{"id":"3c948a77-0acd-4788-bdee-ce4ba431621f","arxiv_id":"2412.19423","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Applying PCA to the time axis of series windows before deep model training keeps average task accuracy while cutting compute and memory, but gains and losses vary strongly by model and dataset.","lead":"This paper tests an old preprocessing trick on modern deep time series models: compress each time window into its most important parts before training. The authors report up to 40% faster training and 30% lower memory use, with roughly unchanged average accuracy on forecasting, classification, and regression benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PCA's 'no accuracy loss' claim hinges on discarded low-variance directions being pure noise; the paper's own Table 2 shows large accuracy drops, so a residual-subspace probe is needed to verify that task-relevant information is not being thrown away.","rationale":"The reader's weakest assumption is exactly the one I would stress: PCA truncation is variance-based, not task-based, and the paper's own results provide direct counterexamples. The strongest counterexamples in Table 2 are large enough that they cannot be waved away as average behavior. A residual-subspace probe would settle whether the loss is due to discarded information or to something else, such as an interaction with a particular model family. Credit is due where the paper has independent support: the comparisons with downsampling, shortening, FFT/DWT, and learned reduction layers, plus the broad model/task sweep, support the narrower claim that PCA is often a cheap and harmless reducer. The appendix also honestly reports PatchTST's accuracy loss and the instance-normalization conflict, which further narrows the scope of 'maintains model accuracy.' Because the central claim is broad and the evidence is single-run, I would keep the reader's CONDITIONAL verdict; the proposed test is a concrete path to either confirming the intended scope or tightening the claim.","tokens_in":20524,"tokens_out":6273,"duration_ms":62274,"concrete_test":"For each dataset in Tables 2 and 3, project held-out windows onto the discarded subspace (eigenvectors k+1..L of the training covariance) and train a simple linear classifier or regressor on those residual coordinates to predict the label or future value. Compare its accuracy or MSE to chance and to the accuracy of the same probe on the top-k PCA coordinates. If residual-only probes exceed chance substantially, especially for SelfRegulationSCP1 or UWaveGestureLibrary, the discarded subspace contains task-relevant signal, so PCA is removing information rather than only noise and the 'no accuracy sacrifice' generalization fails for those cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim rests on the assumption that the top-k eigenvectors of the training-set temporal covariance capture all task-relevant signal, so that the discarded k+1..L subspace is noise. The paper's own Table 2 contradicts this assumption: TimesNet on SelfRegulationSCP1 drops from 0.918 to 0.686, and Linear on UWaveGestureLibrary drops from 0.575 to 0.409 after PCA. These are not small fluctuations; they show that class-discriminative information lived in the truncated subspace for those datasets. The paper provides no analysis of the residual subspace, no error bars, and no protocol for identifying when PCA truncation will hurt. Since the abstract claims PCA improves efficiency 'without sacrificing model accuracy' in general, the entire conclusion is load-bearing on the discarded components being irrelevant; the counterexamples show that this is not a universal property of the method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using Principal Component Analysis (PCA) on sliding windows of time series to reduce the temporal dimension before feeding inputs into deep learning models for time series classification, forecasting, and extrinsic regression. The authors argue that, contrary to the common belief that PCA disrupts temporal dependencies, applying PCA inside windows preserves statistical information and improves computational efficiency without sacrificing accuracy. They report experiments on 13 datasets with four model families (linear, Transformer, CNN, RNN) and compare PCA against shortening, downsampling, FFT, DWT, and learned reduction layers, as well as reporting training time and GPU memory savings.","tokens_in":20637,"tokens_out":5880,"duration_ms":51815,"significance":"If the central claim were fully supported, the paper would offer a simple, model-agnostic preprocessing step that reduces temporal dimensionality and computational cost while preserving accuracy, which would be of practical value to the time series community. The paper has strengths: it covers three tasks, multiple model families, and includes explicit efficiency measurements (Table 16, Figures 4–5) that show meaningful speedups and memory reductions. However, the accuracy claim as stated in the abstract ('without sacrificing model accuracy') is contradicted by the paper's own tables, with several large accuracy drops after PCA. The efficiency results remain credible and useful, so a revised version that honestly qualifies the accuracy claim and provides guidance on when PCA is safe would be a meaningful contribution.","major_comments":[{"comment":"The