{"id":"875a0654-f0e7-4118-a6e6-9f3612fbb31c","arxiv_id":"2508.14069","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors claim Gaussian-interpolated sparse campus load data enables load forecasting, with LSTM achieving a 10.67% MAPE, but the reported accuracy is measured against the very imputations the method creates.","lead":"The paper imputes 62% missing power-plant load data with per-hour Gaussian sampling, then compares ten forecasting models on the filled-in series. It reports LSTM as the most accurate, but the evaluation uses the same synthetic imputations as ground truth, so the headline result does not reflect real forecasting skill.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test set 'actuals' include synthetic values drawn from the same per-hour Gaussian distributions used to train the forecasters, so the reported 10.67% MAPE measures fit to the imputation rule, not accuracy on real load.","rationale":"The reader's weakest assumption is that imputed values are treated as ground truth for training and evaluation; this is exactly the load-bearing concern I identify. If the test set were restricted to genuinely observed hours, the headline MAPE could change dramatically, and the paper provides no such analysis. The circularity is direct: per-hour Gaussian parameters are estimated from observed data, the same distributions generate the test 'actuals,' and the models are trained on the same mixture. This is not an external consensus disagreement but an internal validity problem. The paper's own statistical tests (KPSS rejecting stationarity, chi-squared rejecting Gaussianity) further undermine the imputation model, and the CLT justification in Section 3.1 is not valid for dependent data. No independent evidence (e.g., code release, comparison to other imputation methods) offsets this. I agree with the reader's rejection: the central claim as stated is unsupported. The verdict remains REJECT, which is unchanged from the reader's verdict.","tokens_in":11047,"tokens_out":3420,"duration_ms":36601,"concrete_test":"Re-run the BUET evaluation on the test set after removing all imputed hours; compute MAPE and MAE only on hours that were originally observed. Additionally, hold out a contiguous block of observed hours (e.g., four weeks), mask them as missing, impute them with the paper's per-hour Gaussian procedure trained only on preceding observed data, forecast with the trained LSTM, and compare against the true observed values. If observed-only MAPE is substantially above 10.67% or the masked-holdout imputation degrades accuracy materially, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Gaussian interpolation is a 'suitable option' for load forecasting on a 62% sparse dataset, supported by the LSTM's 10.67% MAPE. The evaluation is circular: in Section 2.2, missing hours are filled by sampling per-hour Gaussian PDFs estimated from observed data, and in Section 2.3 the imputed dataset is split into train/validation/test. The test set therefore consists largely of synthetic draws from the same distributions that generated the training data. A model that simply learns each hour's conditional mean will achieve low MAPE on such a test set even if it has no skill at forecasting actual load. The paper never reports performance restricted to originally observed test hours, so we cannot tell whether the low error reflects real predictive ability. The Dayton experiment (Section 3.7) only validates the LSTM on a clean dataset; it does not validate the imputation method. Internal evidence compounds the concern: Section 3.1 reports a KPSS p-value below 0.05, rejecting the stationarity that WSS requires, and a chi-squared test rejecting Gaussianity; the CLT invocation is unjustified for dependent, non-identically distributed data. No comparison to standard imputation baselines (linear, spline, KNN, MICE) supports the claim that Gaussian interpolation is suitable relative to alternatives. Thus the paper's headline conclusion rests on a circular evaluation and contradicted assumptions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Gaussian-interpolation approach for an hourly electrical load dataset of the BUET campus that is approximately 62% missing. Missing values are imputed by fitting a per-hour Gaussian distribution to the observed values and sampling from it, with samples clipped to [0, 2000] kW. The completed series is used to train and evaluate LSTM, BiLSTM, CNN-LSTM, CNN-BiLSTM, Transformer, Prophet, DLinear, SARIMA, XGBoost, and Random Forest models, with LSTM reported as best at 10.67% MAPE. The authors also train the LSTM on the complete Dayton dataset and report 1.55% MAPE, and conclude that Gaussian interpolation is suitable for load forecasting on sparse data.","tokens_in":11293,"tokens_out":4634,"duration_ms":45945,"significance":"If the reported results were valid, the paper would offer a practical route to forecasting from very sparse load data, and the broad model comparison would be a useful reference. The paper is clearly organized and states its imputation procedure and statistical tests explicitly. However, the