{"id":"7f0f3e8b-8782-4423-ae67-ae7507d0129e","arxiv_id":"2501.07599","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Dissolved oxygen fluctuations in the tidal Thames are q-Gaussian distributed, the fitted width decreases with distance from the sea, and the Informer transformer gives the best multi-hour forecasts.","lead":"Researchers analyzed five years of 15-minute dissolved oxygen readings from nine sites along the tidal River Thames and found that oxygen fluctuations follow heavy-tailed statistical patterns whose width changes with distance from the sea. They also show that a transformer-based model, Informer, forecasts oxygen levels up to 12 hours ahead more accurately than several baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multiplicative-EMD 'best detrending' claim compares q-Gaussian log-likelihoods across different data transformations; without a Jacobian correction the ranking is not meaningful.","rationale":"The paper aims to establish three main results: q-Gaussian fluctuations after detrending, a negative correlation between the fitted width parameter β and distance to the sea, and superior ML forecasting by LGBM/Informer. The reader's weakest assumption was the arbitrary detrending split (f = 6 h, m = 3); that is a legitimate sensitivity concern. However, a more fundamental and less speculative problem sits one step earlier: the evidence for the preferred detrending method is a log-likelihood comparison across data that have been transformed differently. Additive methods fit fluctuations in mg/L, while multiplicative methods fit fluctuations in log-space after Eq. (6). Without a Jacobian correction, the log-likelihood values are not comparable, so Fig. 2 and the abstract's 'multiplicative EMD stands out' claim are unsupported as written. The test is straightforward: apply the Jacobian correction and see whether the ranking survives. The ML forecasting section is more self-contained and includes sensible baselines, though it reports no uncertainty over repeated runs; that is a secondary concern. The β-distance trend, if real, should also receive uncertainty quantification, but the Jacobian issue is the most load-bearing because it directly undermines a headline methodological claim. I do not think this requires changing the overall CONDITIONAL verdict: the paper can be revised by redoing the likelihood comparison and reporting whether the ranking changes. I therefore keep the reader's verdict unchanged while sharpening the required revision.","tokens_in":20160,"tokens_out":7685,"duration_ms":81237,"concrete_test":"Recompute the Fig. 2 log-likelihoods on the original DO scale: for each multiplicative detrending fit, add -Σ log F_i to the reported log-likelihood (or directly maximize the Jacobian-corrected likelihood), then re-rank the four detrending methods per site. If multiplicative EMD is no longer highest for nearly all sites, the central method-comparison claim fails; report the corrected ranking and the size of the Jacobian shift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that multiplicative EMD is the most effective detrending method (Abstract; Fig. 2) is supported only by comparing q-Gaussian log-likelihoods across detrending methods. For additive decompositions (Eq. 1), fluctuations F_t are in original DO units; for multiplicative decompositions (Eq. 2), the fit is performed on log F_t after taking logarithms in Eq. (6). A log-likelihood for log F_t is not comparable to one for F_t: the implied density for the original-scale fluctuation is f_F(F) = f_log(log F)/F, so the likelihood on the original scale differs by a site- and method-dependent Jacobian factor of -Σ log F_t. The paper never applies this correction. Since DO values are O(10) mg/L, the omitted term is large and varies across sites; the ranking in Fig. 2 could change, and the 'highest log-likelihood in nearly all fittings' claim is not established. This is an internal correctness issue, not a matter of competing conventions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes five years of 15-minute water-quality data from nine monitoring sites along the tidal River Thames. It detrends dissolved oxygen (DO) time series with seasonal decomposition and empirical mode decomposition, in both additive and multiplicative forms; fits q-Gaussian distributions to the resulting fluctuations; and reports the fitted shape parameter q and scale parameter beta against distance to the sea. It then trains a LightGBM model for same-time DO prediction with SHAP feature attributions, and compares seven forecasting models, including Informer, for multi-hour DO forecasts. The central