representative of online inputs, the authors should report a quantitative evaluation of the prototype approximation error as a function of k and of distribution shift, perhaps by measuring the reconstruction error of test segments assigned to prototypes or the attention-output discrepancy against full attention.","section":"Sec. VIII-D"}],"minor_comments":[{"comment":"Algorithm 2 line 6 sets V <- P W^Q, but Eq. (14) defines V = P W^V; this is likely a typo and should be corrected.","section":"Sec. VI-A, Eq. (14) and Algorithm 2"},{"comment":"The TimesNet entries for PEMS08 (0.66 at horizon 96 and 0.07 at horizon 336) appear to be missing a decimal point or otherwise misprinted; please check and correct these values.","section":"Table III"},{"comment":"The text states that the ablation study was conducted on the PEMS04 dataset, but Table IV reports results for PEMS08 and Electricity; the text and table should be reconciled.","section":"Sec. VIII-C, Table IV"},{"comment":"The assignment step uses a discrete argmin while the prototype update uses gradient descent; please state explicitly that this is an alternating Lloyd-style procedure and clarify whether Eq. (11) is intended as a bi-level optimization objective.","section":"Sec. V, Eq. (6) and Algorithm 1"},{"comment":"The notation in Definition 2 uses D both for the dataset and for the number of entities, which conflicts with Table I where N denotes the number of entities; please use a distinct symbol for the entity count.","section":"Sec. II, Definition 2 and Table I"},{"comment":"There are several typos and awkward phrasings, including 'expanding this across all query centroids yields will get us' in Sec. VI-A, 'foreacasting' in Sec. VII, and 'even as the perturbation ratio improves' in Sec. VIII-E; a careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible empirical contribution and a substantial amount of experimental work, but the theoretical section appears not to have been checked for dimensional consistency, and the comparison protocol would not meet the benchmarking standard expected by a rigorous journal. I would support a major revision that fixes or removes the theorem and expands or reframes the empirical claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this for the method, not the theorem. FOCUS's idea—offline segment clustering into prototypes, then using those prototypes as queries in attention—is genuinely new relative to the cited literature, and the complexity story is clean: a fixed prototype count k makes online cost O(kl) instead of O(l^2). The paper also does some things right: code is public, ablations isolate contributions, and Sec. VIII-D/E test generalization to unseen segments and outliers, which is more than most forecasting papers bother with.\n\nNow the soft spots. Theorem 1 is not a proof. The JL-lemma claim in Eq. (24) has mismatched dimensions, and the theorem bounds an approximation of P w^T, not the softmax attention output that ProtoAttn actually computes. So the theoretical justification for \"approximates self-attention\" is unsupported. That is a real flaw in a central claim, and the authors should either fix it or delete it and present linear complexity as an architectural fact, which it already is.\n\nThe empirical claim is shakier than the abstract implies. They grid-search p and k for FOCUS per dataset while using \"original configurations\" for all baselines—an asymmetry that can inflate win rates even without any real accuracy advantage. The comparison set also omits iTransformer, TSMixer, TiDE, and ModernTCN, all cited in the paper and all standard in long-horizon forecasting. So the \"26/28 top-1\" headline is only about the included subset. There are no error bars, and horizons are limited to 96 and 336, so the \"long\" part of \"long-range\" is not fully tested.\n\nNone of this makes the method obviously wrong. The ablations suggest each component contributes, the offline prototypes are trained on the same training split used for supervised learning (standard, not circular), and the efficiency numbers in Fig. 6 look credible. It is a plausible forecasting architecture that currently oversells itself.\n\nWho should read it: people working on efficient long-sequence forecasting, and anyone interested in how benchmark choices shape SOTA claims. It deserves peer review, not desk rejection. A good reviewer would send it back for major revision: fix or remove the theorem, rerun against the missing baselines with matched tuning budgets, report variance, and add horizons 192 and 720. If those come back clean, FOCUS becomes a solid contribution. My recommendation: send it to review with those demands.","headline":"Read it for the method, not the theorem: the prototype-attention idea is new and the complexity story is clean, but the approximation proof doesn't hold and the SOTA claim rests on an unfair benchmark.","tokens_in":22403,"tokens_out":4018,"would_cite":false,"duration_ms":44112,"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":"A fixed set of offline-clustered segment prototypes can replace all-pairs attention in multivariate forecasting, giving