{"id":"a0e969f0-c9f8-4dfa-93aa-aba71783cb74","arxiv_id":"2505.10172","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A parameter-light adaptive linear model (ALinear) outperforms larger baselines on long-horizon univariate forecasting benchmarks while using under 1% of their parameters, but the efficiency comparison rests on questionable baseline parameter counts.","lead":"This paper introduces ALinear, a linear model with around 0.2 million parameters that reportedly matches or beats much larger time series models on seven benchmarks. It argues that scaling up parameter counts is unnecessary for forecasting and proposes an adaptive mechanism that reweights trend and seasonal components by forecast horizon.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline parameter counts for TimesNet appear inflated by roughly 1000x; the claim of using less than 1% of large-model parameters rests on this count.","rationale":"The reader's weakest-assumption analysis identified the same load-bearing issue: the reported TimesNet parameter count appears inflated and undermines the central efficiency claim. The stress-test confirms this is the most significant risk: it is a quantitative, checkable claim that directly supports the abstract's headline assertion and the proposed PNP metric. A secondary inconsistency is that ALinear's own parameter count (297K in Figure 1(b)) does not match Eq. 8, which gives 186K for T=96, H=960; this further signals that parameter counting is not reliable in the manuscript. The model's predictive accuracy may still be competitive, and the adaptive horizon mechanism is a reasonable contribution, so the appropriate outcome is a conditional acceptance requiring corrected parameter counts and reproducible code, exactly as the reader recommended. Therefore the reader's verdict does not change.","tokens_in":14869,"tokens_out":4435,"duration_ms":42276,"concrete_test":"Use the official open-source TimesNet implementation with the same univariate setting as the paper (input length 96, hidden dimension 512, top_k=5, horizons 48/96/192/336/720/960) and print sum(p.numel() for p in model.parameters()). Also run the same parameter count for all other baselines in Table 1 using the authors' stated hyperparameters. If TimesNet's count is about 1M rather than 1200M, recompute the 'less than 1%' claim and the PNP values in Figure 4; the quantitative efficiency claim must be revised unless another baseline supplies the ratio.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 4.1 claim that ALinear 'consistently outperforms large-scale models while using less than 1% of their parameters.' This claim depends on Figure 1(b) and Section 4.2 assigning TimesNet 1199.72M parameters, and the Introduction stating that TimesNet 'exceeded above 1,000M parameters.' In the official TimesNet implementation (Wu et al., ICLR 2023), univariate long-term forecasting models typically have on the order of 1M parameters, not 1.2B; the figure appears to overstate TimesNet's parameter count by roughly three orders of magnitude. If TimesNet actually has about 1.2M parameters, ALinear (reported as 297K in Figure 1(b), or 186K from Eq. 8) uses roughly 16-25% of TimesNet's parameters, not less than 1%. Moreover, compared to iTransformer (6.85M) or Autoformer (10.54M), ALinear would be about 3-4% of their parameters, so the blanket 'less than 1%' statement is not supported by the reported numbers. The PNP metric in Eq. 10 uses log(#Parameters), so an error of 10^3 in the baseline count changes PNP substantially (roughly a factor of 1.5 for TimesNet alone) and materially alters Figure 4. The parameter-efficiency advantage, which is central to the scaling-law argument, is therefore overstated until baseline parameter counts are recomputed consistently.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper challenges the applicability of scaling laws to time series forecasting by presenting ALinear, an ultra-lightweight linear model with three main components: a horizon-aware adaptive moving-average decomposition (Section 3.1), component-specific linear projections (Section 3.2), and a horizon-dependent recombination that exponentially decays the seasonal component with forecast horizon (Section 3.3). The authors report that ALinear uses only k-level parameters, outperforms large-scale baselines on seven univariate benchmarks while using less than 1% of their parameters, and introduce a Parameter-Normalized Performance (PNP) metric to compare accuracy per parameter (Section 4.2). They also present ablations, hyperparameter