{"id":"ad33a44f-3257-4351-b000-e864ed57c8a1","arxiv_id":"1908.06836","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A fruit-fly-optimization-tuned Holt-Winters model reports lower MAPE than a default-parameter Holt-Winters, a seasonal index model, and a GA-SVR model on two short electricity series, but the comparison omits standard optimized Holt-Winters baselines.","lead":"The authors pair a fruit fly swarm search with Holt-Winters exponential smoothing to choose three smoothing parameters automatically. On two Chinese city electricity series, the hybrid reports lower average forecast errors than three baseline methods, including when only a few years of training data are available.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Best-performance claim rests on a non-standard Holt-Winters baseline: MHW with fixed α=0.2, β=0.1, γ=0.6, whereas standard practice optimizes these parameters; without that comparison the FOA advantage is unestablished.","rationale":"The paper's central claim is that FOA-tuned Holt-Winters parameters yield the best monthly electricity consumption forecasts, including in small-sample settings. For that claim to be meaningful, the comparison baseline must represent standard practice. The chosen MHW baseline uses arbitrarily fixed smoothing parameters, even though the paper's own cited textbook describes estimating these parameters from data. This is load-bearing because the proposed model's contribution is precisely the parameter-selection mechanism; if standard optimized Holt-Winters performs similarly or better, the FOA step contributes nothing beyond conventional estimation. The test proposed above would settle this directly by comparing FOA-MHW against an optimized Holt-Winters model under identical data splits. The paper also has other support problems, including no code, no data, small test samples of six points, no error bars, and an apparent internal contradiction about SI being best below five years of training data. These reinforce rejection, but the missing fair baseline is the most specific and decisive scientific gap. The reader's verdict of REJECT remains appropriate; my stress-test does not change it.","tokens_in":11768,"tokens_out":7339,"duration_ms":69853,"concrete_test":"Re-run the two case studies with the same 42/6 training/validation split and the same 2018 test set, replacing MHW-default by a standard multiplicative Holt-Winters model whose α, β, and γ are estimated on the optimizational training set (for example, by grid search or an optimizer minimizing one-step-ahead RMSE, or equivalently R's ets with a multiplicative-seasonal, additive-trend specification). Compare test MAPE on both series. If the standard optimized HW attains MAPE close to or below FOA-MHW's 3.65% and 1.89%, the claimed FOA advantage is not demonstrated; if FOA-MHW still wins clearly on both series, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's evidence for FOA-MHW superiority is entirely relative to three baselines, and the key Holt-Winters baseline is MHW with fixed default smoothing parameters (α=0.2, β=0.1, γ=0.6; Section 3.2 step 3). Standard exponential smoothing estimates the smoothing parameters from the data, for example by minimizing one-step-ahead squared error or likelihood, as covered in the paper's own reference [28]. Comparing FOA-tuned parameters against hand-fixed defaults conflates the tuning procedure with the FOA optimizer; the reported MAPE gains over MHW-default (3.65% vs 4.52% in Table 1, and 1.89% vs 3.18% in Table 2) may simply reflect the value of any sensible optimization. The central claim that FOA-MHW 'always achieves the best performance' is therefore not established unless FOA-MHW also beats a fairly optimized Holt-Winters baseline. A secondary internal tension exists, since Section 3.3 states that the SI model has the smallest MAPE for training data below five years, which conflicts with the conclusion's 'always' wording, but the primary methodological gap is the missing standard HW benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid forecasting model, FOA-MHW, which uses the fruit fly optimization algorithm (FOA) to select the three smoothing parameters (α, β, γ) of the multiplicative Holt-Winters (MHW) exponential smoothing method. The model is applied to monthly electricity consumption data from a Chinese city and from a telecommunications/television industry subsector, using a bimonthly reporting frequency that yields six data points per year and 54 points total (2010–2018). The data are split so that the first 42 points are used for optimization and the last 6 points of the training period (2017) serve as a validation set; the test set is the full year 2018. The authors compare FOA-MHW against a seasonal index (SI) model, an MHW model with fixed default parameters (α=0.2, β=0.1, γ=0.6), and a genetic-algorithm-tuned support vector regression (GASVR) model. They report MAPE values of 3.65% and 1.89% for FOA-MHW on the two series versus 4.38%/9.05% (SI), 4.52%/3.18% (MHW-default), and 7.00%/2.71% (GASVR). They also vary the training length down to 3 years and claim that FOA-MHW remains accurate, with a MAPE of 3.58% in the 3-year case. The paper concludes that FOA-MHW 'always achieve the best performance' over the selected benchmarks.","tokens_in":11934,"tokens_out":4374,"duration_ms":39216,"significance":"If the central claim were established, the paper would offer a practical, low-data forecasting tool: automatically tuned Holt-Winters smoothing with MAPE around 2–4% on bimonthly electricity series, which would be useful for