{"id":"059a7d04-6b44-4171-97c1-48d7ddf817c0","arxiv_id":"2505.06310","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"An adaptive Bayesian forecaster that updates the generalised logit shape parameter online improved average CRPS skill over fixed-shape and classical adaptive baselines across 101 UK wind farms.","lead":"A new Bayesian forecaster that adaptively tunes the shape of a data transformation achieved small but consistent probabilistic forecasting gains over existing wind power methods across 101 UK wind farms. It offers grid operators a robust, easy-to-update way to turn noisy power data into calibrated uncertainty estimates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ν-update in §4.2 is not well-defined as written: it conflates precision matrices, inverse-transforms a Cholesky factor as if it were lagged data, and has no posterior justification, so the small Bayes-ν improvement rests on an unvalidated heuristic.","rationale":"Credit where due: the evaluation is extensive (101 usable farms, three test years, consistent ranking direction in Table 1, same ranking in Appendix A), and the sensitivity analysis provides real evidence that the implemented Bayes-ν is robust to perturbations. These facts suggest the underlying approach has value. However, they do not by themselves establish the paper's central methodological claim, because the procedure that supposedly produces the adaptive ν is not fully defined in the text. The reader's weakest assumption correctly targets the ν update; I go slightly further, arguing that the issue is not only 'representativeness of pseudo-observations' but internal dimension and notation incoherence. Repairability matters: if the authors can provide a coherent statement of the update and release code, the conditional acceptance is appropriate; if not, the empirical comparison cannot be audited. Missing significance testing is a secondary but real concern that compounds the fragility of the small Table 1 differences. I therefore keep the reader's CONDITIONAL verdict rather than moving it. The critique is directed at the argument, not at the authors.","tokens_in":18118,"tokens_out":10380,"duration_ms":103638,"concrete_test":"Implement the ν-update exactly as written on one wind farm (e.g., 2021 test year) and print the shapes of P*_{z,t+M}, L, y*, X*_B, y_ν and Y_{B,ν} at the first update step. The method is reproducible only if y_ν is a vector whose length equals the number of rows of Y_{B,ν}, and Y_{B,ν} is a lagged AR design matrix with p+1 columns. If these shapes cannot be reconciled with the AR(p) likelihood (4.16), the published update is not executable, and the reported Bayes-ν results cannot be attributed to the described method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 is the only component that distinguishes Bayes-ν from Bayes, so the claimed superiority of Bayes-ν depends entirely on the ν-update. The update is not a derived posterior procedure, and as written it is internally under-specified. In §4.1, P_z is the M×M observation-error precision matrix, while the posterior precision of θ is P*_θ from (4.2). Section 4.2 then says, 'let P*_z,t+M be the updated precision for the model parameter θ*,' and Cholesky-decomposes that object in (4.14), conflating the two precision objects. Eq. (4.15) then defines x* := L^{-1}_ν(y*) and X*_B := L^{-1}_ν(L), where L is the Cholesky factor of a precision matrix—a square matrix, not a vector of lagged observations. Inverse-transforming L and using it as the AR design matrix in (4.16) is not consistent with any AR(p) likelihood of the form (2.10); no choice of dimensions is specified that makes the objects in (4.16) a valid response vector and design matrix. Even if the notation is repaired, there is no argument that minimising the likelihood of quantities built from posterior first moments optimises any posterior or marginal likelihood for ν. The empirical support for this heuristic is thin: relative to fixed-ν Bayes, Bayes-ν gains only 0.181, 0.072 and 0.275 percentage points of Skill Score in Table 1, with no standard errors or tests, and Table 2 shows Bayes-ν placed in ranks 6-7 forty times versus five for Bayes. Thus the headline claim rests on a mechanism that is both mathematically underspecified and statistically unquantified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an adaptive Bayesian method for very short-term wind power forecasting. The method applies a generalised logit transformation with a shape parameter ν to convert double-bounded wind power data to an unbounded domain, fits an autoregressive model in the transformed domain, and updates the parameters online. The main novelty is an adaptive update of the transformation shape parameter ν (the Bayes-ν method), which is meant to leverage Bayesian posterior summaries to reconstruct 'representative data' and then optimise ν via a combined negative log-likelihood objective. The authors evaluate seven forecasting methods over 101 wind farms in Great Britain (with an auxiliary analysis over all 128 farms) across three test years, using CRPS, CRPS Skill Score, and functional reliability diagrams. The headline empirical claim is that Bayes-ν achieves the lowest average CRPS and the highest Skill