claims are that DO fluctuations are q-Gaussian with power-law tails, that multiplicative EMD is the most effective detrending method, that beta decreases linearly with distance to the sea, and that LightGBM and Informer are the best-performing machine-learning models for regression and forecasting, respectively.","tokens_in":20401,"tokens_out":4243,"duration_ms":43420,"significance":"If substantiated, the q-Gaussian characterization of DO fluctuations and the beta-distance relation would extend superstatistical water-quality analysis from a small river (the River Chess) to a large tidal estuary, and the forecasting comparison would be practically useful for environmental monitoring. The paper has genuine strengths: it uses real multi-site data over five years, describes a clear data-cleaning protocol, provides a public code repository, and fits the proposed distributions by maximum likelihood. However, the load-bearing statistical issues below—arbitrary detrending choices without sensitivity analysis, an invalid likelihood comparison, and unsupported spatial and forecasting generalizations—currently prevent full confidence in the headline claims.","major_comments":[{"comment":"The detrending parameters f = 6 hours and m = 3 modes are explicitly described as 'somewhat arbitrary and has to be made by the user' in the Detrending section. All subsequent q-Gaussian fits, the beta-distance trend, and even the downstream forecast-feature analysis depend on this split between trend and fluctuations, but the paper provides no sensitivity analysis or uncertainty quantification. A robustness check over a plausible range of f and m values, or a principled data-driven criterion for choosing them, is needed to establish that the reported q, beta, and the beta-distance relation are not artifacts of the chosen detrending parameters.","section":"Detrending (Eqs. 4-5), Superstatistical analysis (Fig. 3)"},{"comment":"The claim that multiplicative EMD is 'the most effective detrending method' is based on comparing q-Gaussian log-likelihoods across methods. For additive decompositions (Eq. 1), the fluctuation F_t is in original DO units; for multiplicative decompositions (Eq. 2), the fit is performed on log F_t after applying Eq. (6). These log-likelihoods are not comparable without a Jacobian correction: the implied density for the original-scale fluctuation is f_F(F) = f_log(log F)/F, so the likelihood on the original scale differs by the site- and method-dependent term -sum(log F_t). Because DO values are O(10) mg/L, this term is large and varies across sites, so the ranking in Fig. 2 could change. The authors should redo the comparison on a common scale or explicitly report likelihoods under a consistent transformation.","section":"Detrending (Eqs. 1, 2, 6), Fig. 2"},{"comment":"The paper claims a linear decreasing trend in the fitted width parameter beta versus distance to the sea, but this is based on nine sites with no error bars, correlation coefficient, confidence intervals, or significance test. With such a small sample, the apparent trend could be dominated by one or two sites. Please provide regression statistics (e.g., R^2, slope uncertainty, p-value) and ideally bootstrap or permutation-based confidence intervals, and also report how the trend depends on the detrending choices discussed above.","section":"Superstatistical analysis, Fig. 3 (bottom)"},{"comment":"The paper concludes that 'Informer consistently delivers superior performance' for long-term DO forecasting, but the forecasting experiments are reported for only one site (TBGP). The regression experiments use all nine sites, so the forecasting conclusion is not supported at the same breadth. Either additional sites should be included in the forecasting comparison, or the conclusion should be explicitly restricted to the TBGP site. As reported, Table 2 also gives only five-iteration averages without standard deviations, making it difficult to judge whether the differences between Informer and Repeat or LSTM are statistically meaningful.","section":"Time series forecasting, Table 2"}],"minor_comments":[{"comment":"There is a typo in the prose: 'fluctutions' should be 'fluctuations'. In addition, the site label is inconsistent between 'TCaP' and 'TCap', and 'TChp' in Table 1 appears as 'TChP' elsewhere.","section":"Detrending"},{"comment":"Reference [22], attributed to a superstatistical wind-statistics application, cites 'Highlights from this issue' in Emergency Medicine Journal; this appears to be a citation error and should be corrected.","section":"References"},{"comment":"Data-processing items 4 and 6 both describe