utilities with limited history. The method is clearly described and the data are real, so the work is reproducible in principle. However, the significance is substantially undercut by two evidentiary weaknesses. First, the primary Holt-Winters baseline uses hand-fixed default parameters rather than a conventionally optimized Holt-Winters model, so the reported gains may reflect the value of parameter tuning rather than any special property of FOA. Second, the evaluation rests on only six test points per series, with no uncertainty quantification or significance testing, and the paper's own small-sample experiment (Figure 12) shows the SI model outperforming FOA-MHW below five years of training data, contradicting the unqualified 'always' claim. The paper thus provides a promising algorithmic idea and a plausible result, but the evidence as presented is not sufficient for the strength of the conclusions.","major_comments":[{"comment":"The primary Holt-Winters baseline, MHW with default parameters (α=0.2, β=0.1, γ=0.6), is not a standard Holt-Winters implementation. Standard practice estimates the smoothing parameters from the data, for example by minimizing one-step-ahead squared error or likelihood, as covered by the paper's own reference [28]. The reported MAPE differences between FOA-MHW and MHW-default (3.65% vs 4.52% in Table 1; 1.89% vs 3.18% in Table 2) conflate the FOA optimizer with the act of tuning parameters. To support the claim that FOA provides an advantage over standard Holt-Winters, the authors must add a benchmark that is optimized by a conventional method (e.g., grid search or likelihood maximization) on the same training/validation split, and show that FOA-MHW is at least as accurate as that benchmark.","section":"Section 3.2, step 3 and Tables 1–2"},{"comment":"The evaluation uses only six test points per series (the year 2018, with bimonthly data), and no confidence intervals, significance tests, or repeated subsampling are reported. The observed MAPE differences (e.g., 3.65% vs 4.52% in Table 1) may not be statistically meaningful, and the conclusion that FOA-MHW 'always achieve the best performance' is internally contradicted by Section 3.3, where the SI model is reported to have the smallest MAPE when the training data span fewer than five years (Figure 12). Please add uncertainty quantification and rephrase the conclusion to reflect the conditions under which FOA-MHW actually wins.","section":"Section 3.1, Tables 1–2, and Section 3.3"},{"comment":"The FOA is a stochastic algorithm, but the paper reports only a single run with no seed information, no repeated trials, and no sensitivity analysis for the hyperparameters (sizepop=50, maxgen=20, FR=[5,10]). The reported optimal smoothing parameters (e.g., α=0.5562, β=0.2022, γ=0.3590) and the resulting MAPE values could vary substantially across runs, so the robustness of the claimed advantage is not established. Please report means and standard deviations over multiple independent runs or at least run a sensitivity analysis.","section":"Section 2.3, step 4, and Section 3.2"},{"comment":"There is an inconsistency in the definition of the smell concentration judgment value. In the standard FOA description (equations (11)–(12)), the distance is Dist_i = sqrt(X_i^2 + Y_i^2) and S_i = 1/Dist_i. However, in step 4 of the FOA-MHW procedure, equations (24)–(26) define D_i = 1/sqrt(X_i^2 + Y_i^2), which would make S_i = 1/D_i = sqrt(X_i^2 + Y_i^2), effectively inverting the intended relation and causing the algorithm to prefer locations far from the origin. Since the paper reports convergence, this is likely a typographical error, but as written the algorithm description is not reproducible. Please correct the equations.","section":"Section 2.3, step 4, equations (21)–(26)"}],"minor_comments":[{"comment":"The abstract contains the typo 'the the' in the first line and several grammatical issues; please proofread throughout.","section":"Abstract"},{"comment":"In Table 1, 'July. 2018' should be 'Jul. 2018'; in Figure 12, the x-axis is drawn with the number of training years decreasing from 8 to 3, which is unconventional; consider reversing the axis order for readability.","section":"Section 3.2 and Figure 12"},{"comment":"The introduction states 'this is the first enhanced HW model', but there is a substantial literature on optimizing exponential smoothing parameters with metaheuristics; please soften or verify this claim with appropriate references.","section":"Introduction"},{"comment":"The term 'optimizational training data set' is unusual; consider naming it 'fitting training set' or 'estimation set' to avoid confusion with the validation set.","section":"Section 2.3 and Section 3.1"},{"comment":"Reference [28] is cited for the default parameter values, but the book 'Forecasting: Methods and Applications' does not prescribe a universal default (0.2, 0.1, 0.6); please provide a precise source or state that these are arbitrary settings chosen for the baseline.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely to require a substantially expanded empirical evaluation to satisfy the standards of this journal. The missing optimized-Holt-Winters baseline and the absence of any uncertainty quantification are the two load-bearing gaps. The novelty of applying FOA to Holt-Winters parameter selection is real but modest; the paper's value would hinge on a convincing demonstration that FOA achieves comparable or better accuracy than standard parameter estimation, especially in small-sample situations. I would also ask the authors to reconcile the 'always best' conclusion with the SI model's superior performance in the short-training regime shown in Figure 12. If the authors can add these experiments and temper the claims, the paper could be publishable; otherwise the evidence is too thin."