{"id":"99e387a9-313b-4fe1-b85e-2afd5341dc0d","arxiv_id":"2508.07016","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A framework that retrieves time-lagged correlated series via a shifted dynamic time warping distance, or a learned approximation of it, and uses them as auxiliary inputs to improve time series forecasting.","lead":"This paper proposes a framework that finds other time series shifted in time and related to the one being forecast, then feeds those related series into standard forecasting models as extra inputs. The reason to read it is that this simple idea, time-lagged cross-correlations, is often ignored by deep learning forecasters, and the paper reports consistent error reductions on weather, stock, and real estate data plus a large speedup from a learned approximator.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random split of overlapping sliding windows in §4.1.1/4.1.3 leaks future information into training, so the reported MSE reductions in Table 1 are not validated as forecasting improvements.","rationale":"The reader's verdict of CONDITIONAL is appropriate. My concern is more specific than the reader's weakest assumption: instead of focusing only on the general fragility of the protocol or lack of significance tests, I identify a concrete leakage mechanism in the random split of overlapping sliding windows. This is load-bearing because the entire quantitative case rests on the MSE reductions in Table 1. If those numbers are inflated by leakage, the central claim fails. The proposed chronological-split test would settle this. I do not believe the paper should be rejected outright, because the stock experiment (chronological split) and the ablation (RS vs DTW vs SSDTW) provide some supporting evidence that the selection mechanism can help. But the weather and real estate results need to be revalidated. The verdict remains CONDITIONAL pending that test.","tokens_in":12939,"tokens_out":7891,"duration_ms":77136,"concrete_test":"Re-run the weather and real estate experiments (Table 1) using a chronological split: train on the first 60% of the sliding-window sequence, validation on the next 20%, test on the last 20%, keeping all other settings identical. Report the resulting average MSE reductions for SSDTW and CLE. If the average reductions no longer exceed a few percent or change sign, the reported gains are an artifact of window-overlap leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1.1 (weather) and §4.1.3 (real estate) state that the data are 'randomly divided into training, validation, and test sets in a 6:2:2 ratio.' Because the forecasting protocol creates overlapping sliding windows (e.g., weather: [1,50], [2,51], ...; real estate: first-nine-months → last-three-months), a random split places windows that share almost all input timesteps into different partitions. A test window starting one day after a training window has 49/50 identical inputs and a target that is the very next day; the model can effectively memorize near-duplicates during training. Hence the MSE values in Table 1 reflect in-sample memorization rather than out-of-sample forecasting. This threatens the central accuracy claim (16.01% / 9.95% / 21.29% reductions) because those numbers may not reproduce on a proper chronological split. The stock dataset avoids this because it follows FNSPID's chronological test split (§4.1.2), which makes the inconsistency between datasets a red flag. The paper also gives no variance across random seeds, so we cannot tell whether any remaining gains are significant.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TLCCSP, a framework that augments a target time series with auxiliary series selected by a new SSDTW distance: the minimum DTW distance between the target and the candidate over a small set of time shifts. The top-Ks candidates are concatenated as extra input features to an arbitrary forecasting backbone. To avoid the O(N^2 T^2) cost of SSDTW, a contrastive learning encoder (CLE) is trained so that embedding distances approximate SSDTW distances, and the paper reports that using CLE cuts SSDTW computation by about 99%. Experiments on weather, stock, and real estate data with seven backbones report average MSE reductions of 16.01%, 9.95%, and 21.29% for SSDTW and 17.88%, 6.13%, and 8.62% for CLE relative to single-series forecasting. An ablation on the stock dataset compares SSDTW with random selection and plain DTW; sensitivity studies vary Ks, Ke, and the temperature lambda.","tokens_in":13207,"tokens_out":5811,"duration_ms":62843,"significance":"If the reported gains are real, the framework is a useful plug-in for existing forecasting models and the CLE speedup is practically significant. The ablation against random selection and plain DTW is a good control, and the case study supports the qualitative claim that SSDTW