central claim that PCA improves efficiency 'without sacrificing model accuracy' is contradicted by the paper's own results. In Table 2, TimesNet on SelfRegulationSCP1 drops from 0.918 to 0.686 and Linear on UWaveGestureLibrary drops from 0.575 to 0.409 after PCA. In Table 3, FEDformer is worse with PCA in 21 of 32 metric cells. These are not small fluctuations; they show that class-discriminative and forecast-relevant information can reside in the discarded low-variance subspace. The abstract and the text in §4.1 ('PCA shows better performance in 10 metrics' and 'without performance loss') need to be qualified, for example by restricting the claim to datasets where a validation-based check confirms that the discarded subspace is not informative, or by reporting the proportion of settings where accuracy is maintained rather than an unconditional statement.","section":"Abstract; §4.1, Table 2; §4.2, Table 3"},{"comment":"The theoretical justification contains an incorrect claim: 'PCA preserves higher-order moments, including skewness and kurtosis, because its linear transformation ensures that these higher-order statistical characteristics remain intact.' A linear projection onto the top-k principal components is a rank-k map, and the skewness, kurtosis, and even the variance of the projected data are not invariant under this truncation. The same issue affects the claim about preserving 'relative mean/sum values,' since the projection is applied to mean-centered data. The denoising argument (treating low-variance components as noise) is an assumption, not a consequence of PCA, and the paper's own counterexamples in Table 2 show that this assumption fails on several datasets.","section":"§3.2"},{"comment":"The experimental comparisons report a single run per configuration with no standard deviations, confidence intervals, or significance tests. Given that the 'maintains accuracy' claim rests on small differences (e.g., 'largely unchanged' for Linear and TimesNet in Table 3), the absence of variance information makes it impossible to distinguish genuine equivalence from run-to-run noise. The authors should report results from multiple seeds with error bars, or use a paired statistical test such as the Wilcoxon signed-rank test across the datasets, before claiming that accuracy is preserved.","section":"§4, all result tables"},{"comment":"The comparison for PatchTST is not a like-for-like test of PCA. To apply PCA, the authors remove the instance normalization module from PatchTST and compare against PatchTST also without instance normalization; however, the original PatchTST with instance normalization achieves the best results, and removing instance normalization causes a large deterioration. Since instance normalization is an important component of the model, the claim that PCA is effective for patch-based models is not established. The authors should either integrate PCA with instance normalization in a way that is stable, or clearly state that PCA requires removing a beneficial component and therefore is not a drop-in replacement for all models.","section":"Appendix B.4, Table 12"}],"minor_comments":[{"comment":"The Better Count numbers do not sum consistently with the number of metric cells: for the Linear/Linear* pair, 19+17=36, but there are only 32 cells (4 datasets × 4 horizons × 2 metrics). This suggests a tabulation error that should be corrected.","section":"Table 3, 'Better Count' row"},{"comment":"The conclusion uses the abbreviation 'TSRE' while the rest of the paper uses 'TSER' (time series extrinsic regression); the notation should be made consistent.","section":"§6, Conclusion"},{"comment":"The figures report normalized training/inference time and GPU memory usage, but do not specify whether times are per epoch, per full training run, or averaged over which configurations, nor do they include error bars; the captions should clarify these details.","section":"Figures 4 and 5"},{"comment":"The FFT and DWT baselines are described as selecting the top k frequency components, but the text does not specify how complex-valued FFT coefficients are handled or whether any reconstruction is performed; the extremely poor MSE values (above 2.0) suggest a possible implementation issue that should be documented.","section":"§5, Table 7"},{"comment":"The description of the number of principal components mentions values 16, 48, and 96, but does not provide a complete mapping of which k is used for each dataset and task; a summary table would make the experimental setup reproducible.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal and the novelty of applying PCA to the temporal dimension in a general TSA pipeline is reasonable. The main problem is overclaiming: the abstract and several sections state that PCA preserves accuracy unconditionally, while the paper's own tables show substantial accuracy drops on some datasets. This is fixable by a careful rewriting of the claims and by adding a data-dependent criterion for when the truncation is safe. I do not see grounds for rejection, but the current version should not be accepted without these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: useful empirical study, overclaimed headline. The PCA-as-temporal-compression recipe largely works on average, but 'without sacrificing model accuracy' is not supported by the paper's own tables.