Score in every test year, with the same ranking of methods reproduced in Appendix A on the full dataset.","tokens_in":18537,"tokens_out":5852,"duration_ms":59802,"significance":"If the proposed adaptive Bayesian framework is valid, it would be a practically relevant contribution to probabilistic wind power forecasting, particularly because the case study is unusually large (101–128 wind farms, three test years) and the comparison includes both classical adaptive methods (RLS, NR) and Bayesian alternatives. The manuscript has clear strengths: the data source is described, the evaluation protocol is rolling and out-of-sample, the main ranking is reproduced on an additional dataset, and the use of functional reliability diagrams to summarise calibration across many wind farms is a useful visualisation. The consistency of the average CRPS ranking across all three test years and across the full dataset is a genuine empirical finding. However, the central methodological innovation — the adaptive ν-update — is not derived from first principles and, as written, contains internal inconsistencies that prevent the method from being replicated exactly.","major_comments":[{"comment":"The adaptive ν-update is not a well-defined statistical procedure as written. The text states that P*_{z,t+M} is 'the updated precision for the model parameter θ*', but in Section 4.1 P_z is the M×M observation-error precision matrix, whereas the posterior precision of θ is P*_θ from Eq. (4.2). The Cholesky factor L of the stated matrix is a square matrix; defining x* := L^{-1}_ν(y*) and X*_B := L^{-1}_ν(L) and then treating X*_B as an AR design matrix in Eq. (4.16) is not consistent with any AR(p) likelihood of the form in Eq. (2.10), because the entries of L are not lagged observations and the required dimensions are not specified. Even if the notation is repaired, no argument is given that minimising the negative log-likelihood of quantities constructed from posterior first moments optimises a posterior or marginal likelihood for ν. Because this update is the only component that distinguishes Bayes-ν from Bayes, the claimed superiority of Bayes-ν depends entirely on an unvalidated heuristic. The authors should either provide a rigorous derivation of the update as an approximate Bayesian procedure or clearly label it as an ad hoc rule, and then validate it through simulation or by comparison with a fixed ν tuned on a validation period.","section":"Section 4.2, Eqs. (4.14)–(4.17)"},{"comment":"The headline claim that 'Bayes-ν achieves the lowest average CRPS and the highest Skill Score' is supported only by point estimates of averages. The differences between Bayes-ν and Bayes are small: 0.181, 0.072 and 0.275 percentage points of Skill Score in the three test years. No standard errors, confidence intervals, or significance tests are reported, so it is unclear whether these differences are distinguishable from noise. Table 2 further shows that Bayes-ν is ranked 6 or 7 in 40 cases, compared with only 5 for Bayes, indicating that the adaptive ν-update can substantially degrade performance on some wind farms. The authors should provide uncertainty measures for the average CRPS and Skill Score differences, for example via bootstrapping or Diebold–Mariano-type tests, and discuss the practical significance of the observed average gains in light of the variability across farms.","section":"Section 5.4, Table 1 and Table 2"},{"comment":"The manuscript interprets the differences between the Bayes-versus-AR-Lν and Bayes-ν-versus-Bayes Skill Score gaps as 'verifying the implication' that the benefit of adaptive ν depends on how ν varies in reality. This is not a verification but a descriptive observation about three test years. The multi-peaked Skill Score densities and the negative Skill Scores visible in Figure 4 are attributed to missing data and boundary effects, but no quantitative analysis is provided to support this attribution. The authors should either provide formal evidence for the stated cause of the negative Skill Scores or temper the claim. Similarly, the conclusion that adaptive ν estimation 'further enhances' performance should be phrased as conditional on the specific datasets and should acknowledge the inconsistency across wind farms.","section":"Section 5.4, Figure 4 and discussion"},{"comment":"The sensitivity analysis is conducted on a single wind farm with an unspecified disturbance magnitude and limited details on the simulation setup (e.g., the distribution of the added noise, the number of Monte Carlo draws, and the exact definition of the disturbances to μ, σ_z and P). Since the paper uses robustness as a key selling point of the Bayesian method, this analysis should be described precisely enough to be reproduced, and the conclusions should be tempered to reflect that they are based on one farm. The claim that disturbances to P and ν have effects 'extremely small, at levels of 10^-6% and 10^-3%' needs an explanation of how such small numbers arise and whether they are practically meaningful.","section":"Section 5.7, Sensitivity analysis"}],"minor_comments":[{"comment":"There is a typo in Eq. (2.5): 'ω_0 =' appears mid-sentence and the sentence is incomplete. The text defining the