the removal of the August 2022 oxygen-injection period; these could be consolidated to avoid redundancy.","section":"Data processing"},{"comment":"The Methods section says the number of dropped EMD modes is 'varied based on each site's trajectories, as illustrated in the code,' but the per-site m values are not reported in the manuscript; please state them in the text or a table so the analysis is reproducible without inspecting the code.","section":"Detrending methods"},{"comment":"The forecast-error metrics in Table 2 are averaged over five iterations, but no standard deviations or confidence intervals are reported; adding these would help assess whether the reported improvements are robust.","section":"Time series forecasting"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is interdisciplinary and seems better suited to an applied statistics or water-science venue than to a core ML venue, since the machine-learning components are largely standard applications. The superstatistical analysis and the possibility of a spatial beta-distance relationship are the most distinctive contributions; those parts need the statistical strengthening described in the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a useful empirical paper, not a breakthrough, and it deserves a proper referee, but two of its central statistical claims are shakier than the text admits. What's new: the first superstatistical analysis of the tidal Thames, with nine sites over 45 km; a multiplicative EMD detrending variant; a claimed beta-versus-distance-to-sea trend; and an Informer forecasting benchmark. The paper extends the authors' earlier River Chess work rather than inventing a new framework, which is fine. It does some things well: it is upfront about limitations (small number of sites, rainfall resolution, arbitrary detrending choices), it provides code, and the forecasting comparison includes sensible baselines. The SHAP analysis is a reasonable use of standard tools. The main soft spot is the log-likelihood comparison in Fig. 2. Multiplicative detrending fits a q-Gaussian to log F_t, while additive methods fit F_t directly. A log-likelihood for log F_t is not comparable to one for F_t unless you account for the Jacobian, i.e. the implied density on the original scale differs by a factor of 1/F_t, contributing -Σ log F_t to the log-likelihood. DO values are order 10 mg/L, so this term is large and site-dependent. The abstract's claim that multiplicative EMD is 'most effective' is not established by the evidence as presented. This is an internal correctness issue, not a competing convention. Second, the beta-distance correlation uses nine sites, no confidence intervals, no significance test, and the splitting into trend and fluctuation depends on f=6h and m=3, which the paper itself calls somewhat arbitrary. Different choices could change or remove the trend. The causal language about geographical influence goes beyond what a nine-point correlation supports. The forecasting section is better: multiple horizons, multiple baselines, and the Informer result is plausible. But it is one site, five runs averaged without variance, and the attention-weight interpretation is post hoc. Bottom line: the qualitative finding -- Thames DO fluctuations are heavy-tailed and well described by q-Gaussians -- likely survives. The quantitative claims about the best detrending method and the beta-distance trend need UQ and significance testing. Useful for water quality ML and superstatistics readers. Send it to review, but expect major revision.","headline":"Useful empirical extension to a big tidal river, but the headline comparisons rest on incomparable log-likelihoods and the beta-distance trend lacks uncertainty quantification.","tokens_in":20894,"tokens_out":2242,"would_cite":false,"duration_ms":21630,"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":"After detrending, dissolved-oxygen fluctuations in the tidal Thames are heavy-tailed q-Gaussians whose fitted width decreases with distance to the sea.","keywords":["dissolved oxygen","superstatistics","q-Gaussian distribution","empirical mode decomposition","LightGBM","Informer","river water quality","time series forecasting"],"falsifier":"Re-run the detrending and q-Gaussian fitting with $f=3$ hours or $f=12$ hours, or with $m=2$ or $m=4$ dropped modes; if the fitted $\\beta$ no longer shows a monotone decrease with distance to the sea, or if another heavy-tailed family fits the same fluctuations as well, the paper's central claims would be refuted.","tokens_in":19982,"feed_emoji":"💧","tokens_out":6387,"duration_ms":57781,"temperature":0.7,"pith_summary":"This