{"id":"35ea643b-e137-4b7a-b010-11bba3dc9c85","arxiv_id":"2505.02606","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Wavelet-based lossy compression can reduce smart-grid sensor data by up to 99% without degrading XGBoost forecasts, but linear models break under the train/test mismatch.","lead":"This paper tests whether compressing sensor data with wavelet transforms before feeding it to forecasting models hurts prediction accuracy. It finds that XGBoost stays accurate even at 99% compression, while a linear model fails when trained on smoothed data and tested on noisy data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Deployment-consistent evaluation missing: the headline tradeoff is measured with test data uncompressed while models are trained on compressed data.","rationale":"The reader's verdict is CONDITIONAL, and the protocol issue is the reason. The paper describes compression for storage and transmission; at inference time the inputs would themselves be compressed and decompressed before being fed to a forecaster. The reported protocol, training on compressed data and testing on uncompressed data, measures robustness to distribution shift rather than the practical compression/forecast tradeoff. The XGBoost stability under this harsher shift is encouraging, but it does not quantify the deployment tradeoff and it creates the misleading OLS collapse. Running the matched-compression protocol is cheap and would either strengthen the paper, by showing XGBoost remains stable and OLS recovers, or require a reinterpretation of the sensitivity claims. The other weaknesses noted by the reader, such as absent code, missing error bars, and the XGBoost-only NMI guidance, are real but secondary; they do not displace the protocol as the most load-bearing issue. Since the concern does not force a verdict change beyond the existing CONDITIONAL, the verdict remains UNCHANGED.","tokens_in":17866,"tokens_out":6430,"duration_ms":82712,"concrete_test":"Re-run the Section 3.2 experiment with the same DWT compression applied to the test segments before scoring, using identical r_lossy and wavelet for train, validation, and test, including bior6.8 at r=0.6 and r=0.99. If OLS RMSE drops from thousands to the baseline scale, the headline OLS collapse is an artifact of the mismatched protocol. If XGBoost remains stable under the matched protocol, the core XGBoost claim should be restated under the deployment-consistent protocol it actually supports.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that wavelet compression up to r=0.99 preserves forecasting accuracy—rests on Section 3.2's protocol: models are 'trained and validated using datasets subjected to lossy compression and tested on the corresponding uncompressed data.' That is not the protocol a deployed edge-compression pipeline would use: when historical context and covariates are compressed for storage or transmission, the inference-time inputs are also reconstructed from the same lossy representation. Training on compressed data and testing on raw data measures robustness to distribution shift, not the compression/forecast tradeoff. The extreme OLS failures in Table 2 (e.g., RMSE 6372.656 for bior6.8 at r=0.6) are largely a consequence of this shift: the model learns smooth short-lag inputs and then receives noisy raw ones at test. The paper itself notes that smoothing test data could mitigate this, but it does not adopt that check. The XGBoost result may survive a deployment-consistent protocol, but as reported the quantitative tradeoff and the OLS sensitivity result are not validated for the intended use case.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates the impact of discrete wavelet transform (DWT) based lossy compression on the forecasting accuracy of three time series models (OLS, XGBoost, TiDE) for a smart grid case study of a seawater supply system in Hirtshals, Denmark. Compression is performed by thresholding wavelet coefficients, with the lossy compression rate defined as the fraction of zeroed coefficients. Five biorthogonal wavelets and rates up to 0.999 are considered, and Normalized Mutual Information (NMI) is used to quantify information retention. The results show that XGBoost is largely unaffected by compression up to r=0.99, while OLS is highly sensitive to smooth wavelets and high rates, and TiDE shows intermediate variability. The paper claims that wavelet compression can substantially reduce storage and transmission requirements without sacrificing forecasting accuracy.","tokens_in":18026,"tokens_out":7285,"duration_ms":79111,"significance":"If the claims hold, the paper offers a practically useful empirical