inflated distribution is also a little repetitive and should be rewritten.","section":"Section 2.1, Eq. (2.5)"},{"comment":"There are duplicate entries: Gneiting (2008a) and Gneiting (2008b) are identical, and Messner and Pinson (2019a) and (2019b) are identical. These references should be merged or corrected.","section":"References"},{"comment":"'Coordinate descant optimisation' should read 'coordinate descent optimisation'. Also, the outlier exclusion rule wq > σ_w^2 appears ad hoc; the authors should provide more justification for this specific threshold.","section":"Section 5.3, Implementation"},{"comment":"The note under Table 2 says 'For Rank 1 to 3, the best result is highlighted in bold', but the caption is ambiguous about the meaning of bold across different columns. Clarify how the bold entries should be read.","section":"Section 5.4, Table 2"},{"comment":"The sentence 'The differences in average Skill Scores across the three test datasets between Bayes and AR-Lν (respectively 0.41%, 0.01%, 0.23%) are both greater and smaller then that between Bayes-ν and Bayes' contains a typo ('then' should be 'than') and is confusing because the percentages are not clearly aligned with the test years.","section":"Section 5.4, discussion"},{"comment":"The notation P_z,t is used both for an M×M diagonal matrix of observation-error precisions and, later, for a scalar Gamma-distributed precision parameter. This dual use makes Section 4.2 harder to follow and should be cleaned up with distinct symbols for the matrix and the scalar.","section":"Section 4.1, Eqs. (4.2)–(4.11)"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is large and the basic ranking of methods is consistent across test years and datasets, which is a real strength. The core methodological novelty, however, is not yet rigorous enough for publication: Section 4.2 requires either a proper derivation or an explicit re-framing as a heuristic with supporting validation. The small average gains of Bayes-ν over Bayes, together with the absence of uncertainty quantification, make me suggest a major revision rather than acceptance at this stage. The authors also should be asked to report the exact number of wind farms used in the main analysis (101) consistently with the abstract (which says 'over 100 wind farms')."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth your time, but keep your expectations calibrated. The paper does two things well. First, it sets up a clean Bayesian conjugate updating scheme for an AR model on generalized-logit-transformed wind power data, with forgetting factors that let the prior adapt. That part is sensible and clearly explained. Second, the empirical evaluation is genuinely extensive: 101 farms, three holdout years, consistent ranking across those years and in the appendix with all 128 farms, plus a useful functional reliability diagram that summarises calibration across farms. The violin plots and rank tables give a fair picture of where each method wins and loses. If the fixed-ν Bayes method were all they claimed, I would be happy to see it in print.\n\nThe problem is the extra claim that adaptively updating ν is a meaningful improvement. The adaptive ν-update in Section 4.2 is not well-defined as written. The notation conflates the observation-error precision matrix with the posterior precision of θ; the Cholesky factor L of a precision matrix is then inverse-transformed and used as if it were a matrix of lagged observations, which does not line up dimensionally with any AR(p) likelihood. The reconstructed data are built from posterior first moments without any argument that the resulting objective relates to a marginal likelihood or posterior for ν. This is a heuristic, presented without derivation.\n\nThe empirical support is thin. The skill-score gain of Bayes-ν over Bayes is 0.18, 0.07, and 0.27 percentage points across the three years. No standard errors, no tests, no intervals. And Table 2 shows Bayes-ν placed 6th or 7th forty times versus five for Bayes, so the adaptive version is not uniformly better; it wins often but also loses more often. For a claimed algorithmic contribution, that is not convincing.\n\nTo be fair, the fixed-ν Bayesian method and the comparison framework are contributions in themselves. The paper is honest about the limitations of adaptive methods and the effect of data quality. The writing is clear. I do not see circularity in the central comparison: the benchmarks are external and the holdout years are genuine. The reconstructed-data objective is an internal fit, not a comparison trick.