linear-time inference and top-1 accuracy in 26 of 28 settings.","keywords":["multivariate time series forecasting","long-range dependencies","offline clustering","time series prototypes","linear-complexity attention","low-rank approximation","dual-branch architecture","efficient inference"],"falsifier":"Run FOCUS on a test set deliberately composed of segment shapes absent from training, such as synthetic ramps with steepness beyond any prototype or event patterns created by concatenating unseen motifs, and record whether MSE rises sharply as the fraction of unseen segments increases; if error degrades much faster than for a full self-attention model, the fixed-prototype low-rank assumption is the point of failure.","tokens_in":21262,"feed_emoji":"📈","tokens_out":11532,"duration_ms":108060,"temperature":0.7,"pith_summary":"FOCUS claims that the long-range dependencies a multivariate forecaster needs can be captured by a small fixed set of segment prototypes discovered once offline, so online modeling only has to relate incoming segments to these prototypes rather than to each other. If true, this replaces the quadratic cost of self-attention with linear cost while preserving, and in the paper's experiments improving, accuracy. The core evidence is ProtoAttn, which computes attention weights per prototype and maps the results back to segments with an assignment matrix, together with a low-rank approximation theorem showing that the gap from full self-attention stays small when the segment matrix has rank no larger than the prototype count. The paper reports top-1 results on 26 of 28 benchmark settings and lower FLOPs and peak memory than the compared models.","feed_headline":"Offline clusters cut forecast attention to linear time","feed_subtitle":"FOCUS learns a fixed set of segment patterns offline, then models only input-to-pattern dependencies.","key_machinery":"The load-bearing object is the prototype set $C = \\{c_j\\} \\subset \\mathbb{R}^p$, one representative segment pattern per cluster, obtained offline by minimizing the weighted loss $L = L_{\\text{rec}} + \\alpha L_{\\text{corr}}$ that combines Euclidean reconstruction error with negative Pearson correlation to each assigned segment. Online, the assignment matrix $A \\in \\mathbb{R}^{l\\times k}$ links each input segment to its nearest prototype, and ProtoAttn computes the attention map between embedded prototypes and input segments as $A\\,\\mathrm{softmax}(C_Q K^\\top/\\sqrt{d})\\,V$, so all segments in the same bucket share the same attention weights. Runtime is linear in $l$ because $k$ is constant, and Theorem 1, built on the Johnson–Lindenstrauss lemma, is the argument that this prototype-based computation is close to full self-attention when the input segment matrix is low rank. The dual-branch extractor and parallel fusion module push the same mechanism along time and entity dimensions.","core_discovery":"The central claim is that a fixed set of $k$ offline-clustered segment prototypes can act as the query side of attention: each incoming segment is assigned to its nearest prototype to form an assignment matrix $A$, and the attention output is $A\\,\\mathrm{softmax}(C_Q K^\\top/\\sqrt{d_k})\\,V$, where $C_Q$ are embedded prototypes and $K,V$ come from the input segments. Because $k$ is fixed and independent of input length, complexity becomes $O(kl)$ in the number of segments $l$. Theorem 1 argues that if the segment matrix $P$ has rank at most $r$, the prototype-based low-rank approximation $AC$ approximates $P$ under projection by any column of $W_Q W_K^\\top$ up to relative error $\\epsilon$ with high probability, using $k = O(\\log r / \\epsilon^2)$ prototypes. The same mechanism runs separately along temporal and entity dimensions in a dual-branch extractor, and a parallel fusion module with $m$ fixed readout queries combines the two feature sets before projecting to the forecast. The paper claims this design is both more accurate and cheaper than all-pairs attention, and in the experiments FOCUS ranks first in 26 of 28 settings.","pith_inferences":["A useful diagnostic the paper does not report would be to track each test segment's distance to its nearest prototype; if forecasting error climbs as that distance grows, practitioners would have a direct signal for when the offline prototype set needs refreshing.","The benefit of a fixed prototype set is regime-dependent: when the lookback is short enough that the prototype count $k$ is comparable to the segment count $l$, the constant overhead of $k$ prototypes can dominate and the linear-time advantage largely disappears.","The assignment matrix doubles as an event label for every moment in history, so the same mechanism could power downstream anomaly detection or explanation tools without additional supervision.","A natural extension would be a slowly updated prototype set learned online; the paper only studies fixed offline prototypes, but its own non-stationarity experiment suggests a moderate refresh could extend the method's validity without losing the linear-complexity guarantee."],"forward_implications":["Forecasting on very long histories