sensitivity results, and an analysis of component drift across horizons (Sections 4.3 and 4.4).","tokens_in":15236,"tokens_out":5591,"duration_ms":50661,"significance":"If the central efficiency claim survived scrutiny, this would be a practically important result: an ultra-lightweight model matching or beating much larger architectures would strengthen the case against indiscriminate parameter scaling in time series forecasting and offer a deployable baseline. The proposed horizon-dependent recombination idea and the PNP metric are useful contributions, and the paper includes an ablation study and a cross-horizon component analysis. However, the headline quantitative claims rest on baseline parameter counts that appear to be inflated by roughly three orders of magnitude for one central baseline, and on accuracy comparisons that are partly contradicted by entries in the paper's own Table 1. The manuscript also reports no error bars despite stating that experiments were repeated with five seeds. The contribution is potentially valuable but is not yet established as presented.","major_comments":[{"comment":"The reported TimesNet parameter count of 1199.72M appears to be inflated by roughly 1000x. The official TimesNet implementation for univariate long-term forecasting typically uses about 1M parameters, not 1.2B. This number is load-bearing: the abstract and Section 4.1 claim 'less than 1% of their parameters', and Figure 4's PNP scores use log(#parameters) in Eq. (10). With ALinear reported at 186K-297K parameters, the ratio to a ~1.2M-parameter TimesNet would be roughly 16-25%, not below 1%. Please recompute all baseline parameter counts using a consistent counting methodology (including whether per-variable univariate models, embeddings, and projection layers are included) and update the quantitative efficiency claims accordingly.","section":"§1, Figure 1(b), §4.2"},{"comment":"The claim that ALinear 'consistently outperforms existing methods' is contradicted by Table 1 itself. For example, at ETTm1 with pred_len=48, TimesNet (MSE 0.019), iTransformer (0.020), PatchTST (0.020), and FEDformer (0.021) all outperform ALinear (0.036); similar counterexamples appear at ETTm2 pred_len=48 and weather pred_len=48. The statement in Section 4.1 that ALinear achieves optimal results in 71.4% of scenarios needs a precise definition of 'scenario' and must be re-audited against the table. The headline conclusion and the abstract's 'consistently outperforms' formulation should be qualified to reflect the actual comparison outcomes.","section":"§4.1, Table 1"},{"comment":"The text describes the exponential decay as a 'learnable spectral filter' and a 'progressive frequency attenuation strategy', but the decay rate λ=δ/H is fixed by hyperparameter δ and the horizon H; no parameter in Eq. (5) is learned. This mischaracterizes the mechanism. In addition, the ablation 'w/o adaptive' in Table 2 removes the adaptive recombination of Eq. (6), not the decay itself, so the attribution of gains to 'progressive frequency decay' is not directly established by the reported ablations.","section":"§3.3, Eq. (5)"},{"comment":"The parameter counts are internally inconsistent. Equation (8) gives Θ_ALinear = 2HT + 2H + 4, which for T=96 and H=960 equals 186,244 parameters, while Figure 1(b) reports ALinear as 297.02K parameters. This discrepancy must be resolved before the parameter-efficiency comparison can be assessed, and it also affects the claim that ALinear uses 'k-level' parameters.","section":"§3.5, Eq. (8), Figure 1(b)"},{"comment":"The paper states that all experiments were repeated five times with different random seeds, but Table 1 reports only mean MSE and MAE values with no standard deviations, confidence intervals, or significance tests. Several comparisons are very close (e.g., ETTh1 pred_len=48 ALinear MSE 0.042 vs PatchTST 0.042; ETTh2 pred_len=336 ALinear 0.223 vs Transformer 0.223). Without error bars, the claim of consistent superiority over baselines is not statistically supported.","section":"§4.1, Experimental setup"}],"minor_comments":[{"comment":"The model name is written inconsistently as 'Alinear' in the abstract and 'ALinear' elsewhere; please standardize the spelling.","section":"Abstract, throughout"},{"comment":"Figure 2 contains literal '???' placeholders in the kernel size boxes, which suggests the figure is unfinished and should be completed before publication.","section":"Figure 2"},{"comment":"The parameter labels contain typographical errors (e.g., 'FEDformer-16,83M' and '16,95M') and the horizontal axis