result: robust forecasting models such as XGBoost can tolerate aggressive lossy wavelet compression, enabling data volume reduction in IoT-based smart energy systems. The study is clearly written, uses a real-world dataset, and systematically explores multiple wavelets and compression rates. The use of NMI as a task-agnostic information-retention measure is a sensible idea, and the observation of XGBoost's robustness is potentially valuable. However, the central quantitative claims are not yet validated because the evaluation protocol measures distribution-shift robustness rather than the deployment compression tradeoff, and the compression-rate metric is a sparsity fraction rather than an actual storage saving. With these issues addressed, the paper could be a solid applied contribution.","major_comments":[{"comment":"The evaluation protocol trains and validates all models on lossy-compressed data but tests them on the corresponding uncompressed data. This is not the protocol a deployed edge-compression pipeline would use: at inference time, the inputs to the model would also be reconstructed from the lossy representation. As written, the experiments measure robustness to a train/test distribution shift, not the compression/forecasting tradeoff claimed in the abstract. The extreme OLS failures in Table 2 (e.g., bior6.8 at r=0.6, RMSE 6372.656) are largely a consequence of this mismatch, since the model learns smooth short-lag inputs and then receives noisy ones at test. The manuscript mentions that smoothing the test data could mitigate this but does not adopt that check, and no alternative protocol is reported. Please either add a deployment-consistent evaluation (e.g., testing on reconstructed data from the same compression pipeline) or re-frame the existing results as a distribution-shift robustness study.","section":"Section 3.2"},{"comment":"The lossy compression rate r_lossy = N0/NY is defined as the fraction of wavelet coefficients that are zeroed after thresholding. This is not the actual storage reduction achieved: the retained non-zero coefficients and their positions must still be encoded, and the approximation coefficients at the coarsest decomposition level are always stored. As a result, a reported r_lossy of 0.99 does not imply a 100x reduction in file size, and the comparison in Section 3.2 with Brotli's r_lossless=0.36 (a file-size reduction) is apples-to-oranges. The central quantitative claims about achievable compression (abstract, Section 4) rest on this metric. Please report a true bitrate- or file-size-based compression ratio, or explicitly relabel r_lossy as a sparsity fraction and temper the storage-claim language accordingly.","section":"Section 2.2.2"},{"comment":"The manuscript does not specify how the threshold epsilon is selected to achieve each target r_lossy, nor the DWT decomposition depth used. These details determine the set of retained coefficients and are essential for reproducibility and for interpreting the differences between wavelets. Without them, the reported r_lossy values are underdetermined and the experiment cannot be replicated. Please state the threshold-selection rule (e.g., quantile of coefficient magnitudes) and the decomposition level, or provide the code that implements the compression pipeline.","section":"Sections 2.2.1-2.2.2"}],"minor_comments":[{"comment":"Aggregate RMSE/MAE values are reported without confidence intervals or standard errors, despite strong variability across the 12 test datasets visible in Figure 4. Provide an uncertainty measure for the headline comparisons.","section":"Section 3.2 / Table 2"},{"comment":"Model specifications (lag lengths p and q, XGBoost hyperparameters, TiDE architecture and training details) are not reported; if library defaults were used, state this explicitly.","section":"Section 2.3"},{"comment":"The notation I(˜Y0;Y) in the NMI definition is confusing because Y and ˜Y0 both denote the original signal; clarify that the denominator is H(Y) when r=0.","section":"Section 2.2.3"},{"comment":"The x-axis is logit-scaled in Figure 5 but linear in Figure 6; use a consistent scale for easy comparison.","section":"Section 3.3 / Figures 5-6"},{"comment":"The abstract's claim that XGBoost maintains stable performance across diverse compression rates is too strong given the large error increases at r=0.999 (e.g., bior1.5 RMSE 0.402 vs. 0.041 baseline); qualify the claim to 'up to r=0.99'.","section":"Abstract / Table 2"},{"comment":"The elbow-method recommendation is a heuristic based on the same data used to fit the beta curves and observe the RMSE-NMI relationship; it is not an independent, validated selection criterion. Consider presenting it as an observation or a hypothesis for future work.