\n\nMy recommendation: this deserves peer review. The evaluation methodology and the fixed-ν Bayesian method are valuable enough that a competent referee should spend time on it. But the ν-update needs to be re-derived or clearly framed as a pragmatic heuristic with diagnostics, and the claims about Bayes-ν's superiority need uncertainty quantification. Without those, the adaptive method should be treated as exploratory rather than a proven improvement.","headline":"The fixed-ν Bayesian method and the large-scale evaluation are solid and worth publishing, but the adaptive ν-update is under-specified and its claimed gains are small, unquantified, and uneven.","tokens_in":19096,"tokens_out":1624,"would_cite":true,"duration_ms":18741,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper's central claim is that Bayes-ν, an adaptive Bayesian AR method with online updating of the generalized logit shape parameter, outperforms six benchmarks for very short-term wind power forecasting.","keywords":["probabilistic forecasting","wind power","generalised logit transformation","adaptive Bayesian estimation","very short-term forecasting","continuous ranked probability score","functional reliability diagrams","online parameter updating"],"falsifier":"Take a wind farm where Bayes-ν wins and re-run it with the reconstruction term L*(ν) removed from the ν objective, so ν is updated only from new observations: if the full method no longer beats fixed-ν Bayes, the reconstruction is the deciding mechanism. Alternatively, simulate a double-bounded series with a known time-varying ν and check whether the updated ν̂ tracks the true ν; if it does not, the ν-update is not recovering the intended quantity.","tokens_in":17897,"feed_emoji":"🌬️","tokens_out":8531,"duration_ms":82976,"temperature":0.7,"pith_summary":"The paper sets out to show that very short-term probabilistic wind power forecasts improve when the transformation used to unbounded the double-bounded data is itself adapted online. Its method, Bayes-ν, pairs the generalized logit transformation with a conjugate Bayesian AR model and re-estimates the transformation's shape parameter at each update from pseudo-observations recovered from the posterior. In a four-year, 101-farm British case study it achieves the lowest average CRPS and the highest skill score relative to persistence in every test year, and its calibration is more consistent between farms than recursive least squares or Newton-Raphson alternatives. A sympathetic reader would take the contribution to be a practical, stable way to keep probabilistic forecasts well calibrated as wind farm behaviour drifts.","feed_headline":"Bayes-ν beats seven wind forecast rivals on 101 farms","feed_subtitle":"Retuning the generalized logit shape parameter each step lowers CRPS and improves calibration in very short-term forecasts.","key_machinery":"The generalized logit transformation $L_\\nu(x)=\\ln\\left(x^\\nu/(1-x^\\nu)\\right)$ maps double-bounded wind power data $x\\in(0,1)$ to the real line, with the shape parameter $\\nu$ controlling skewness near zero and rated power; $\\nu=1$ recovers the ordinary logit. The forecasting model is an AR($p$) process in the transformed domain with a Gaussian prior on the coefficients and a Gamma prior on the error precision, giving conjugate Bayesian updates with forgetting factors. The load-bearing trick for adaptive $\\nu$ is to reconstruct a small set of pseudo-observations from the posterior: after updating the precision matrix, the paper applies a Cholesky decomposition $P^* = k^2 LL^\\top$, forms $y^* = L^\\top \\mu^*$, transforms these back to the original domain, and then chooses $\\hat\\nu$ by minimizing $L^*(\\nu)+\\tilde L(\\nu)$, a sum of negative log-likelihoods from the reconstructed data and from the newly received data. The updated value is smoothed as $\\nu^*_{t+M}=(1-\\gamma)\\nu_{t+M}+\\gamma\\hat\\nu_{t+M}$, which lets the transformation track non-stationarity without discarding its previous state.","core_discovery":"On its own terms, the paper's central discovery is that the shape parameter ν of the generalised logit transformation should be treated as a time-varying quantity and updated online, not fixed across the forecast horizon. The method alternates between conjugate Bayesian updates of the AR coefficients and error precision and a point estimate of ν that minimizes a combined negative log-likelihood of the new observations and of pseudo-observations reconstructed from the posterior precision via a Cholesky decomposition. The empirical core is Table 1: Bayes-ν has the lowest average CRPS (3.651%, 3.975%, 3.644%) and the highest skill score (3.925%, 3.504%, 4.604%) across the 2021–2023 test datasets, and it ranks first 129 times across all farm-year scenarios, whereas NR ranks first 91 times but falls to ranks 6 or 7 on 128 occasions. The authors also show, through functional reliability diagrams, that all methods tend to produce over-dispersed predictive densities, with the Bayesian methods having narrower 50% envelopes of calibration curves across farms.","pith_inferences":["An unstated consequence is that the Cholesky-based pseudo-observation reconstruction is not wind-specific, so the same update rule could be applied to any double-bounded series with boundary inflation, such as solar power or battery state-of-charge data.","A testable extension is to compute the rolling ν trajectory for each farm and align it with known operational events; the paper observes surges and declines in ν without tying them to causes, making the trajectory a candidate real-time anomaly signal.","A sharper comparison would treat ν fully Bayesian rather than through the point-wise heuristic; if the reconstructed pseudo-observations are biased, that comparison should show the advantage of Bayes-ν over fixed-ν Bayes shrinking.","A practical diagnostic suggested by the paper's own sensitivity