{"id":"0c7801a7-753d-4a02-82e9-4a83df012ebc","arxiv_id":"2412.01557","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A lightweight timestamp-encoding module improves long-term forecasting accuracy when blended with existing backbones, especially linear models on electricity and traffic data.","lead":"This paper adds a small module (TimeSter) that encodes calendar timestamps into an extra forecast, then blends it with a standard time series model. On electricity and traffic benchmarks, the blended linear model cuts mean squared error by roughly 23% over a plain linear projector.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-dataset/per-horizon time-feature selection appears test-based, so the headline 23% improvement on Electricity/Traffic may be optimistic and may not replicate under a validation protocol.","rationale":"The reader's verdict (CONDITIONAL) is appropriate, but the reader's weakest_assumption focuses on stationarity of calendar effects in Eq. (2). That is a genuine theoretical limitation and is partially self-acknowledged in the Weather discussion, yet the empirical claim on the seven benchmarks is still meaningful within the tested time spans. The more load-bearing concern is methodological: the per-dataset, per-horizon feature selection appears to be performed on the test set, since Table 5 lists test MSE for all feature combinations and Table 9 selects the best ones. Because the magnitude of the feature-choice effect is comparable to the headline 23% improvement, this selection bias could substantially shrink the claimed gain under an out-of-sample protocol. The paper's ablations are otherwise informative and the integration across backbones (Tables 3 and 4) gives useful evidence that timestamp features help, but the headline number needs a hold-out validation check. This aligns with the reader's rationale, which noted the missing validation protocol, though it was not listed as the weakest assumption; hence 'partial' agreement. My verdict is UNCHANGED because the reader already set CONDITIONAL, and this concern is exactly the kind of addressable conditionality that a revision should resolve.","tokens_in":24354,"tokens_out":5575,"duration_ms":51701,"concrete_test":"Re-run the full Table 3 comparison using a proper validation protocol: split each training set (e.g., last 20% of the training period) as validation, select the time-feature subset and 𝛽/ksize per dataset and horizon using only validation MSE, then retrain on train+val and report test MSE for both RLinear and RLinear+TimeSter. If the average Electricity/Traffic improvement over RLinear drops below roughly 10%, or if TimeSter no longer beats CycleNet or SOFTS on those datasets, the reported 23% is substantially an artifact of test-set feature selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on the comparison in Table 3 between RLinear and RLinear+TimeSter, where TimeSter uses the \"optimal\" time-related features listed in Table 9 for each dataset and prediction length. The paper never defines a validation split for selecting these features. Section 5.2.1 reports test-set MSE for all eight feature combinations at horizon 720 (Table 5), and Table 9's choices coincide with the best-performing rows of Table 5 (plus the Min_H addition in Table 8 for ETTm1). This means the same test labels that produce the headline numbers are used to choose which timestamps enter the model. The bias can be large: on Electricity at T=720, H alone gives MSE 0.249 while the selected H_D_S gives 0.198, a 20% swing; on ETTh1, H_D_M_S gives 0.581 versus H at 0.464. The 23% average improvement on Electricity and Traffic is therefore not an unbiased estimate of TimeSter's benefit; it is the improvement of the best-tested feature set, not of a fixed, pre-specified model. A separate limitation, acknowledged for Weather and inherent to Eq. (2), is the omission of any year/trend feature, making the timestamp branch rely on calendar stationarity across years; this is a deployment concern but does not by itself invalidate the in-benchmark measurements. The feature-selection issue is more load-bearing because it directly affects the size of the headline gain.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies whether explicit timestamp features (hour of day, day of week, month, season, minute) can improve long-term multivariate time series forecasting. It introduces TimeSter, a module that encodes historical timestamp features with two nonlinear layers, a Conv1d, and a linear layer, then predicts via a single linear projector; the TimeSter output is combined with a backbone forecaster through a fixed weighting β. The main model TimeLinear uses a linear backbone and simplified RevIN. The authors report that TimeSter reduces the average MSE of RLinear by about 23% on Electricity and Traffic, improves several other backbones (FITS, PatchTST, ModernTCN, TimesNet, iTransformer), and that TimeLinear is competitive with or better than more complex state-of-the-art models on seven benchmark datasets with 100k–1M parameters. The paper includes ablations of feature subsets, module components, decoder variants, longer look-back windows, robustness over three seeds, and an ACF-based explanation