is partially garbled; please clean up the figure and its captions.","section":"Figure 1(b)"},{"comment":"The parameter count in Eq. (8) counts projection matrices as 2HT + 2H but does not explicitly list k1, k2, v1, v2 in the 'decomp.' and 'recomb.' terms; please clarify exactly which parameters are included in each term.","section":"§3.5, Eq. (8)"},{"comment":"The y-axis of Figure 5 is labeled only as 'Value' without units or a definition; please indicate what quantity is plotted (e.g., mean seasonal component magnitude).","section":"§4.3, Figure 5"},{"comment":"The manuscript does not state whether code or trained models will be released; for a paper whose central claim is an empirical efficiency comparison, providing the implementation and random seeds is important for verification.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the reported TimesNet parameter count of 1199.72M, which appears to be off by about three orders of magnitude. Because the paper's central thesis is the 'less than 1% parameters' claim and the PNP metric uses log(#parameters), this is not a cosmetic error. Additionally, several rows of Table 1 show that ALinear is not the best model, which conflicts with the abstract's 'consistently outperforms' claim. These issues are fixable by recomputing baseline counts, reporting error bars, and revising the claims, but they require substantial changes to the paper's central narrative. I would not recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is ALinear's horizon-aware design: an adaptive kernel size for the moving-average decomposition and an exponentially decaying seasonal term with learned recombination weights. That is a reasonable, cheap extension of DLinear, and the authors deserve credit for spelling out the mechanism clearly. The PNP metric is also a fair idea in principle, even if its implementation inherits the parameter-count problem below.\n\nThe soft spot is load-bearing. The paper repeatedly claims ALinear uses less than 1% of the parameters of large models, and points to TimesNet having over 1,000M parameters (Introduction and Figure 1(b)). I checked the public TimesNet implementation; a univariate forecasting setup has on the order of 1M parameters, not 1.2B. If the real count is near 1M, then ALinear (roughly 186K–300K by the paper's own numbers) uses maybe 16–30% of TimesNet's parameters, and roughly 3–4% of iTransformer's or Autoformer's. Still efficient, but not the 100x-plus gap the abstract and Section 4.1 sell. The PNP numbers in Figure 4 shift accordingly because they use log(#Parameters). This is not a cosmetic issue; it is the quantitative basis for the scaling-law argument.\n\nThere are smaller problems worth naming. The paper reports 5 seeds but gives no error bars, so Table 1's many bold entries have no variance attached. The parameter count from Eq. 8 (2HT+2H+4, about 186K for T=96, H=960) does not match the 297K in Figure 1(b). Section 3.3 calls the decay a learnable spectral filter, but delta is a fixed hyperparameter. The ablation is only on Weather, which limits how much weight I can put on the component-wise claims. No code is released, so reproducing Table 1 would take real effort.\n\nStill, the architecture is coherent, the paper engages with the right literature, and the empirical sweep across seven datasets is useful. This is not a paper that should be ignored; it just needs the parameter counts recomputed, the headline claim reworded, error bars added, and ideally the code released. I would send it to serious peer review with those requirements rather than desk-reject it.\n\nWho gets value: researchers working on lightweight forecasting or on the question of whether linear baselines dominate transformers. The paper is worth reading as a design proposal, but the efficiency comparison should not be quoted until the baseline counts are corrected.","headline":"The ALinear design is a sensible lightweight follow-on to DLinear, but the paper's headline 'less than 1% of parameters' claim rests on a TimesNet parameter count that looks inflated by roughly 1000x and needs correction before the efficiency story holds.","tokens_in":15695,"tokens_out":2120,"would_cite":false,"duration_ms":22663,"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 linear model with a few hundred thousand parameters outperforms large-scale transformer-based forecasters on seven benchmark datasets.","keywords":["time series forecasting","scaling law","parameter efficiency","horizon-aware decomposition","adaptive linear model","long-term forecasting","trend-seasonal