discussion is to estimate ν on a rolling window offline and switch on the adaptive update only when the rolling estimate moves beyond a threshold, focusing the mechanism on genuinely changing regimes."],"forward_implications":["An operator using Bayes-ν can expect lower average CRPS and a higher skill score than persistence, AR, RLS, NR, and fixed-ν Bayes in comparable very short-term settings, based on 101 UK farms over three test years.","The added value of adaptive ν is conditional on ν actually drifting; the paper notes that when the shape parameter is stable, most of the improvement comes from the Bayesian update itself, and the adaptive update adds less.","Because RLS and NR occasionally collapse to near-persistence skill levels while the Bayesian methods stay consistently ranked high, Bayes-ν is the safer choice when a single bad forecast is costly.","Functional reliability diagrams give a farm-level view of calibration, and the paper uses them to show that all methods over-disperse, with Bayesian methods showing less farm-to-farm variation in calibration bias.","The method's robustness under parameter disturbances comes from its stepwise interactive update, where the same precision matrix drives both parameter updates and the reconstruction that retunes ν."],"supporting_citations":[{"why":"Introduces the generalised logit-normal distribution and thresholding for wind power data, and provides both the skewness argument and the RLS baseline.","marker":"Pinson (2012)"},{"why":"Supplies the Newton-Raphson adaptive update for ν that the proposed method extends and uses as the NR benchmark.","marker":"Pierrot and Pinson (2021)"},{"why":"Defines the logit-normal distribution whose density underpins the likelihood in Eq. (2.3).","marker":"Aitchison and Shen (1980)"},{"why":"Provides evidence that conditional wind power distributions are skewed near zero and rated power, motivating the shape parameter.","marker":"Dowell and Pinson (2015)"},{"why":"Establishes CRPS as a strictly proper scoring rule, the paper's primary evaluation metric.","marker":"Gneiting and Raftery (2007)"},{"why":"Provides functional boxplots used to construct the functional reliability diagrams that assess calibration across farms.","marker":"Sun and Genton (2011)"},{"why":"Documents why persistence is a strong benchmark for very short-term forecasts, defining the reference for the skill score.","marker":"Tawn and Browell (2022)"}],"fun_headline_variants":["Adaptive logit shape sharpens Bayesian wind forecasts","Updating ν online cuts CRPS on 101 UK wind farms","Bayes-ν: online shape update wins wind forecast contest","Time-varying logit transform boosts very short-term wind power","Retune ν each step for better probabilistic wind forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's extra step assumes that small artificial data samples reconstructed from the updated model summary stand in for the real wind power data when re-estimating the shape parameter; if those samples are not representative, the claimed edge over fixed-ν Bayes shrinks or disappears.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive logit shape sharpens Bayesian wind forecasts","Updating ν online cuts CRPS on 101 UK wind farms","Bayes-ν: online shape update wins wind forecast contest","Time-varying logit transform boosts very short-term wind power","Retune ν each step for better probabilistic wind forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000622,"raw_usage":{"total_tokens":2899,"prompt_tokens":978,"completion_tokens":1921,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":1839}},"tokens_in":594,"tokens_out":1921,"duration_ms":13801,"temperature":1.0,"reasoning_tokens":1839,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:11:12.581035+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a wind farm where Bayes-ν wins and re-run it with the reconstruction term L*(ν) removed from the ν objective, so ν is updated only from new observations: if the full method no longer beats fixed-ν Bayes, the reconstruction is the deciding mechanism. Alternatively, simulate a double-bounded series with a known time-varying ν and check whether the updated ν̂ tracks the true ν; if it does not, the ν-update is not recovering the intended quantity.","supporting_citations":[{"cited_title":"and Pinson, P","cited_arxiv_id":null,"evidence_quote":"Supplies the Newton-Raphson adaptive update for ν that the proposed method extends and uses as the NR benchmark."},{"cited_title":"and Shen, S","cited_arxiv_id":null,"evidence_quote":"Defines the logit-normal distribution whose density underpins the likelihood in Eq. (2.3)."},{"cited_title":"and Pinson, P","cited_arxiv_id":null,"evidence_quote":"Provides evidence that conditional wind power distributions are skewed near zero and rated power, motivating the shape parameter."},{"cited_title":"and Genton, M","cited_arxiv_id":null,"evidence_quote":"Provides functional boxplots used to construct the functional reliability diagrams that assess calibration across farms."},{"cited_title":"and Browell, J","cited_arxiv_id":null,"evidence_quote":"Documents why persistence is a strong benchmark for very short-term forecasts, defining the reference for the skill score."}],"review_version":1}