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The practical take: the central claim that FOA-MHW always beats the benchmarks is not supported, because the key Holt-Winters baseline uses fixed arbitrary parameters (α=0.2, β=0.1, γ=0.6) instead of standard optimized parameters. The paper never compares against a properly fitted Holt-Winters model, so the reported gains likely just reflect the value of tuning, not the specific optimizer. The stress-test note is right, and the paper's own Section 3.3 undermines the 'always' wording by showing the SI model does better below five years of training data.\n\nWhat is actually new: applying fruit fly optimization to select Holt-Winters smoothing parameters, with two real-data case studies and a small-sample analysis. That combination may be new, and the idea is reasonable. The paper is clearly structured, and the authors are honest that the data is proprietary and that the SI model beats them in one regime, even if the conclusion overstates it.\n\nSoft spots beyond the baseline: the FOA equations in Section 2.3 are internally inconsistent. Equations (24)-(26) define D_i = 1/sqrt(...), then S_i = 1/D_i, which gives S_i = sqrt(...), not a value in [0,1] as required for smoothing parameters. This makes the algorithm as written impossible to reproduce. There are only six test points per series, no significance tests or intervals, and no code or data. These are real problems, but the tuning idea is salvageable.\n\nThe paper is not a waste of desk space. It tackles a practical problem and the empirical setup, apart from the weak baseline, is reasonable. But as it stands, the evidence does not support the headline claims, and the algorithm description needs fixing.\n\nFor peer review: yes, send it out, but expect heavy revision. A referee should demand a comparison against standard optimized Holt-Winters (e.g., via maximum likelihood or grid search), rolling-origin evaluation, corrected equations, and ideally shared data or code. I would not cite it in current form, but it is worth a serious look as an applied contribution.\n\nThe reader's REJECT verdict is fair in substance; the paper is not publishable as-is. But the idea is not bad, and with the right comparisons it could become a legitimate applied paper.","headline":"FOA-tuned Holt-Winters is a plausible idea, but the paper's evidence rests on a straw-man baseline and an internally inconsistent algorithm description.","tokens_in":12564,"tokens_out":3214,"would_cite":false,"duration_ms":31775,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Auto-selecting Holt-Winters smoothing parameters with a fruit fly search keeps monthly electricity forecasts accurate even when only three years of training data are available.","keywords":["monthly electricity consumption forecasting","Holt-Winters exponential smoothing","multiplicative Holt-Winters model","fruit fly optimization algorithm","small-sample forecasting","parameter optimization","seasonal time series","mean absolute percentage error"],"falsifier":"On the same 54 bimonthly points, fit multiplicative Holt-Winters by minimizing one-step-ahead error on the training portion and evaluate on the same six test months; if that standard optimized version matches or beats the reported 3.65% and 1.89% MAPEs, the fruit fly search is not what carries the accuracy.","tokens_in":11475,"feed_emoji":"⚡","tokens_out":7196,"duration_ms":64380,"temperature":0.7,"pith_summary":"This paper proposes a hybrid forecasting model, FOA-MHW, that uses the fruit fly optimization algorithm to choose the three smoothing parameters ($\\alpha$, $\\beta$, $\\gamma$) of multiplicative Holt-Winters exponential smoothing. On 54 bimonthly electricity-consumption points from a city in China, with 48 used for training and 6 for testing, the tuned model reports mean absolute percentage errors of 3.65% for total city consumption and 1.89% for a single industry, both better than the seasonal-index model, the default-parameter Holt-Winters model, and the GA-SVR model. The central claim is that automatic parameter selection keeps Holt-Winters accurate even when only three years of training data are available, with MAPE 3.58%. If correct, this gives utilities a simple, low-data forecasting option that does not require the large training sets typical of machine-learning methods.","feed_headline":"Fruit-fly tuning keeps monthly electricity forecasts near 2-4% error","feed_subtitle":"Auto-selected Holt-Winters parameters beat seasonal-index and machine-learning baselines on bimonthly utility data.","key_machinery":"The load-bearing mechanism is the FOA-MHW estimator, in which the fruit fly optimization algorithm (a swarm heuristic that scores candidate positions by a smell-concentration value) searches the three smoothing constants of the multiplicative Holt-Winters recursion, using the root mean square error on a validation segment as the fitness. The algorithm maps each candidate triple ($\\alpha$, $\\beta$, $\\gamma$) to a fruit fly location, then iteratively moves the swarm toward the location with lowest validation RMSE. This tuning step, not a change in the forecasting equation, is what the paper credits for the improved accuracy.","core_discovery":"The paper's central claim is that the accuracy of multiplicative Holt-Winters for bimonthly electricity consumption