decomposition","frequency attenuation"],"falsifier":"A direct replication of Table 1 using the released checkpoints would settle the claim: if ALinear's MSEs are not reproduced, or if the actual TimesNet checkpoint has tens of millions of parameters rather than 1,199.72M, the headline claim would not stand.","tokens_in":14680,"feed_emoji":"📉","tokens_out":10858,"duration_ms":95560,"temperature":0.7,"pith_summary":"This paper asks whether the scaling law that drives large language and vision models—more parameters, better performance—also applies to time series forecasting. It argues the answer is no and presents ALinear, an ultra-lightweight linear model that, in its experiments, outperforms ten established forecasting models on seven univariate datasets while using less than 1% of their parameters. The design principle is that forecasting changes character with the prediction horizon: trend stays predictable longer, while seasonal and high-frequency components become noise, so the model learns to rebalance components and attenuate high frequencies as the horizon grows. If the result holds, it would redirect research effort from parameter expansion toward horizon-aware efficiency and make accurate long-horizon forecasting feasible in resource-constrained settings.","feed_headline":"A tiny linear model beats billion-parameter forecasters","feed_subtitle":"It rebalances trend and seasonality by horizon, beating big models with under 1% of their parameters.","key_machinery":"The load-bearing object is ALinear's horizon-parameterized decomposition-recombination loop. The trend component is extracted by a moving average whose kernel size depends on the forecast horizon, $\\alpha(H)=\\min(\\max(k_1+k_2 H, w_{\\min}), w_{\\max})$, with learnable $k_1,k_2$; the seasonal part is the residual $X-T_H$. Each component gets its own linear projection ($\\hat T_H = W_T T_H + b_T$, $\\hat S_H = W_S S_H + b_S$), and the seasonal forecast is then attenuated by an exponential decay $\\hat S_H(t)\\exp(-\\lambda t)$ with $\\lambda = \\delta/H$, so high-frequency content fades as the horizon grows. The two branches are recombined with horizon-dependent weights $\\beta_T(H)=\\sigma(v_1+v_2 H)$ and $\\beta_S(H)=1-\\beta_T(H)$. This single mechanism replaces attention: it lets the model put more weight on trend and suppress seasonality at long horizons, which is exactly the behavior the paper says is missing from fixed-decomposition baselines.","core_discovery":"The paper's central claim is that the scaling law that governs large language and vision models does not transfer to time series forecasting. The evidence is ALinear, a linear model with $2HT+2H+4$ parameters (about 297K in the reported configuration), which the authors say beats ten state-of-the-art forecasting models—including Transformers, Autoformer, Crossformer, PatchTST, and TimesNet—on seven univariate benchmarks across horizons 48 to 960, while using less than 1% of the baselines' parameters. The paper further claims the advantage grows with horizon: at 960 steps ALinear's MSE is on average 23.7% lower than the second-best model, and the model's learned trend-seasonal balance shifts with dataset characteristics rather than following a fixed schedule. It also introduces Parameter-Normalized Performance, $\\mathrm{PNP}=100/(\\mathrm{Metric}\\cdot\\log(\\#\\text{parameters}))$, as a parameter-aware evaluation metric, on which ALinear ranks first.","pith_inferences":["Beyond the seven benchmarks, if the result transfers, the useful notion of a scaling law for forecasting would be relative to the amount of predictable signal in the data, not to raw data size; one should then see similar gains for other horizon-adaptive linear models on noisy industrial series.","A Fourier-domain variant that applies horizon-dependent spectral masks rather than exponential time-domain decay should reproduce ALinear's gains, which would identify frequency attenuation, not the linear projection, as the active mechanism.","A natural test the paper does not run is multivariate forecasting; applying the same adaptive decomposition channel-by-channel would show whether the dataset-dependent component drift persists when variables interact.","The PNP metric's log-parameter denominator encodes a policy choice: a tenfold parameter cut at equal error is always worth a fixed efficiency bonus, so adopting PNP as a reporting standard would change which models are called state of the art."],"forward_implications":["On the seven univariate benchmarks tested, ALinear's roughly 0.3-million-parameter model