\n\nWhat's new is the breadth: a systematic check of PCA on the temporal axis of sliding windows across 13 datasets, four model families, and three tasks, with comparisons against shortening, downsampling, FFT/DWT, and learned reduction layers. The efficiency numbers are plausible, and the paper honestly reports settings where PCA hurts (FEDformer, several classification datasets). That honesty is a real point in its favor.\n\nThe soft spots are real. The biggest is the abstract's blanket claim. Their Table 2 shows TimesNet on SelfRegulationSCP1 dropping from 0.918 to 0.686 and Linear on UWave dropping from 0.575 to 0.409 after PCA. Those are not noise; they show task-relevant signal lived in the truncated subspace. The paper should either narrow the claim to 'average accuracy is roughly maintained' or provide a diagnostic for when truncation is unsafe. Second, the related work misses decades of temporal PCA in functional data analysis, singular spectrum analysis, and climate EOF analysis; that is a novelty overclaim and must be fixed. Third, the theory section claims PCA preserves higher-order moments, which is wrong for skewness and kurtosis under linear projection. Fourth, there is no code, no seeds, no error bars; the efficiency claim rests on single runs. Fifth, the ETTh1 Linear baseline looks anomalous (0.189 vs 0.063 at horizon 96) - likely a typo, but it needs checking.\n\nThe central recipe - fit PCA on training windows, keep top-k components, feed the shorter series into the model - is simple and likely to be a useful default for many pipelines, provided users accept some datasets where it hurts. This paper is for practitioners who want a cheap preprocessing trick and are willing to verify on their own data, and for reviewers who want a case study in aligning claims with evidence. It deserves peer review, but it needs a careful revision.","headline":"Useful empirical recipe for temporal PCA compression, but the 'no accuracy loss' claim is contradicted by the paper's own tables and the theory section has a clear error.","tokens_in":21265,"tokens_out":3693,"would_cite":false,"duration_ms":31128,"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":"PCA on sliding time windows preserves accuracy while cutting training cost by up to 40%.","keywords":["principal component analysis","temporal dimensionality reduction","time series classification","time series forecasting","time series extrinsic regression","model efficiency","deep learning","denoising"],"falsifier":"A replication on a classification dataset whose class labels are carried by low-amplitude, rare, or transient features, so that the top variance directions are class-blind, would falsify the no-accuracy-sacrifice claim if PCA preprocessing produces a large, reproducible accuracy drop; the paper's own results on SelfRegulationSCP1 with TimesNet (0.918 to 0.686) and UWaveGestureLibrary with Linear (0.575 to 0.409) are already candidate instances of this failure.","tokens_in":20206,"feed_emoji":"⚡","tokens_out":9046,"duration_ms":68599,"temperature":0.7,"pith_summary":"This paper argues that Principal Component Analysis, normally avoided for time series because it scrambles the order of time steps, can be applied safely to the temporal dimension if the series are first cut into sliding windows. It claims that projecting each window onto the top-k principal components of the temporal covariance, fitted once on the training set, retains the information that forecasting, classification, and regression models actually use, while shortening inputs from, for example, 336 to 48 steps. Across Linear, Transformer, CNN, and RNN models and three tasks, the paper reports that accuracy is maintained in about half the settings and improved in many others, while training and inference speed up by 10-40% and GPU memory drops by up to 30%. If correct, PCA becomes a parameter-free, drop-in preprocessing step that cheaply compresses the temporal axis before any deep time series model is applied.","feed_headline":"PCA preprocessing cuts time series model costs up to 40%","feed_subtitle":"Projecting each window onto top principal components keeps accuracy while cutting training time and GPU memory.","key_machinery":"The mechanism is standard PCA applied to the window dimension of a dataset. The training windows are mean-centered, the temporal covariance matrix $C = \\frac{1}{n-1} D_{\\mathrm{centered}}^T D_{\\mathrm{centered}}$ is computed, and each window is projected onto the top $k$ eigenvectors: $D_{\\mathrm{pca}} = D_{\\mathrm{centered}} V_k$. The covariance and eigenvectors are estimated once on the training set and applied to validation and test windows without re-estimation. The paper's justification is that windowing preserves enough temporal structure, the projection discards low-variance noise while retaining statistical characteristics like means, peaks, and higher-order moments, and the specific trend or periodicity of a series is not what models need, since coherent patterns are enough. For patch-based models, the same projection is applied to each patch separately, such as length-16 patches reduced to 2 components, before concatenation.","core_discovery":"The paper's central claim is that PCA preprocessing of sliding series windows is a valid and efficient form of temporal dimensionality reduction for deep time series models. The