for feature selection.","tokens_in":24613,"tokens_out":7653,"duration_ms":69401,"significance":"If the headline result is reproducible under a pre-specified protocol, this is a useful contribution to the LTSF empirical literature. The proposed module is simple, has a clear intuition, and is evaluated across multiple architectures; the paper ships code, reports three-seed robustness for the main model, and provides an ACF analysis that makes the feature choices falsifiable. The main caveat is that the headline 23% gain currently depends on choosing the best time-feature subset on the test set, which inflates the expected improvement of a fixed model. With a validation-based selection protocol or a fixed feature set, the quantitative claim would be much stronger.","major_comments":[{"comment":"No validation protocol is defined for the per-dataset/per-horizon choice of time-related features. The selections in Table 9 coincide with the best-performing rows on the test set (Table 5 for T=720, with the Min_H addition for ETTm1 from Table 8): for Electricity at T=720, H_D_S gives 0.198 while H alone gives 0.249, and for Traffic, H_D gives 0.512 while H gives 0.646. Because the same test labels are used both to select features and to report the 23% averaged improvement, the headline number is the best-tested configuration, not an unbiased estimate for a pre-specified model. Please fix the feature set per dataset (or per horizon) before accessing test labels, or use a hold-out validation split for selection, and report the resulting estimates.","section":"§5.2.1, §5.3.5, Tables 5, 8, 9"},{"comment":"TimeSter encodes only historical timestamps U, while the future timestamp features P defined in Section 2 are not used. The fixed linear map in Eq. (2) therefore assumes the calendar patterns (hour, weekday, month, season) in the training window are stationary across the forecast span, and the deliberate omission of a year feature means inter-annual trends or regime shifts cannot be represented. The Weather discussion in §5.2.1 acknowledges the one-year-span problem for seasonality, but the paper should state this as a general limitation of the timestamp branch and scope the generalization claims accordingly; a constructive test would be to evaluate on a train/test year split with a year shift.","section":"§4.1, Eq. (2), §5.2.1"},{"comment":"The phrase 'consistent state-of-the-art performance across seven real-world datasets' overstates the results. In Table 2, TimeLinear is behind MSD-Mixer on ETTm1 (0.385 vs 0.361) and ETTm2 (0.273 vs 0.271), behind MSD-Mixer on Weather (0.251 vs 0.233), and behind SOFTS on Traffic (0.480 vs 0.409). The more modest abstract wording 'on par with or exceed' is supported, but 'consistent SOTA' is not, and should be revised or substantiated with a statistical comparison.","section":"Abstract, §5.1.1, Table 2"}],"minor_comments":[{"comment":"The main comparisons report only mean values; Table 10 gives standard deviations for TimeLinear but not for baselines. Add error bars or standard-deviation columns for the key comparisons, or state that the differences are within seed noise.","section":"Tables 2–4, Table 10"},{"comment":"The repeated 'A vg' headings should be 'Avg'.","section":"Tables 3 and 4"},{"comment":"'our mode' should read 'our model'.","section":"§5.2.3"},{"comment":"The text refers to 'the optimal time-related features' without specifying the selection criterion; please state the performance metric and the split used to define optimality.","section":"§5.3.5"},{"comment":"The figure referenced in the introduction does not define the baselines it compares; please add a caption or legend sufficient to interpret the comparison.","section":"Figure 1(a)"},{"comment":"Timespans such as '2016.7 - 2018.6' should be given as explicit dates or ranges for clarity.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The central technical idea is reasonable and the ablations are thorough, but the headline quantitative claim rests on test-set feature selection with no validation protocol. If the authors can provide a validation-based selection or fix the feature sets and the gains persist, the paper could be acceptable after a revision. I also recommend that the journal require the SOTA claim to be aligned with Table 2 and that baseline variance be reported on an equal footing with TimeLinear."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core empirical claim survives reading: adding a timestamp-only branch to a linear backbone consistently improves MSE across datasets, and the 23% average gain on Electricity and Traffic is in the right ballpark. But the headline number is the best of eight feature sets chosen after seeing test results, not a pre-specified model, so quote it with a caveat.\n\nWhat is actually new: TimeSter is a lightweight timestamp encoder plus linear decoder, blended with any backbone via a fixed weight. The systematic per-dataset and per-horizon study of timestamp combinations (hour, day, month, season, minute) and the ACF analysis tying feature selection to periodicity is a solid contribution. The module is simple and parameter-efficient, and the gains appear consistently across RLinear, FITS, PatchTST, ModernTCN, TimesNet, and iTransformer. The robustness table with three seeds is reassuring, and the code link is there.