is controlled mainly by its smoothing parameters, and that fruit fly optimization can find those parameters from a validation segment. In the two reported experiments, the tuned FOA-MHW model posts the lowest mean absolute percentage error among all four models, and the error stays nearly flat as training data shrink from eight years to three, reaching 3.58% at three years. The selected parameters in the city case ($\\alpha=0.5562$, $\\beta=0.2022$, $\\gamma=0.3590$) are treated as the reason the model absorbs irregular seasonality better than the fixed-default baseline.","pith_inferences":["The paper leaves untested the standard practice of estimating Holt-Winters parameters directly from the training data by error minimization; a head-to-head test against that optimized Holt-Winters baseline would show whether the fruit fly search itself is needed.","Because only three parameters are searched, a simple grid search or coordinate descent over ($\\alpha$, $\\beta$, $\\gamma$) might reproduce the same MAPE, which would place the value in parameter tuning rather than in the swarm heuristic.","The same tuning scheme could plausibly extend to additive or damped Holt-Winters and to monthly series with 12-point seasons, but the evidence here only covers $L=6$ bimonthly data."],"forward_implications":["When smoothing parameters are chosen by FOA, the same Holt-Winters recursion reports MAPE of 3.65% on city consumption and 1.89% on the industry series, outperforming SI, MHW-default, and GASVR.","Reducing the training window from eight years to three leaves FOA-MHW near 3.58% MAPE, while the default-parameter Holt-Winters and GASVR errors rise as data shrink.","The method inherits the low data requirement of Holt-Winters and adds only a bounded optimization search, so it can be deployed where machine-learning baselines lack sufficient training samples.","The reported results apply to bimonthly series with six observations per year, where the multiplicative seasonal form fits the data."],"supporting_citations":[{"why":"Supplies the Holt-Winters formulation, initialization, and the default smoothing parameters used as the MHW-default baseline.","marker":"[28]"},{"why":"Introduces the fruit fly optimization algorithm that performs the parameter search.","marker":"[27]"},{"why":"Provides the Holt-Winters seasonal forecasting method that the model builds on.","marker":"[25]"},{"why":"Provides the Winters seasonal exponential smoothing formulation used by the method.","marker":"[26]"},{"why":"Supplies the FOA implementation pattern for seasonal electricity consumption forecasting.","marker":"[29]"},{"why":"Defines the seasonal-index benchmark model compared against FOA-MHW.","marker":"[30]"},{"why":"Defines the GA-SVR machine-learning benchmark model compared against FOA-MHW.","marker":"[31]"}],"fun_headline_variants":["Fruit-fly-tuned Holt-Winters improves small-sample power forecasts","Fruit fly optimization sharpens Holt-Winters for tight electricity data","Bimonthly power demand forecast improved by fruit-fly parameter pick","Small-data electricity forecasting gets a boost from fruit fly tuning","Fruit-fly-selected Holt-Winters parameters cut monthly forecast error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The advantage is measured against a Holt-Winters model with fixed default smoothing parameters, not against the standard practice of estimating those parameters from the data, so the comparison may understate how good plain Holt-Winters already is.","fun_headline_variants_meta":{"raw":{"variants":["Fruit-fly-tuned Holt-Winters improves small-sample power forecasts","Fruit fly optimization sharpens Holt-Winters for tight electricity data","Bimonthly power demand forecast improved by fruit-fly parameter pick","Small-data electricity forecasting gets a boost from fruit fly tuning","Fruit-fly-selected Holt-Winters parameters cut monthly forecast error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000588,"raw_usage":{"total_tokens":2740,"prompt_tokens":907,"completion_tokens":1833,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":1741}},"tokens_in":523,"tokens_out":1833,"duration_ms":13879,"temperature":1.0,"reasoning_tokens":1741,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:29:36.880381+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the same 54 bimonthly points, fit multiplicative Holt-Winters by minimizing one-step-ahead error on the training portion and evaluate on the same six test months; if that standard optimized version matches or beats the reported 3.65% and 1.89% MAPEs, the fruit fly search is not what carries the accuracy.","supporting_citations":[{"cited_title":"Makridakis, S","cited_arxiv_id":null,"evidence_quote":"Supplies the Holt-Winters formulation, initialization, and the default smoothing parameters used as the MHW-default baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the fruit fly optimization algorithm that performs the parameter search."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Holt-Winters seasonal forecasting method that the model builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Winters seasonal exponential smoothing formulation used by the method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FOA implementation pattern for seasonal electricity consumption forecasting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the seasonal-index benchmark model compared against FOA-MHW."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the GA-SVR machine-learning benchmark model compared against FOA-MHW."}],"review_version":1}