central evaluation does not support the claims: the test set on the BUET data consists mainly of synthetic draws generated by the same per-hour Gaussian model used to create the training targets, so the reported MAPE values measure agreement with the imputation rule rather than forecasting skill. The internal stationarity and normality tests contradict the assumptions used to justify the imputation, and the Dayton experiment does not validate the imputation method. The headline conclusion therefore rests on circular evidence.","major_comments":[{"comment":"The evaluation protocol in Sections 2.2 and 2.3 is circular. Section 2.2 fills every missing hour by sampling X = mu_h + sigma_h Z, where (mu_h, sigma_h) are estimated from the observed values; Section 2.3 then splits the completed dataset into training, validation, and test sets. Because the 'actual' values in the test set are synthetic draws from the same per-hour distributions that generated the training labels, a model that learns the hourly conditional mean will achieve low MAPE without any real forecasting ability. The paper never reports performance restricted to the originally observed test hours, so the 10.67% MAPE in Section 3.4 cannot be interpreted as load-forecasting accuracy.","section":"2.2-2.3"},{"comment":"The statistical evidence reported in Section 3.1 contradicts the WSS/Gaussian assumptions that motivate the imputation. The KPSS test returns p<0.05, rejecting stationarity, which is incompatible with the WSS assumption invoked in Section 2.2; the chi-squared test rejects Gaussianity, and the appeal to the Central Limit Theorem is not justified for a strongly autocorrelated, non-identically distributed series (Fig. 3). The paper states these results but does not resolve the inconsistency before using the Gaussian-imputed data as ground truth.","section":"Section 3.1"},{"comment":"The Dayton experiment in Section 3.7 does not validate the interpolation method. The 1.55% MAPE is obtained on a completely observed dataset and therefore provides no evidence about Gaussian imputation on the sparse BUET data. The statement that this result 'validates our approach' (Section 3.7) overstates what the experiment can show.","section":"Section 3.7"},{"comment":"No comparison is made with alternative imputation methods under the same forecasting protocol. Section 2.2 asserts that linear, polynomial, or spline interpolation 'could not be used', but no experiments support this claim; without baselines such as linear interpolation, splines, KNN, or MICE, the paper's conclusion that Gaussian interpolation is 'suitable' is not supported.","section":"Section 2.2"}],"minor_comments":[{"comment":"MAPE is undefined when y_i = 0, which can occur after clipping negative samples to zero in Section 2.2; the authors should specify how such cases are handled in the error computation.","section":"Equation (3)"},{"comment":"The paper uses WSS to justify per-hour Gaussian parameters, but WSS requires time-invariant mean and autocovariance; per-hour varying (mu_h, sigma_h) is better described as a periodic or cyclostationary assumption, and this distinction should be clarified.","section":"Section 2.2"},{"comment":"Section 3.7 reports MAE as '31.06%' without units; MAE for load data should be expressed in kW or as a percentage only if clearly defined.","section":"Section 3.7"},{"comment":"The 'supplementary section S1' referenced at the start of Section 2 is not included in the submitted manuscript; either provide the supplementary material or remove the reference.","section":"Section 2"},{"comment":"Figure 2 is schematic and the caption does not describe the workflow shown; a short explanation in the caption or text would improve readability.","section":"Figure 2"}],"recommendation":"reject","confidential_remarks":"The circular evaluation is the decisive issue: the BUET test set is generated by the same Gaussian imputation rule used to create the training targets, so the reported MAPE values cannot support the paper's central claim. The Dayton result is irrelevant to validating the imputation, and the paper's own tests contradict its modeling assumptions. I see no local fix that would preserve the conclusions; a fundamentally different evaluation design, including performance on originally observed test hours and comparisons with other imputation methods, would be needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read of arXiv:2508.14069. The headline is that the key evaluation is circular. The BUET test set 'actuals' are Gaussian samples drawn from per-hour distributions fitted on the same data used to train the models, so the reported 10.67% MAPE tells you how well the LSTM learns the imputation rule, not how well it forecasts real load. That's a load-bearing flaw, and it's not buried in the supplement; it's Sections 2.2–2.3 of the method.\n\nWhat the paper does well: it's clearly written, the experimental sweep is broad (ten models), and it honestly reports the KPSS rejection of stationarity and the chi-squared rejection of Gaussianity, even if it then waves them away with an unjustified appeal to the Central Limit Theorem. The Dayton experiment does show that an LSTM gets a 1.55% MAPE on a clean 2004–2018 dataset, but that only validates the off-the-shelf model, not the interpolation scheme.