\n\nThe main soft spot is feature selection. Table 9 reports \"optimal\" time-related features per dataset and horizon, but the paper never describes a validation split for choosing them. The choices line up with the best rows of Table 5 at T=720, plus the minute feature added after Table 8 for ETTm1. On Electricity at T=720, hour alone gives 0.249 MSE while the selected hour-day-season set gives 0.198, a 20% swing. So the abstract's 23% improvement is the gain of the best-tested feature set, not of a fixed, pre-specified model. I do not think this invalidates the qualitative conclusion—some timestamp features clearly help—but it makes the headline number optimistic.\n\nTwo smaller points. Equation (2) maps historical timestamp features to the horizon with a fixed linear layer and omits any year or trend feature, so the timestamp branch assumes calendar stationarity across years; the authors acknowledge the Weather one-year limitation but not this general deployment concern. And Table 2's baselines are mostly copied from prior papers with no error bars, though Tables 3 and 4, which matter for the module comparison, are rerun and consistent.\n\nMy take: send it to review. The idea is simple, the study is systematic, and the weakness is addressable. Ask for a validation protocol for feature selection, rerun the headline numbers with pre-specified feature sets, and add error bars to Table 2.","headline":"TimeSter's core claim holds—timestamp features genuinely help—but the headline 23% gain is the best of several feature sets chosen on test labels, so treat it as optimistic until a validation protocol is added.","tokens_in":25172,"tokens_out":2464,"would_cite":true,"duration_ms":22662,"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":"The paper claims that explicit time-stamp encoding, added as a lightweight branch, cuts a linear forecaster's average MSE by about 23% on Electricity and Traffic and lets a 100k-parameter model match much larger networks across seven…","keywords":["time series forecasting","time-related features","time stamp encoding","long-term forecasting","linear models","seasonality","temporal feature selection","multivariate time series"],"falsifier":"Hold out a later year whose seasonal pattern is shifted relative to the training years, for example a dataset spanning a policy change or an unusually warm year, and train TimeLinear on the earlier years: if the MSE gain over the plain linear backbone disappears or becomes negative, the calendar-stationarity assumption is the load-bearing part.","tokens_in":24114,"feed_emoji":"🕒","tokens_out":4758,"duration_ms":41386,"temperature":0.7,"pith_summary":"The paper argues that the calendar information attached to each data point—hour of day, day of week, month, and season—is itself a powerful forecasting signal that most modern models leave underused. To make that signal explicit, it introduces TimeSter, a small encoder that turns historical timestamps into pseudo-observations and projects them to the future, and TimeLinear, which adds TimeSter's forecast to a plain linear backbone. On Electricity and Traffic this addition reduces average MSE by about 23% relative to the linear model alone, and across seven datasets TimeLinear matches or beats larger Transformer- and convolution-based models with far fewer parameters. The claim matters because it suggests a cheap, architecture-agnostic way to improve long-term forecasts.","feed_headline":"Time-of-day and season features cut forecast error by 23%","feed_subtitle":"A lightweight timestamp branch lets a simple linear model beat far larger networks on seven benchmarks.","key_machinery":"The load-bearing object is the TimeSter encoder-decoder pair coupled with a weighted late fusion. TimeSter encodes historical timestamps $U$ through two nonlinear layers, a 1D convolution, and LayerNorm into pseudo-observations $X_U$, then applies a single linear layer $Y_U = W X_U + b$ to produce a future forecast; the final prediction is $Y' = \\beta Y_B + (1-\\beta) Y_U$ with the backbone forecast $Y_B$, plus simplified reversible instance normalization. This machinery isolates stationary calendar cycles, which repeat every day, week, month, or season, from the residual non-cyclical part the backbone must learn, which is why a linear backbone can suddenly compete with much larger models.","core_discovery":"The paper's central discovery is that time stamps are not just auxiliary metadata but a first-class predictive signal: given only the hour, weekday, month, and season of a future point, a single linear layer can already forecast strongly periodic series competitively, and adding that timestamp forecast to a backbone's observation-based forecast yields large and consistent gains. TimeSter models the conditional distribution of observations given timestamps and decodes it with one linear map, so its contribution is added to the backbone's as a fixed weighted sum. The paper further finds that the benefit is largest for channel-independent backbones and for datasets with clear daily, weekly, or seasonal autocorrelation, and that the