and CLE retrieve lagged correlated sequences. However, the main evidence is not yet convincing: on two of the three datasets the evaluation protocol uses a random split of overlapping sliding windows, which can leak the target into training and invalidate the reported reductions. The absence of any variance or significance information further weakens the quantitative claims. The work is therefore promising but requires a corrected evaluation before the central forecasting claim can be accepted. Strengths: the method is simple, model-agnostic, and the computational-cost reduction is explicitly measured; the paper does not oversell the SSDTW-vs-CLE comparison and even notes cases where CLE outperforms SSDTW.","major_comments":[{"comment":"The weather and real estate datasets use sliding windows and then a random 6:2:2 train/validation/test split. Weather windows [1,50], [2,51], ... share 49 of 50 input days; real estate windows share 8 of 9 input months. A random split puts near-duplicate windows into different partitions, so the model can memorize the target rather than forecast it. The reported weather and real estate reductions (16.01% and 21.29%) are therefore not validated as out-of-sample improvements. A chronological split, as apparently used for the stock dataset, is required, and all three datasets should be evaluated under the same protocol.","section":"§4.1.1 and §4.1.3, Table 1"},{"comment":"No variance, number of runs, or significance statistics are reported for any entry. Many effects are small or negative (e.g., Stock LSTM SSDTW 1.68 vs Single 1.65; Real Estate TimesNet SSDTW 0.88 vs 0.87; Real Estate Transformer CLE 1.57 vs 1.55), so the average reductions in Table 1 cannot be distinguished from random seed variation. This is not a presentational issue: the central claim is quantitative, and the paper should provide repeated-run statistics or, at minimum, error bars for the headline numbers.","section":"Table 1, §4.1"},{"comment":"SSDTW selects series by minimizing DTW over a set of shifts tau, but Eq. (3) feeds the raw auxiliary histories {S*_t} to the forecaster without the winning shift. Thus the framework does not actually use the estimated lag in prediction; it only uses the identity of the selected series. If the lag itself is informative, an experiment with shifted auxiliary inputs or explicit lag features is needed; if only the selection matters, the claim that TLCCSP 'captures time-lagged cross-correlations' should be softened. This distinction is central to the paper's stated mechanism.","section":"§3.1 Eq. (3), §3.2 Eq. (4)-(5)"}],"minor_comments":[{"comment":"The abstract says CLE 'further decreases' MSE on weather by 17.88%, but Table 1's Δ column compares CLE to Single, not to SSDTW. In stock and real estate, CLE is on average worse than SSDTW. Please rephrase to avoid implying a stacked improvement.","section":"Abstract, Table 1"},{"comment":"The statement that SSDTW and CLE 'consistently rank first or second' is contradicted by Table 1: e.g., Stock LSTM SSDTW is worse than Single, and Real Estate TimesNet SSDTW is worse than Single. Please qualify the claim.","section":"§4.4"},{"comment":"The positive and negative sample sets are described with the same index i and with conditions i ≤ Ke and i > N−Ke. This is ambiguous about the ordering of SSDTW distances; please define rank explicitly from 1 (nearest) to N (farthest).","section":"Eq. (6)"},{"comment":"Typos and inconsistent terminology: 'TLCCSPP' in the conclusion, 'Conclution', 'appliations', 'sequencess', and 'candidate stock' in a weather context where the objects are cities. Please copyedit.","section":"Throughout"},{"comment":"The CLE encoder is described only as a '3-block convolutional neural network' from ADATIME [40]; no layer sizes, pooling, or training hyperparameters for the encoder are given. Please provide enough detail for reproducibility.","section":"§3.3, §4.3"}],"recommendation":"major_revision","confidential_remarks":"The random-split leakage in §4.1.1 and §4.1.3 is a serious validity problem: the weather and real estate results in Table 1 may not survive a chronological split. I recommend the editor require the authors to rerun those experiments with a proper chronological train/validation/test split and to report repeated-seed statistics before considering the paper for publication. If the gains disappear under the corrected protocol, the paper should be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a straightforward engineering wrapper—compute a minimum-DTW distance over a few time shifts, pick the nearest series, feed them as extra features—and that is a fair thing to test. The ablation against random selection and plain DTW is the best part of the paper; it supports the claim that the selection criterion, not just added dimensionality, drives the gains. The CLE encoder is a reasonable contrastive approach, and the ~99% reduction in pairwise correlation cost is plausible, though the 980-hour figure is an estimate, not a measured runtime.