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is an empirical case study; the core XGBoost-robustness observation is interesting and likely publishable, but the two methodological issues (evaluation protocol and compression metric) are significant enough to require a major revision. The authors have been transparent about data and methods; the use of LLM as a writing aid is disclosed. I recommend major_revision with the requested additional analyses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent applied study with a modest new contribution. The main thing you should know is that the central claim—XGBoost stays accurate under heavy wavelet compression—is probably true, but the paper's evaluation protocol doesn't actually test the deployment scenario it claims to address.\n\nWhat's new: systematic comparison of five biorthogonal wavelets across compression rates 0.4–0.999 for three forecasting models on a real smart-grid-like dataset. DWT compression for forecasting is not new (refs 9 and 23 cover that), and the models are off-the-shelf. The new bits are the NMI-based heuristic for picking a compression rate and the explicit robustness framing. Those are legitimate extensions, not a method breakthrough.\n\nWhat's done well: the experiments are clearly described. The NMI metric is a sensible way to quantify information retention. The authors are honest about limitations, including the potential loss of information for downstream tasks and the fact that smoothing the test data could mitigate OLS sensitivity. They acknowledge LLM writing assistance, which is refreshing. The XGBoost robustness result is consistent across Table 2, even with the protocol caveat.\n\nSoft spots: the train-on-compressed/test-on-uncompressed protocol (Section 3.2) is the big one. In a real edge-compression pipeline, if you compress for storage or transmission, the inference-time inputs are also reconstructed from the same lossy representation. Training on compressed and testing on raw data measures distribution shift, not the compression/forecast tradeoff. The paper itself notes smoothing test data as a mitigation but doesn't adopt it. The headline OLS collapse (RMSE 6373 at bior6.8, r=0.6) is mostly an artifact of this shift. XGBoost may well survive a deployment-consistent protocol, but as reported the quantitative tradeoff is not validated for the intended use case.\n\nOther issues: no error bars on the aggregate metrics, no hyperparameters for XGBoost or TiDE, no code or data release. The NMI elbow heuristic is demonstrated only for XGBoost. These are fixable but need to be addressed.\n\nWho is this for: practitioners in smart grid or IoT settings who want a rough recipe for how much wavelet compression a tree-based forecaster can tolerate. That recipe is probably serviceable, but I wouldn't generalize beyond this site.\n\nRecommendation: this deserves peer review—it's a serious empirical study with a clear question—but it needs a major revision first. The authors should add a deployment-consistent evaluation, report error bars and hyperparameters, and either release code/data or explain why not. I'd write a supportive but demanding referee report.","headline":"A competent, clearly-written applied study whose main robustness claim is likely true but is measured with a train-on-compressed/test-on-uncompressed protocol that doesn't match real deployment.","tokens_in":18619,"tokens_out":3017,"would_cite":false,"duration_ms":27845,"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":"Careful wavelet-based lossy compression can remove 99% of a smart-grid sensor stream without degrading forecast skill, and normalized mutual information tells you where the safe limit is.","keywords":["time series forecasting","lossy compression","discrete wavelet transform","biorthogonal wavelets","normalized mutual information","smart grid","XGBoost","demand forecasting"],"falsifier":"Run the identical experiment with the test set compressed at the same rate and with the same wavelet as the training set, and compare the linear model's RMSE with Table 2; if Bior6.8 at $r=0.6$ no longer gives errors around 6372, the train/test mismatch rather than the wavelet caused the collapse. A second check is to test on an independent dataset whether the fitted NMI elbow, the point where $1-I_r(\\alpha,\\beta)$ starts falling steeply, predicts the compression rate at which XGBoost RMSE begins to climb.","tokens_in":17649,"feed_emoji":"⚡","tokens_out":7133,"duration_ms":85637,"temperature":0.7,"pith_summary":"The paper tries to establish that wavelet-based