authors' argument is that windowing preserves partial temporal structure, so the PCA projection does not destroy the dependencies the model needs; instead it acts as a denoiser, keeping the top variance directions and discarding low-variance components as noise. On the paper's evidence, the reduced representation keeps the mean, peak, and higher-order moment information that models rely on, so accuracy is preserved while the input is shortened. The headline numbers are up to 40% faster Informer training and inference, a 30% GPU memory reduction for TimesNet, and comparable or better accuracy in roughly half of the classification, forecasting, and regression comparisons.","pith_inferences":["Inference: The accuracy drops the paper itself reports on SelfRegulationSCP1 and UWaveGestureLibrary suggest the method will fail when discriminative information lives in small-variance directions; a class-conditional or supervised PCA variant would be a natural extension.","Inference: Because PCA is fit on training windows, it implicitly assumes the temporal covariance is stationary; on drifting or non-stationary streams an adaptive basis would need to be re-estimated periodically.","Inference: The speedups are partly a model-length effect, so the same preprocessing should transfer to other architectures whose cost scales with input length, including modern linear and state-space sequence models."],"forward_implications":["Any fixed-window TSA model can take the PCA-projected series as a drop-in input, shortening 336-step windows to 48 steps, with comparable average accuracy and 10-40% faster training and inference.","Because PCA adds no trainable parameters and is fit once on the training set, its efficiency gains do not come at the cost of model complexity, unlike adding a linear or 1D-CNN reduction layer.","For patch-based models such as PatchTST, per-patch PCA preserves much of the forecasting performance only when instance normalization is removed, so the interaction between normalization and PCA preprocessing must be handled explicitly.","Compared with shortening the input, downsampling, FFT, and DWT reductions, PCA is the only one of these methods that keeps forecasting accuracy near the unreduced baseline on ETT datasets.","RNN models benefit most on a relative basis, with training time falling to about one-fourth and inference time to about one-third of the unreduced times."],"supporting_citations":[{"why":"Supplies the PCA procedure used for the reduction.","marker":"[Pearson, 1901]"},{"why":"Supplies the Informer model, the ETT datasets, and the forecasting protocol.","marker":"[Zhou et al., 2021]"},{"why":"Supplies the TimesNet model and the evaluation setup used across the three tasks.","marker":"[Wu et al., 2023]"},{"why":"Supplies the Linear model and its published forecasting results used as a baseline.","marker":"[Zeng et al., 2023]"},{"why":"Supplies the PatchTST architecture and motivates the per-patch PCA treatment, including the instance normalization interaction.","marker":"[Nie et al., 2022]"},{"why":"Supplies the FEDformer baseline used in the comparisons.","marker":"[Zhou et al., 2022]"},{"why":"Supplies the time series extrinsic regression datasets and task definition.","marker":"[Tan et al., 2021]"},{"why":"Supplies the UEA archive classification datasets.","marker":"[Bagnall et al., 2018]"},{"why":"Supplies the GRU baseline for the RNN experiments.","marker":"[Chung et al., 2014]"},{"why":"Supplies the LSTM baseline for the RNN experiments.","marker":"[Hochreiter, 1997]"}],"fun_headline_variants":["PCA on time windows speeds up models, cuts GPU memory","Windowed PCA trims time series cost: up to 40% faster","PCA cuts time steps, not accuracy: up to 40% speedup","Temporal PCA: shorter inputs, same accuracy, 40% faster","Windowing + PCA: 40% faster training without accuracy loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The task-relevant signal in every window lies in the top-k variance directions of the temporal covariance estimated once on the training set, and that covariance stays the same between training and test windows.","fun_headline_variants_meta":{"raw":{"variants":["PCA on time windows speeds up models, cuts GPU memory","Windowed PCA trims time series cost: up to 40% faster","PCA cuts time steps, not accuracy: up to 40% speedup","Temporal PCA: shorter inputs, same accuracy, 40% faster","Windowing + PCA: 40% faster training without accuracy loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000449,"raw_usage":{"total_tokens":2275,"prompt_tokens":965,"completion_tokens":1310,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":1216}},"tokens_in":581,"tokens_out":1310,"duration_ms":9053,"temperature":1.0,"reasoning_tokens":1216,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:37:52.797692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A replication on a classification dataset whose class labels are carried by low-amplitude, rare, or transient features, so that the top variance directions are class-blind, would falsify the no-accuracy-sacrifice claim if PCA preprocessing produces a large, reproducible accuracy drop; the paper's own results on SelfRegulationSCP1 with TimesNet (0.918 to 0.686) and UWaveGestureLibrary with Linear (0.575 to 0.409) are already candidate instances of this failure.","supporting_citations":[],"review_version":1}