\n\nWhere it falls down is evaluation hygiene. The weather dataset is built from overlapping 50-day windows and then randomly split 6:2:2. That puts near-duplicate windows in train and test, so the test loss reflects memorization of the same evolving series rather than true forecasting. If that is the actual protocol—and the text says it is—the 16.01% weather MSE reduction is not a valid out-of-sample result. The stock dataset follows a chronological split and is the cleaner experiment. The real estate section is ambiguous; if the split is by plot with one window each, leakage isn't there. The paper needs to specify and switch to chronological splits.\n\nAlso missing: any variance across seeds or significance tests. Several CLE-vs-SSDTW differences are a few percent; without error bars those are not interpretable. The encoder is trained with a contrastive loss, not a regression loss to SSDTW distances, so the claim that embeddings approximate SSDTW distances is only qualitatively supported. And the imputation of real estate data with TimesNet before correlation computation is a confounder, though probably minor.\n\nIn short: the core idea is worth exploring and the paper is readable, but the headline numbers are not established. The right fix is a chronological split, error bars, and either code or a real runtime measurement. This deserves a serious referee because the method is simple enough to be useful and the ablation is genuinely informative; it just needs major revision before the empirical claims can be trusted.","headline":"Useful wrapper idea and a solid ablation, but the weather split leaks future information, so the headline MSE gains are not trustworthy as forecast improvements.","tokens_in":13711,"tokens_out":4519,"would_cite":false,"duration_ms":45610,"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":"This paper argues that appending the top few time-lagged cross-correlated series to a forecaster's input lowers prediction error on weather, stock, and real estate data, and that a contrastive encoder can approximate the required correlatio","keywords":["time series forecasting","time-lagged cross-correlation","dynamic time warping","contrastive learning","auxiliary sequence selection","SSDTW","scalable similarity search","forecasting accuracy"],"falsifier":"Run the same seven backbones on a held-out forecast period where all correlations are computed strictly from earlier data (weather: 2019 only; stock: through 2018 only; real estate: 2022 only). If Table 1's MSE reductions disappear or random selection matches them, the central claim fails.","tokens_in":12787,"feed_emoji":"📈","tokens_out":10771,"duration_ms":95147,"temperature":0.7,"pith_summary":"This paper tries to establish that a forecasting model can be improved simply by giving it, as extra input features, a small set of other time series that tend to move after a delay. The authors define a shift-aware similarity, SSDTW, as the minimum dynamic time warping distance between the target series and a candidate series shifted by a few days or months. Across weather, stock, and real estate data and seven forecasting backbones, adding the top $K_s$ series by this measure reduces mean squared error relative to forecasting from the target alone. To make the search scalable, the paper trains a contrastive learning encoder whose distances approximate SSDTW, and reports that this preserves most of the accuracy gain while cutting correlation-computation time by about 99%.","feed_headline":"Lagged neighbors cut forecasting error across three domains","feed_subtitle":"Feeding the most shift-correlated series to a forecaster cuts MSE 16%, 10%, 21%; an encoder runs the search ~99% faster.","key_machinery":"The central mechanism is Sequence Shifted Dynamic Time Warping (SSDTW), defined as $\\mathrm{SSDTW}(A,S)=\\min_{\\tau\\in\\mathcal{T}}\\mathrm{DTW}(A,S_\\tau)$, where $\\mathcal{T}$ is a small domain-specific set of shift windows. It makes the similarity between two series depend on whether one moves after the other, not just on matched shape. The second mechanism is a contrastive learning encoder: a three-block convolutional network trained with the most SSDTW-correlated series as positives and the least as negatives, so that cosine distance in embedding space approximates SSDTW distance and replaces the expensive dynamic-programming search.","core_discovery":"The paper's central claim is that time-lagged cross-correlations between series are a reusable source of forecasting signal. It defines the Sequence Shifted Dynamic Time Warping distance as $\\mathrm{SSDTW}(A,S)=\\min_{\\tau\\in\\mathcal{T}}\\mathrm{DTW}(A,S_\\tau)$, where $\\mathcal{T}$ is a small set of shift windows chosen per dataset. Selecting the top $K_s$ candidate series by this distance and concatenating them to the target's history as auxiliary inputs lowers MSE relative to single-series forecasting: averages of 16.01% on weather, 9.95% on stock, and 21.29% on real estate across seven backbone models. The paper further claims a contrastive learning encoder whose embedding distances approxi","pith_inferences":["Beyond the paper: the shift sets $\\mathcal{T}$ are hand-picked per dataset; a natural extension is learning the shift set from data rather than choosing {1,3,5,10}, {5,10,20,30}, or {1,2,3}.","Beyond the paper: the reported gains are cell averages without variance or significance information, so a paired resampling analysis across random seeds would clarify whether the improvements are systematic or carried by a few backbones.","Beyond the paper: the paper notes CLE sometimes beats SSDTW selection, which suggests the embedding may encode more than the minimum-DTW score; testing whether embedding distances track lead-lag direction or correlation strength would clarify the mechanism."],"forward_implications":["Any existing forecaster can be upgraded by appending the top $K_s$ lag-correlated series to its input, with no change to the backbone architecture.","The contrastive encoder makes this selection cheap enough to update continuously on large candidate pools, enabling real-time retrieval of correlated series for financial and weather applications.","The framework is domain-agnostic: the same procedure works with day-level, multi-day, or month-level shift sets, as long as the shift set matches the data's timescale.","Choosing too many auxiliary series degrades accuracy, so the selection count $K_s$ is a meaningful hyperparameter rather than a free lunch."],"supporting_citations":[{"why":"Supplies the stock dataset, the 49-day-in/1-day-out window protocol, and the baseline model configuration used for stock experiments.","marker":"[15]"},{"why":"Supplies the weather dataset used for both correlation computation and forecasting.","marker":"[14]"},{"why":"Supplies the real estate dataset, with missing values imputed before correlation computation.","marker":"[16]"},{"why":"Supplies the TimesNet backbone used in experiments and the imputation model applied to missing real-estate entries.","marker":"[3]"},{"why":"Supplies the three-block convolutional encoder architecture used by the contrastive learner.","marker":"[40]"},{"why":"Supplies the DTW algorithm that SSDTW extends by taking the minimum over time shifts.","marker":"[38, 39]"}],"fun_headline_variants":["Lagged cross-correlations cut forecasting error","Shift-correlated series improve forecasting","Use lagged neighbors to forecast with less error","Time-lagged correlation search boosts forecast accuracy","Lagged series pairing reduces MSE across domains"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the selected lag-correlated series carry information about the target's future that the target's own history does not already contain, and that the correlation periods used to select them do not leak future information into the forecast.","fun_headline_variants_meta":{"raw":{"variants":["Lagged cross-correlations cut forecasting error","Shift-correlated series improve forecasting","Use lagged neighbors to forecast with less error","Time-lagged correlation search boosts forecast accuracy","Lagged series pairing reduces MSE across domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1444,"prompt_tokens":821,"completion_tokens":623,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":558}},"tokens_in":565,"tokens_out":623,"duration_ms":6250,"temperature":1.0,"reasoning_tokens":558,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:22:14.781762+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same seven backbones on a held-out forecast period where all correlations are computed strictly from earlier data (weather: 2019 only; stock: through 2018 only; real estate: 2022 only). If Table 1's MSE reductions disappear or random selection matches them, the central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the stock dataset, the 49-day-in/1-day-out window protocol, and the baseline model configuration used for stock experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the weather dataset used for both correlation computation and forecasting."},{"cited_title":"RETQA: A Large-Scale Open-Domain Tabular Question Answering Dataset for Real Estate Sector","cited_arxiv_id":"2412.10104","evidence_quote":"Supplies the real estate dataset, with missing values imputed before correlation computation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TimesNet backbone used in experiments and the imputation model applied to missing real-estate entries."}],"review_version":1}