paper tries to establish that the short-term swings in dissolved oxygen (DO) along the tidal River Thames are not Gaussian noise but heavy-tailed fluctuations described by q-Gaussian distributions, a family from superstatistics that arises when a system's local variance itself fluctuates. It argues that the right way to isolate those swings is a multiplicative detrending step, in which the measured signal is log-transformed and the slow component is removed by empirical mode decomposition; this method gives the highest log-likelihood in nearly all nine monitoring sites. The paper also claims that the fitted width parameter $\\beta$ of the q-Gaussian decreases linearly with distance to the sea, meaning DO variability is larger inland and smaller near the estuary. On the practical side, it finds that LightGBM predicts current DO best among regression models, with temperature, pH, and time of year as the dominant features, and that the Informer transformer slightly beats all baselines for forecasting DO 3 to 12 hours ahead. If these claims hold, water managers gain both a statistical description of extreme low-oxygen events and a forecasting model that points to specific critical hours.","feed_headline":"Thames oxygen swings are heavy-tailed q-Gaussians","feed_subtitle":"Five years of 15-minute sensor data show dissolved-oxygen fluctuations are q-Gaussian, with variability shrinking toward the sea.","key_machinery":"The load-bearing object is the q-Gaussian distribution, a two-parameter density with shape index $q$ and scale parameter $\\beta$ that generalizes the Gaussian and reduces to it at $q=1$; it is the exact marginal distribution obtained when a Gaussian with fluctuating inverse variance is integrated over a $\\chi^2$ distribution. The paper's other central mechanism is the detrending split: seasonal decomposition with a six-hour filter or empirical mode decomposition with three dropped modes, applied additively or multiplicatively via a log transform, decides what counts as fluctuation and therefore shapes every fitted $q$ and $\\beta$. For forecasting, the Informer's ProbSparse self-attention, which selects the most active queries before computing attention, is the mechanism credited with capturing the half-day periodicity of DO and producing the lowest long-horizon errors.","core_discovery":"The central discovery is that, once the seasonal and tidal trend is removed, dissolved-oxygen fluctuations at all nine Thames sites are well approximated by q-Gaussian densities with $q>1$, i.e. power-law tails that make extreme deviations more probable than a Gaussian would allow. The paper identifies multiplicative empirical mode decomposition as the detrending method that yields the best q-Gaussian fits, and reports a systematic spatial pattern: the scale parameter $\\beta$ decreases roughly linearly with distance to the sea, so sites farther inland show wider DO fluctuations. In addition, the paper presents LightGBM as the best same-time regressor for DO and the Informer model as the best multi-hour forecaster, with attention weights that concentrate on the morning-to-early-afternoon and late-evening-to-early-morning windows of the most recent half-day.","pith_inferences":["Because the reported $q$ values often exceed $5/3$, the fitted model has infinite variance at those sites; a practical consequence not spelled out in the paper is that confidence intervals and risk metrics based on the sample standard deviation will systematically understate the chance of extreme DO drops.","The same multiplicative-EMD plus q-Gaussian pipeline could be tested on other observables, such as electrical conductivity, on the same Thames data; a negative result there would show how far the superstatistical pattern extends beyond DO.","A direct test of the sea-distance mechanism would use the same analysis on a river without tidal seawater influence, or on the non-tidal Thames upstream, to check whether the $\\beta$ gradient is driven by seawater mixing rather than by distance alone.","A natural follow-up would be to train the Informer on all nine sites and compare attention patterns across sites, testing whether the half-day windows are generic rather than specific to the one site used for the forecasting evaluation."],"forward_implications":["If the q-Gaussian description is right, the probability of extreme low-oxygen events is higher than a Gaussian model would predict, and the second moment may not exist at sites with $q>5/3$.","If the $\\beta$-versus-distance trend is real, inland sites are where DO variability is largest, so those are the sites where monitoring and intervention