\n\nThe soft spots are not minor. The WSS assumption is contradicted by the paper's own KPSS result; the Gaussianity is rejected but dismissed without a substantive argument; no interpolation baseline (linear, spline, KNN, MICE) is compared, so the 'Gaussian interpolation is suitable' claim is ungrounded. The CLT invocation is doubly wrong: the draws are independent, so averaging wouldn't apply to a single sample, and the data are dependent and non-identically distributed across hours. The clipping to [0, 2000 kW] also means the 'Gaussian' samples are not truly Gaussian at the boundaries, another unexamined inconsistency.\n\nWho gets value from this? Someone cataloguing evaluation pitfalls in time-series imputation could cite it as a cautionary case. But as a contribution to load forecasting, the method is standard stochastic imputation, and the evaluation doesn't demonstrate any benefit.\n\nMy verdict: desk reject. The authors would need to re-run on originally observed test hours only, compare with standard imputation baselines, and stop claiming WSS when their own tests reject it. I'd tell them to go back to the data.","headline":"Circular evaluation sinks the paper: the BUET test 'actuals' are Gaussian samples from the same per-hour distributions used for training, so the reported MAPE measures fit to the imputation rule, not forecasting skill.","tokens_in":11911,"tokens_out":2258,"would_cite":false,"duration_ms":22658,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Gaussian imputation lets a 62%-sparse load record train a usable LSTM forecaster.","keywords":["load forecasting","sparsity","Gaussian interpolation","LSTM","BiLSTM","Transformer","wide-sense stationary","missing data imputation"],"falsifier":"Hold out a random subset of originally observed hours before imputation, train the LSTM exactly as in the paper, and score it only on those held-out observed hours; if the MAPE on genuinely observed load departs substantially from 10.67%, the published figure is an artifact of the Gaussian imputation. A second check is to refit the same pipeline on a complete real load record from the same plant and ask whether forecasts of true future hours reach comparable error.","tokens_in":10805,"feed_emoji":"⚡","tokens_out":11042,"duration_ms":104727,"temperature":0.7,"pith_summary":"This paper asks whether a real hourly electrical load record with about 62% of entries missing can still be used for load forecasting. The answer it defends is yes: if the series is treated as wide-sense stationary at the hourly scale, each hour's missing values can be imputed by sampling from a Gaussian whose mean and variance are estimated from the values that were observed. After that imputation, an LSTM network forecasts the completed series with roughly 10.7% mean absolute percentage error, better than nine alternative forecasting models. The authors conclude that Gaussian interpolation is a workable option for sparse, mildly non-stationary load data.","feed_headline":"Gaussian imputation turns 62%-sparse load data into forecasts","feed_subtitle":"An LSTM reaches about 10.7% MAPE on the filled-in hourly series, beating nine rival models.","key_machinery":"The load-bearing mechanism is hourly Gaussian imputation: for each hour of the day at which the plant operates, estimate $\\mu$ and $\\sigma$ from the existing load values for that hour, then fill every missing entry by drawing $X = \\mu + \\sigma Z$ with $Z \\sim \\mathcal{N}(0,1)$, clipping samples above the plant's generating capacity to that capacity and negative samples to zero. This single rule produces the complete series on which all ten forecasters are trained and scored, so the reported accuracy numbers are accuracy on the imputed record, not on the original sparse record.","core_discovery":"The paper's central claim is that per-hour Gaussian imputation converts a record that is roughly 62% empty into a series on which standard load forecasters train successfully, provided one accepts the data as wide-sense stationary within each hour. The empirical core is a model comparison on the interpolated series: LSTM reaches 10.67% MAPE, ahead of Transformer (11.02%), BiLSTM and DLinear (13.13%), Random Forest (14.56%), Prophet (17.73%), XGBoost (18.11%), and SARIMA (20.40%). The same LSTM on a dense public hourly load benchmark yields 1.55% MAPE, which the authors read as confirming that the interpolation, rather than the model choice, dominates the error on the sparse campus data.","pith_inferences":["Editorial inference: the method would be tested more honestly by holding out originally observed hours and ignoring imputed entries when computing error; the current protocol cannot separate skill at forecasting real load from skill at reproducing Gaussian draws.","Editorial inference: the approach transfers naturally to other sparse infrastructure time series such as water, gas, or traffic demand, where gaps are