lossy compression of high-frequency sensor data can cut storage and transmission volume dramatically without destroying forecasting accuracy, as long as the wavelet, compression rate, and model are chosen with care. It tests five biorthogonal wavelets at compression rates from 0.4 to 0.999 on seawater-intake forecasts from a Danish port, comparing a linear regression model, XGBoost, and TiDE. XGBoost's forecast error stays nearly flat up to 99% compression, while the linear model collapses under smooth wavelets and TiDE remains competitive but variable. Normalized mutual information, fitted to a regularized incomplete beta curve, tracks how much signal survives compression and gives a practical warning that predictive error is about to spike. The paper also states its own limitation: lossy compression can irreversibly discard features that later prove useful for other downstream tasks such as anomaly detection and diagnostics.","feed_headline":"Wavelet compression shrinks sensor data 99% without forecast loss","feed_subtitle":"XGBoost stays accurate to 99% compression; mutual information shows where to stop.","key_machinery":"The carrying object is the Discrete Wavelet Transform (DWT) with biorthogonal spline wavelets, indexed as BiorNr.Nd, where the analysis and synthesis filters are dual rather than identical, allowing perfect reconstruction with asymmetric designs. Thresholding small detail coefficients produces a sparse representation, and the lossy compression rate is defined as $r_{\\mathrm{lossy}}=N_0/N_Y$, the fraction of zeroed coefficients. The guiding metric is Normalized Mutual Information, $\\mathrm{NMI}(r)=I(\\tilde{Y}_r;Y)/I(\\tilde{Y}_0;Y)$, whose decline with compression is compressed into two parameters by fitting $1-I_r(\\alpha,\\beta)$, a regularized incomplete $\\beta$ function. This machinery supplies both the data reduction and the early-warning signal: the DWT creates the compressed training data, while the NMI curve, especially its sharp drop near $r\\to 1$, predicts where forecast error for the stable gradient-boosted model begins to climb.","core_discovery":"The central claim is that Discrete Wavelet Transform compression can keep the information needed for accurate one-to-six-hour forecasts of water intake level even when 99% of the wavelet coefficients are zeroed, and that Normalized Mutual Information (NMI) quantifies when compression starts erasing task-relevant information. With XGBoost, RMSE remains close to its uncompressed baseline from $r=0.4$ through $r=0.99$ across all five biorthogonal wavelets, with noticeable degradation only at $r=0.999$. The linear model performs best on uncompressed data but its error explodes with smooth wavelets, reaching an RMSE near 6372 for Bior6.8 at $r=0.6$, because training on smoothed short-lag inputs leaves it sensitive to high-frequency perturbations in the uncompressed test data. TiDE shows no clear pattern but stays competitive except at the most extreme compression. The paper frames NMI as a dimensionless, bounded measure of retained signal strength, fits it as $\\mathrm{NMI}(r)=1-I_r(\\alpha,\\beta)$, and recommends an elbow-style rule on the NMI or RMSE curve to choose a safe compression rate before errors rise sharply. It explicitly acknowledges that lossy compression may irreversibly discard features useful for other downstream applications.","pith_inferences":["The dramatic linear-model failures under Bior3.9 and Bior6.8 likely depend on the train-on-compressed, test-on-uncompressed protocol; if live data were compressed at the same rate as the training data, a regularized linear model might remain usable at moderate compression. The paper mentions smoothing the test data as a possible mitigation but does not test it.","The NMI-to-RMSE relationship for XGBoost suggests a testable edge-sensor design rule: compute a rolling NMI estimate on compressed batches and reduce the compression rate whenever NMI nears its sharp-drop region, rather than fixing a single rate in advance.","Because the error threshold couples NMI with model class, a linear model would likely require a more conservative NMI floor than XGBoost; one concrete test is to find, for each model, the NMI value at which RMSE rises by a fixed factor and use that as a model-specific compression budget.","The method should port to other high-frequency forecasting domains such as wind power or electricity prices, but the optimal wavelet order will probably shift with signal roughness, so the portable finding is the NMI-guided selection procedure rather than any specific Bior order."],"forward_implications":["If the claim holds, a one-minute seawater-supply sensor