would matter most.","The Informer's attention pattern implies that forecasts can be interpreted: the model relies on morning and late-evening windows and on the 16th to 26th quarter-hour of the previous half-day, which could be used to schedule oxygen injection or treatment releases.","If LightGBM with SHAP is used in practice, temperature and pH sensors should be prioritized, since they dominate same-time DO prediction across sites."],"supporting_citations":[{"why":"Defines superstatistics and shows that integrating Gaussians with fluctuating inverse variance yields q-Gaussian marginals; this is the conceptual foundation for the paper's fluctuation model.","marker":"[3,4]"},{"why":"Establishes that water-quality time series in the River Chess follow superstatistics; the result the Thames study extends to a larger tidal river.","marker":"[29]"},{"why":"Introduces the q-Gaussian/Tsallis distribution family used for all fluctuation fits.","marker":"[45]"},{"why":"Applies LGBM and feature-importance analysis to River Chess water quality; supplies the baseline and method comparison for the Thames regression results.","marker":"[30]"},{"why":"Introduces LightGBM, the gradient-boosting model reported as best for same-time DO prediction.","marker":"[47]"},{"why":"Introduces SHAP values, the attribution method used to rank temperature, pH, and time of year.","marker":"[48]"},{"why":"Introduces the Informer model and its ProbSparse self-attention mechanism, the forecasting architecture the paper finds best for long horizons.","marker":"[50]"},{"why":"Provides the telemetry data from the nine Thames monitoring sites that all analyses are based on.","marker":"[42]"},{"why":"Documents the August 2022 oxygen-injection event whose data are excluded to avoid bias.","marker":"[61]"}],"fun_headline_variants":["Thames oxygen swings follow q-Gaussian distributions","Inland Thames sites show wider oxygen fluctuations","Informer best for Thames oxygen forecasting","q-Gaussian tail explains Thames oxygen extremes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis assumes that a six-hour seasonal filter and three dropped EMD modes correctly separate 'trend' from 'fluctuation'; the paper itself calls this choice somewhat arbitrary, and every fitted q-Gaussian parameter and the beta-distance trend would change if that split were made differently.","fun_headline_variants_meta":{"raw":{"variants":["Thames oxygen swings follow q-Gaussian distributions","Inland Thames sites show wider oxygen fluctuations","Informer best for Thames oxygen forecasting","q-Gaussian tail explains Thames oxygen extremes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000622,"raw_usage":{"total_tokens":2916,"prompt_tokens":1015,"completion_tokens":1901,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":1846}},"tokens_in":631,"tokens_out":1901,"duration_ms":15404,"temperature":1.0,"reasoning_tokens":1846,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:05:29.335276+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the detrending and q-Gaussian fitting with $f=3$ hours or $f=12$ hours, or with $m=2$ or $m=4$ dropped modes; if the fitted $\\beta$ no longer shows a monotone decrease with distance to the sea, or if another heavy-tailed family fits the same fluctuations as well, the paper's central claims would be refuted.","supporting_citations":[{"cited_title":"M., Rhys, H","cited_arxiv_id":null,"evidence_quote":"Establishes that water-quality time series in the River Chess follow superstatistics; the result the Thames study extends to a larger tidal river."},{"cited_title":"Possible generalization of boltzmann-gibbs statistics","cited_arxiv_id":null,"evidence_quote":"Introduces the q-Gaussian/Tsallis distribution family used for all fluctuation fits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Applies LGBM and feature-importance analysis to River Chess water quality; supplies the baseline and method comparison for the Thames regression results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces LightGBM, the gradient-boosting model reported as best for same-time DO prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Informer model and its ProbSparse self-attention mechanism, the forecasting architecture the paper finds best for long horizons."},{"cited_title":"Water quality monitoring systems & services","cited_arxiv_id":null,"evidence_quote":"Provides the telemetry data from the nine Thames monitoring sites that all analyses are based on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the August 2022 oxygen-injection event whose data are excluded to avoid bias."}],"review_version":1}