also non-random and the same synthetic-fit risk would appear.","Editorial inference: when imputation is misspecified, comparing models on the imputed series ranks their ability to track the imputation rule rather than their ability to track true load, so model rankings should be re-verified on dense data."],"forward_implications":["A roughly 62%-sparse hourly load record can be made trainable for short-horizon forecasting by per-hour Gaussian imputation, with no strict stationarity required.","On the imputed series, recurrent architectures, especially LSTM, outperform classical statistical, tree, and shallow-linear baselines, so sequence structure carries most of the predictive signal.","The large gap between the sparse-campus error (10.67% MAPE) and the dense-benchmark error (1.55% MAPE) implies that interpolation quality, not model family, dominates the forecasting error.","The completed series shows high autocorrelation over long lags, so the forecasting models learn from smooth, synthetic structure rather than from a naturally dense record."],"supporting_citations":[{"why":"Defines wide-sense stationarity, the assumption that licenses estimating per-hour Gaussian parameters from the observed values.","marker":"[29]"},{"why":"Supplies the Gaussian density and the mean-and-scale sampling rule used to fill every missing load value.","marker":"[31]"},{"why":"Donates the dense public hourly load benchmark on which the same LSTM is retrained for external comparison.","marker":"[45]"},{"why":"Supplies the unit-root test whose p-value the paper cites as evidence of stationarity.","marker":"[49]"},{"why":"Supplies the stationarity test whose rejection the paper must reconcile with its wide-sense stationarity assumption.","marker":"[50]"},{"why":"Supplies kernel density estimation, the non-parametric comparison used to prefer the Gaussian over the Rayleigh fit.","marker":"[53]"},{"why":"Supplies the chi-squared goodness-of-fit test whose rejection of Gaussianity the paper overrides.","marker":"[54]"},{"why":"Supplies the central limit theorem invoked to justify treating the distribution as Gaussian after the chi-squared test rejects it.","marker":"[55]"}],"fun_headline_variants":["Gaussian fill makes 62% sparse load data forecastable","LSTM and Gaussian imputation beat rivals on 62% sparse load","Impute with Gaussian, forecast with LSTM: 62% sparse data works","Gaussian interpolation + LSTM: best on highly sparse load","Sparse load data? Gaussian imputation plus LSTM wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each missing load value can be modeled as an independent draw from a per-hour Gaussian whose mean and variance come from the observed hours, with those synthetic draws then treated as ground truth for training and scoring forecasters; the paper's own stationarity test rejects stationarity and its chi-squared test rejects Gaussianity, so if missing periods follow different dynamics or retain temporal autocorrelation, the reported errors measure fit to the imputation rule rather than fit to real load.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian fill makes 62% sparse load data forecastable","LSTM and Gaussian imputation beat rivals on 62% sparse load","Impute with Gaussian, forecast with LSTM: 62% sparse data works","Gaussian interpolation + LSTM: best on highly sparse load","Sparse load data? Gaussian imputation plus LSTM wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00042,"raw_usage":{"total_tokens":2138,"prompt_tokens":897,"completion_tokens":1241,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1149}},"tokens_in":513,"tokens_out":1241,"duration_ms":9393,"temperature":1.0,"reasoning_tokens":1149,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:34:49.878897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a random subset of originally observed hours before imputation, train the LSTM exactly as in the paper, and score it only on those held-out observed hours; if the MAPE on genuinely observed load departs substantially from 10.67%, the published figure is an artifact of the Gaussian imputation. A second check is to refit the same pipeline on a complete real load record from the same plant and ask whether forecasts of true future hours reach comparable error.","supporting_citations":[{"cited_title":"Edition, A","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian density and the mean-and-scale sampling rule used to fill every missing load value."},{"cited_title":"Molla, Pjm hourly energy consumption data, https://www.kaggle.com/ datasets/robikscube/hourly-energy-consumption/data (2018)","cited_arxiv_id":null,"evidence_quote":"Donates the dense public hourly load benchmark on which the same LSTM is retrained for external comparison."},{"cited_title":"Kwiatkowski, P","cited_arxiv_id":null,"evidence_quote":"Supplies the stationarity test whose rejection the paper must reconcile with its wide-sense stationarity assumption."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies kernel density estimation, the non-parametric comparison used to prefer the Gaussian over the Rayleigh fit."}],"review_version":2}