matches or beats models with millions to over a billion parameters across horizons from 48 to 960 steps.","The accuracy advantage grows with horizon: at 960 steps ALinear averages a 23.7% lower MSE than the second-best model.","Because ALinear's time and space complexity scale as the product of input length and horizon, the reported accuracy is available for real-time and resource-constrained deployment without specialized hardware.","The learned trend-seasonal balance is dataset-dependent, so fixed-decomposition linear baselines such as DLinear are expected to lose ground at longer horizons.","Under the proposed PNP metric, parameter efficiency becomes an explicit evaluation axis, and ALinear's PNP-MSE and PNP-MAE of 105 and 68.9 rank first among compared models."],"supporting_citations":[{"why":"the linear-decomposition baseline that ALinear extends with horizon-adaptive mechanisms.","marker":"[8]"},{"why":"the large model whose reported 1,199.72M parameter count anchors the 'less than 1%' efficiency claim.","marker":"[7]"},{"why":"the attention architecture whose scaling the paper challenges and a Table 1 baseline.","marker":"[1]"},{"why":"the sparse-attention long-sequence baseline that deteriorates most at ultra-long horizons.","marker":"[2]"},{"why":"the decomposition-with-attention model that motivates ALinear's separate trend and seasonal paths.","marker":"[3]"},{"why":"the frequency-domain transformer whose uniform frequency processing contrasts with ALinear's progressive attenuation.","marker":"[6]"},{"why":"the recent transformer baseline ALinear must beat to support its state-of-the-art claim.","marker":"[26]"},{"why":"the cross-dimension transformer baseline used in the efficiency comparison and Table 1.","marker":"[37]"},{"why":"the scaling-law premise from language modeling that the paper sets out to question.","marker":"[9]"},{"why":"the forecasting text that grounds the claim that trend stays predictable while seasonality becomes stochastic at longer horizons.","marker":"[11]"}],"fun_headline_variants":["Scaling law fails in time series forecasting","Tiny linear model beats huge forecasters with 1% params","ALinear: 1% parameters, 23% better long-horizon MSE","Scaling law doesn't transfer to time series forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative claim that ALinear uses less than 1% of a large model's parameters rests on the reported parameter counts for the baselines, especially TimesNet's stated 1,199.72M; if those counts are wrong, the efficiency comparison loses its basis.","fun_headline_variants_meta":{"raw":{"variants":["Scaling law fails in time series forecasting","Tiny linear model beats huge forecasters with 1% params","ALinear: 1% parameters, 23% better long-horizon MSE","Scaling law doesn't transfer to time series forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000322,"raw_usage":{"total_tokens":1823,"prompt_tokens":973,"completion_tokens":850,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":779}},"tokens_in":589,"tokens_out":850,"duration_ms":7894,"temperature":1.0,"reasoning_tokens":779,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:14:33.029014+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct replication of Table 1 using the released checkpoints would settle the claim: if ALinear's MSEs are not reproduced, or if the actual TimesNet checkpoint has tens of millions of parameters rather than 1,199.72M, the headline claim would not stand.","supporting_citations":[{"cited_title":"Are transformers effective for time series forecasting?","cited_arxiv_id":null,"evidence_quote":"the linear-decomposition baseline that ALinear extends with horizon-adaptive mechanisms."},{"cited_title":"Informer: Beyond efficient transformer for long sequence time-series forecasting,","cited_arxiv_id":null,"evidence_quote":"the sparse-attention long-sequence baseline that deteriorates most at ultra-long horizons."},{"cited_title":"Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting,","cited_arxiv_id":null,"evidence_quote":"the decomposition-with-attention model that motivates ALinear's separate trend and seasonal paths."},{"cited_title":"Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting,","cited_arxiv_id":null,"evidence_quote":"the frequency-domain transformer whose uniform frequency processing contrasts with ALinear's progressive attenuation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the forecasting text that grounds the claim that trend stays predictable while seasonality becomes stochastic at longer horizons."}],"review_version":1}