optimal timestamp features shift with dataset and prediction horizon.","pith_inferences":["The paper omits the year feature deliberately, so the method implicitly assumes calendar effects are stationary across years; a natural extension, which the paper does not test, is a year-aware or debiased timestamp embedding.","A testable boundary: on series with weak or drifting periodicity, such as financial returns or sensor noise, the timestamp branch should add little, so measuring the gain against autocorrelation strength would delimit where TimeSter helps.","The fixed mixing weight $\\beta$ could be made time-varying, letting the model down-weight calendar priors during regime shifts, an option the paper leaves unexplored.","Because TimeSter is a small separate branch, it could double as a lightweight domain-calibration module: re-fitting only the timestamp embedding on new data might adapt a pretrained backbone to a new seasonality without retraining the whole model."],"forward_implications":["Adding TimeSter to a plain linear projector cuts average MSE by about 23% on Electricity and Traffic and improves results across all seven benchmark datasets.","TimeSter helps across architectures, with larger gains for channel-independent backbones such as RLinear, FITS, and PatchTST than for channel-dependent ones such as ModernTCN, TimesNet, and iTransformer.","On strongly periodic data, timestamps alone can outperform observation-only linear forecasting for a 720-step horizon, implying that calendar cycles dominate the predictable signal in such series.","The optimal timestamp feature set changes with dataset and forecast length: hour suffices for daily cycles, while month and season become useful for 336- and 720-step horizons.","The benefit of explicit timestamps shrinks as the look-back window grows, because longer history already contains the cycles the timestamp branch encodes."],"supporting_citations":[{"why":"RLinear is the linear backbone that TimeSter is combined with to form TimeLinear, and it is the main baseline TimeLinear improves upon.","marker":"[18]"},{"why":"CycleNet supplies the periodic-pattern baseline and the autocorrelation-function analysis used to justify timestamp feature selection.","marker":"[19]"},{"why":"PatchTST is one of the backbones used to show that TimeSter generalizes beyond linear models.","marker":"[26]"},{"why":"GLAFF is the main timestamp-aware baseline whose adaptive-weight approach TimeSter is compared against and outperforms.","marker":"[34]"},{"why":"iTransformer provides a timestamp-as-token baseline and serves as a channel-dependent backbone in the generalization study.","marker":"[20]"},{"why":"Autoformer supplies the Electricity dataset and an early example of adding timestamps through position encoding.","marker":"[40]"},{"why":"RevIN is the normalization method whose simplified variant TimeSter and TimeLinear build on.","marker":"[14]"}],"fun_headline_variants":["Timestamps alone forecast periodic series; adding them cuts error 23%","Lightweight time-stamp module enhances linear forecasting by 23%","Hour, weekday, and season features reduce forecast MSE by 23%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The timestamp branch assumes the calendar pattern learned from the training years repeats unchanged in the forecast period, because the year feature is deliberately omitted and a fixed linear map is used; a multi-year trend or a shift in seasonal timing would break that contribution.","fun_headline_variants_meta":{"raw":{"variants":["Timestamps alone forecast periodic series; adding them cuts error 23%","Lightweight time-stamp module enhances linear forecasting by 23%","Hour, weekday, and season features reduce forecast MSE by 23%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000952,"raw_usage":{"total_tokens":4024,"prompt_tokens":874,"completion_tokens":3150,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":3089}},"tokens_in":490,"tokens_out":3150,"duration_ms":22352,"temperature":1.0,"reasoning_tokens":3089,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:17:26.399243+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a later year whose seasonal pattern is shifted relative to the training years, for example a dataset spanning a policy change or an unusually warm year, and train TimeLinear on the earlier years: if the MSE gain over the plain linear backbone disappears or becomes negative, the calendar-stationarity assumption is the load-bearing part.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CycleNet supplies the periodic-pattern baseline and the autocorrelation-function analysis used to justify timestamp feature selection."},{"cited_title":"Proceedings of the VLDB Endowment 17, 9 (2024), 2363– 2377","cited_arxiv_id":null,"evidence_quote":"GLAFF is the main timestamp-aware baseline whose adaptive-weight approach TimeSter is compared against and outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Autoformer supplies the Electricity dataset and an early example of adding timestamps through position encoding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RevIN is the normalization method whose simplified variant TimeSter and TimeLinear build on."}],"review_version":1}