becomes practical: doubling the lookback doubles rather than quadruples compute, so the model can exploit the accuracy gains of longer context windows that the paper observes on PEMS08.","The expensive pattern-discovery cost is paid once per dataset offline and then amortized across all online instances and entities, making per-forecast inference cheap.","The learned dependency map between prototypes is interpretable: the paper's case study shows the model finding long-range couplings such as morning traffic rises depending on the previous night's decline.","On resource-constrained devices, the lower FLOPs and peak memory give an accuracy-per-compute advantage over self-attention models, which the paper positions as the practical motivation for the design."],"supporting_citations":[{"why":"Defines the patch-based Transformer baseline and the all-pairs segment-dependency paradigm that FOCUS replaces, and is the strongest segmentation-based comparison in the experiments.","marker":"[62]"},{"why":"Supplies the cross-dimension Transformer baseline that motivates FOCUS's dual temporal/entity branches and is a direct accuracy target.","marker":"[89]"},{"why":"Represents the graph-neural-network family of spatiotemporal baselines in the eight-model comparison.","marker":"[79]"},{"why":"Represents the adaptive-graph dilated-convolution baseline family in the comparison.","marker":"[80]"},{"why":"Supplies a temporal-2D-variation baseline and is one of the models FOCUS must beat on every dataset.","marker":"[75]"},{"why":"LightCTS is the lightweight correlated-series baseline whose efficiency orientation is closest to FOCUS among the comparison models.","marker":"[47]"},{"why":"DLinear is the pure-linear baseline that tests whether the nonlinear prototype dependency model earns its accuracy advantage.","marker":"[87]"},{"why":"Provides the Johnson–Lindenstrauss low-rank projection bound that Theorem 1 uses to show prototype-based attention approximates full self-attention.","marker":"[74]"}],"fun_headline_variants":["FOCUS uses cluster prototypes for linear-time forecasting","Offline clustering makes forecast attention scale linearly","Clustered prototypes replace pairwise attention in forecasts","Prototype-based attention cuts forecasting cost to linear","Segment clustering yields accurate linear-time forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a small fixed set of offline-discovered segment patterns can faithfully represent every segment that appears online; the test data's patterns do not drift away from the training patterns, and the number of distinct recurring patterns does not grow with sequence length.","fun_headline_variants_meta":{"raw":{"variants":["FOCUS uses cluster prototypes for linear-time forecasting","Offline clustering makes forecast attention scale linearly","Clustered prototypes replace pairwise attention in forecasts","Prototype-based attention cuts forecasting cost to linear","Segment clustering yields accurate linear-time forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1356,"prompt_tokens":1004,"completion_tokens":352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":283}},"tokens_in":620,"tokens_out":352,"duration_ms":4205,"temperature":1.0,"reasoning_tokens":283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:58:11.940092+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FOCUS on a test set deliberately composed of segment shapes absent from training, such as synthetic ramps with steepness beyond any prototype or event patterns created by concatenating unseen motifs, and record whether MSE rises sharply as the fraction of unseen segments increases; if error degrades much faster than for a full self-attention model, the fixed-prototype low-rank assumption is the point of failure.","supporting_citations":[{"cited_title":"A time series is worth 64 words: Long-term forecasting with transformers","cited_arxiv_id":null,"evidence_quote":"Defines the patch-based Transformer baseline and the all-pairs segment-dependency paradigm that FOCUS replaces, and is the strongest segmentation-based comparison in the experiments."},{"cited_title":"Connecting the dots: Multivariate time series forecasting with graph neural networks","cited_arxiv_id":null,"evidence_quote":"Represents the graph-neural-network family of spatiotemporal baselines in the eight-model comparison."},{"cited_title":"Graph wavenet for deep spatial-temporal graph modeling","cited_arxiv_id":null,"evidence_quote":"Represents the adaptive-graph dilated-convolution baseline family in the comparison."},{"cited_title":"Timesnet: Temporal 2d-variation modeling for general time series analysis","cited_arxiv_id":null,"evidence_quote":"Supplies a temporal-2D-variation baseline and is one of the models FOCUS must beat on every dataset."},{"cited_title":"Lightcts: A lightweight framework for correlated time series forecasting.Proceedings of the ACM on Management of Data, 1(2):1– 26, 2023","cited_arxiv_id":null,"evidence_quote":"LightCTS is the lightweight correlated-series baseline whose efficiency orientation is closest to FOCUS among the comparison models."}],"review_version":1}