stream can be stored or transmitted at 1% of its raw size while an XGBoost forecaster keeps six-hour intake-level errors close to its uncompressed baseline.","Because lossless Brotli compression only reaches $r=0.36$, lossy wavelet compression becomes practically useful only above that rate; the study's tested range of $0.4$ to $0.999$ covers the regime that matters.","Wavelet choice is task-dependent: rough wavelets such as Bior1.1 preserve sharp transitions, while smooth wavelets such as Bior6.8 retain medium-frequency oscillations but introduce Gibbs artifacts near discontinuities, so the wavelet should be matched to the signal's structure.","NMI curves fitted as $1-I_r(\\alpha,\\beta)$ can be combined with an elbow rule to select a compression rate before the sharp RMSE rise, yielding an operational stopping criterion for data pipelines.","The paper's stated limitation stands: lossy compression may permanently discard features needed later for anomaly detection, diagnostics, or alternative models, so retention decisions should account for the broader ecosystem of data uses."],"supporting_citations":[{"why":"Supplies the XGBoost regressor whose stable performance across compression rates carries the paper's main empirical claim.","marker":"[6]"},{"why":"Supplies the TiDE neural architecture used as the third forecasting model.","marker":"[11]"},{"why":"Provides the survey taxonomy of time-series compression methods that frames DWT as a functional approximation approach.","marker":"[7]"},{"why":"Establishes the biorthogonal wavelet construction and its perfect-reconstruction, asymmetric-filter properties.","marker":"[8]"},{"why":"Supplies the multiresolution decomposition theory underlying the DWT implementation.","marker":"[25]"},{"why":"Provides the detailed wavelet signal-processing background for decomposition and threshold-based compression.","marker":"[26]"},{"why":"Supplies the k-nearest-neighbor mutual information estimator used to compute NMI.","marker":"[21]"},{"why":"Provides the regularized incomplete beta functions used to fit the NMI-versus-compression curves.","marker":"[12]"},{"why":"Supplies the elbow-point detection method recommended for choosing the compression rate from RMSE or NMI curves.","marker":"[31]"}],"fun_headline_variants":["99% data compression keeps XGBoost forecasts accurate","Wavelet compression: 99% less data, same forecast accuracy","Smart grid: compress sensor data 99% without forecast loss","NMI elbow reveals safe compression for forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that training a model on compressed data but testing it on uncompressed data mirrors real deployment; if the live data were also compressed at the same rate, the extreme failures of the linear model under smooth wavelets would probably disappear.","fun_headline_variants_meta":{"raw":{"variants":["99% data compression keeps XGBoost forecasts accurate","Wavelet compression: 99% less data, same forecast accuracy","Smart grid: compress sensor data 99% without forecast loss","NMI elbow reveals safe compression for forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1370,"prompt_tokens":1103,"completion_tokens":267,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":719,"completion_tokens_details":{"reasoning_tokens":200}},"tokens_in":719,"tokens_out":267,"duration_ms":3503,"temperature":1.0,"reasoning_tokens":200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:46:24.300108+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical experiment with the test set compressed at the same rate and with the same wavelet as the training set, and compare the linear model's RMSE with Table 2; if Bior6.8 at $r=0.6$ no longer gives errors around 6372, the train/test mismatch rather than the wavelet caused the collapse. A second check is to test on an independent dataset whether the fitted NMI elbow, the point where $1-I_r(\\alpha,\\beta)$ starts falling steeply, predicts the compression rate at which XGBoost RMSE begins to climb.","supporting_citations":[{"cited_title":"Time series compression survey","cited_arxiv_id":null,"evidence_quote":"Provides the survey taxonomy of time-series compression methods that frames DWT as a functional approximation approach."},{"cited_title":"A Wavelet Tour of Signal Processing","cited_arxiv_id":null,"evidence_quote":"Provides the detailed wavelet signal-processing background for decomposition and threshold-based compression."},{"cited_title":"https://dlmf.nist.gov/, Release 1.2.4 of 2025-03-15","cited_arxiv_id":null,"evidence_quote":"Provides the regularized incomplete beta functions used to fit the NMI-versus-compression curves."}],"review_version":1}