{"id":"a0d4127e-a3a3-4924-9fe4-71f1dbb6b5a2","arxiv_id":"1908.03493","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"GMSD(1,1), a discretized Simpson-rule grey model, exactly simulates homogeneous exponential sequences and gives slightly lower prediction error than GM(1,1), DGM(1,1), and GMSC(1,1) on three real datasets.","lead":"This paper constructs a discrete grey forecasting model, GMSD(1,1), that uses Simpson's rule for the background value and derives the exact recurrence its forecasts follow. It proves the model reproduces purely exponential sequences without error and reports small accuracy gains over three existing grey models on Chinese economic data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical 'outperforming' claim rests on tiny, statistically untested MAPE differences; theory only covers exact exponentials.","rationale":"The reader's verdict of CONDITIONAL is well aligned with my analysis: the mathematical derivation and unbiasedness proof for homogeneous exponential sequences are internally consistent and correct, but the empirical evidence for the central practical claim is thin and not statistically substantiated. My stress-test focuses more strongly than the reader did on the fragility of the 'outperforming' assertion: the reported differences are minuscule (0.01–0.25 percentage points), the test sets are extremely small, GMSD is actually the worst in in-sample fit across all three applications, and no uncertainty quantification is provided. The theoretical result about unbiasedness, while correct, does not discriminate GMSD from DGM(1,1), which is also unbiased for exponentials; hence the practical advantage rests entirely on the weak empirical numbers. This is exactly the kind of concern that a CONDITIONAL verdict should carry: the paper can be accepted in principle, but the empirical claim needs code/data and more robust validation. I do not see an internal inconsistency or a mathematical error that would warrant REJECT. The absence of code/data is a reproducibility issue, addressable by a major revision, which is what CONDITIONAL means. Thus I keep the reader's verdict unchanged. My agreement is 'partial' because the reader's weakest_assumption emphasizes the model specification and the restricted scope of unbiasedness, while I identify the lack of statistical significance in the empirical comparison as the single most load-bearing issue for the paper's bottom-line claim. I still regard the theoretical contribution as sound and worth publishing, so the overall verdict remains CONDITIONAL.","tokens_in":14702,"tokens_out":34762,"duration_ms":290887,"concrete_test":"Obtain the authors' code or raw data for the three applications and recompute the out-of-sample MAPE for GMSD and GMSC using exactly the same initialization and fitting protocol. Then compute bootstrap 95% confidence intervals for the MAPE differences (or a paired test on the absolute errors). If the confidence intervals include zero, as expected for differences of 0.01–0.25 percentage points on 3–6 test points, the 'outperforming' claim should be weakened to 'comparable performance' in the manuscript.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central practical claim is that GMSD(1,1) outperforms GM(1,1), DGM(1,1), and GMSC(1,1) in real applications (Section 6). This claim is supported only by point estimates of MAPE on very small test sets, with no confidence intervals, significance tests, or uncertainty quantification. In Section 5.1 (GDP), the improvement over GMSC is 7.6118% vs 7.6220% MAPE_pred — a difference of 0.01 percentage point. In Section 5.2 (freightage), it is 6.3579% vs 6.3810% — 0.02 points. In Section 4.2 (electricity), the improvement is 3.2669% vs 3.5137% — 0.25 points, but the test set has only 3 observations. These differences are within the noise of a handful of data points and could easily flip with one small perturbation. Moreover, GMSD has the worst in-sample MAPE_simu in all three applications (e.g., 1.7387% vs 1.5675% for GM(1,1) in the electricity example), suggesting its advantage is not from better fit but from an artifact of the short test windows. The theoretical unbiasedness result (Section 3.4) applies only to exact homogeneous exponential sequences, and DGM(1,1) also possesses this property; thus it does not explain why GMSD should be preferred on non-exponential real data. The absence of code and data makes it impossible to verify whether the reported numbers are reproducible or whether the baseline implementations (e.g., forcing the first two fitted values to match the data, as suggested by Table 2 and Table 4) are fair. Without statistical support, the headline claim of 'outperforming' is not established beyond a heuristic preference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a discrete grey model, GMSD(1,1), whose background value is constructed with the Simpson numerical integration formula, and whose time response function is the exact solution of the resulting second-order difference equation. The parameters are estimated by least squares, the discrete model is contrasted with the continuous GMSC(1,1) model, and a proof is given that GMSD(1,1) is unbiased for homogeneous exponential sequences. Numerical examples compare GMSD(1,1) with GM(1,1), DGM(1,1), and GMSC(1,1) on a synthetic exponential sequence and on three real datasets (China's electricity consumption, Lanzhou's GDP, and Lanzhou's freightage), with the empirical sections claiming that GMSD(1,1) provides the most accurate predictions.","tokens_in":14965,"tokens_out":7764,"duration_ms":79569,"significance":"The theoretical core of the paper is sound and useful: the discrete time response is internally consistent with the fitted difference equation, the unbiasedness theorem in Section 3.4 is a genuine algebraic result, and Section 3.3 gives a clean diagnostic, the function phi(a), showing why the continuous GMSC(1,1) model is inconsistent when |a| is not small. These contributions are modest but real within the grey-systems literature. The main weakness is that the paper's headline practical claim, that GMSD(1,1) 'outperforms' the other three models, rests on very small point estimates of MAPE on short test sets without any uncertainty quantification. The absence of code and data also makes the numerical comparisons difficult to verify. If the empirical claims are strengthened or suitably weakened, the paper would be a solid contribution to specialized journals in this area.","major_comments":[{"comment":"The central empirical claim that GMSD(1,1) outperforms GM(1,1), DGM(1,1), and GMSC(1,1) is not supported by the evidence as presented. The reported prediction advantages are extremely small: in Table 5, MAPE_pred is 7.6118% versus 7.6220% for GMSC; in Table 7, it is 6.3579% versus 6.3810%; and in Table 3, it is 3.2669% versus 3.5137% on only three test observations. No confidence intervals, significance tests, or sensitivity analyses are given, and the in-sample MAPE_simu of GMSD(1,1) is the worst among all four models in every application (e.g., Table 3: 1.7387% versus 1.5675% for GM(1,1)). These differences are within the noise of a handful of data points. The authors should either provide proper uncertainty quantification and a test for equality of forecast errors, or substantially weaken the concluding statement to say that GMSD(1,1) performs comparably on these particular datasets rather than claiming general superiority.","section":"Section 5; Tables 4-7"},{"comment":"The unbiasedness proof rests on the closed-form least-squares solution stated in Eq. (22), but the derivation is presented with the phrase 'After some calculations' and the algebra is not shown. This is the load-bearing step that identifies the estimated parameters with the values 3(1-q^2)/(1+4q+q^2) and 3rq(1+q)/(1+4q+q^2). The authors should provide the full derivation, including the computation of B^T B and its inverse, and state the conditions under which the inverse exists (for instance, the number of fitting points n and the range of q). Without this, a central result of the paper is not verifiable from the manuscript text.","section":"Section 3.4; Eq. (22)"},{"comment":"The numerical implementation of the benchmark models is not transparent. Equation (18) defines the GMSC time response using x(0)(1) as the initial value, and in general this does not reproduce x(0)(2) from the raw data; nevertheless, the GMSC columns in Tables 3, 4, and 6 show exact agreement with the actual value at k=2 (e.g., 28588.0000, 567.0400, and 5973.0000). At the same time, the GMSD response in Eq. (14) uses x(1)(1) and x(1)(2) as initial conditions, so by construction it interpolates the first two original data points. The authors should specify the exact initial-condition convention and the fitted time index set for each of the four models, and state explicitly whether any fitted values are set equal to the observed data by construction. Without this information, the reported advantages could be artifacts of different initialization schemes rather than of the Simpson-based background value.","section":"Section 3.3; Eq. (18); Tables 3, 4, 6"}],"minor_comments":[{"comment":"The text says 'parameters a and b are the provided determined of Eq.(22)'; this should be reworded as 'the provided values determined by Eq. (22)'. Also, Figure 2 depends on r, which is described only as randomly generated in [1,15]; please report a seed or the exact generated values so that the figure is reproducible.","section":"Section 4.1; Eq. (26)"},{"comment":"The Introduction states that the paper 'study the forecast stability problem' of the discrete GMSD(1,1) model, but no dedicated stability analysis appears in Section 3 or elsewhere. Section 3.3 discusses parameter dependence and inconsistency of GMSC(1,1), which is related but not identical to forecast stability; the text should be aligned with what is actually studied.","section":"Introduction, Section 3"},{"comment":"The derivation of Eq. (14) is presented for k = 2, 3, ..., n-1, whereas Eq. (15) is stated for k = 1, 2, ..., n-1. Please clarify how the boundary cases k=1 and k=n are defined, since the sums and powers in the formulas involve k-2 and could otherwise be ambiguous.","section":"Equations (14)-(15)"},{"comment":"There are several typos and formatting inconsistencies, for example 'thw development' in Section 1 and the inconsistent spacing of 'GMSC' and 'GMSD' (e.g., 'GM SC(1,1)' versus 'GMSD(1,1)'). A careful proofreading pass is needed.","section":"Throughout"},{"comment":"No code or data availability statement is provided. Given that the paper makes quantitative comparisons on real datasets, the raw data and the implementation of the four models should be made available, or at least the data should be provided in an appendix, so that the reported MAPE values can be checked.","section":"Numerical reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely within the scope of the journal and the theoretical construction is sound. The main risk is that the empirical 'outperforming' claim is overstated and supported only by tiny, statistically untested MAPE differences; the comparison also appears to use different initialization conventions across models. I would support publication after the authors either add uncertainty quantification and clarify the benchmark implementations, or restrict the conclusions to a more modest claim. The missing derivation of Eq. (22) should also be supplied, as it is central to the unbiasedness proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a small, internally coherent paper that does something real. It takes the continuous GM(1,1) with Simpson background value (GMSC) and constructs a genuinely discrete version (GMSD) whose recurrence is the exact solution of the discretized whitening equation. That removes the inconsistency the continuous version has for larger |a|, and they prove the model reproduces homogeneous exponential sequences exactly. The consistency analysis in Section 3.3 is the best part. The unbiasedness calculation checks out; I tried it with a small symbolic example and the formula collapses to r q^{k+1} as claimed.\n\nThe empirical section is the weak part. On the electricity, GDP, and freightage data, GMSD's advantage over GMSC is 0.01–0.25 percentage points of MAPE, on test sets of 3 to 6 points, with no confidence intervals or significance tests. These differences are noise. Worse, GMSD has the worst in-sample MAPE in all three applications, which suggests the small prediction advantage is an artifact of short test windows rather than a modeling edge. The unbiasedness theorem doesn't help here because it only applies to exact exponential sequences, and DGM(1,1) already has that property for the standard background value. There is no code or data, so I could not verify the tables.\n\nI would not describe this as a major advance. It is a useful incremental improvement inside a niche subfield. The theory is honest and the parameter derivation is legible. Section 6's note about extensions to GM(1,n) and GMC(1,n) is appropriate. The self-citations are for background and not a problem.\n\nWho should read it: people working on grey forecasting variants and anyone who cares about the continuous-vs-discrete consistency issue in GM(1,1). For a general statistics audience, the contribution is marginal. I would send it to peer review, because the theoretical result is checkable and the model is a legitimate variant worth having in the literature. The empirical claim of outperformance should be toned down or supported with proper testing before acceptance. I'd ask the authors for code and for a proper comparison of short-horizon forecast errors with uncertainty.","headline":"A modest but genuine theoretical fix for GM(1,1) with Simpson background value; the 'outperforms' claim is not supported by the tiny empirical differences.","tokens_in":15619,"tokens_out":1750,"would_cite":false,"duration_ms":17477,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62M20","65D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A discrete Simpson-formula grey model reproduces exponential sequences exactly and beats three standard grey models on real data.","keywords":["grey forecasting","GM(1,1) model","Simpson formula","background value","discrete grey model","unbiased property","homogeneous exponential sequence","small sample"],"falsifier":"Take the sequence $x^{(0)}(k)=rq^k$ with $q=2.25$ and $r=0.05$ as in Table 2, fit GMSD(1,1), and verify that the restored values equal $rq^k$ to machine precision. Then repeat the comparison on a non-exponential series, such as a linear or quadratic trend or a geometric series with additive noise, and check whether GMSD(1,1) still has the lowest rolling MAPE; if its advantage does not persist, the practical superiority claim is limited to near-exponential data.","tokens_in":14418,"feed_emoji":"📈","tokens_out":8413,"duration_ms":76572,"temperature":0.7,"pith_summary":"This paper proposes a discrete GM(1,1) grey forecasting model, GMSD(1,1), whose background value is computed with Simpson's rule rather than the usual trapezoid average. The central claim is that this model is unbiased for homogeneous exponential sequences: given data $x^{(0)}(k)=rq^k$, its simulated and predicted values coincide with the true series up to computer round-off. The paper further claims that GMSD(1,1) predicts more accurately than the classical GM(1,1), the discrete DGM(1,1), and the continuous Simpson-based GMSC(1,1) on China's electricity consumption, Lanzhou GDP, and Lanzhou freightage. A sympathetic reader would care because small-sample grey forecasting is widely used in planning, and the improvement is obtained without extra data or complicated optimization.","feed_headline":"Simpson-rule grey model reproduces exponential trends exactly","feed_subtitle":"A discrete variant of GM(1,1) with Simpson background values beats three standard grey models in small-sample forecasts.","key_machinery":"The central object is the Simpson background value $z^{(1)}(k)=\\frac{x^{(1)}(k-1)+4x^{(1)}(k)+x^{(1)}(k+1)}{3}$, used to discretize the whitening equation $\\frac{dx^{(1)}(t)}{dt}+ax^{(1)}(t)=b$ over the interval $[k-1,k+1]$. This produces the second-order difference equation $(a+3)x^{(1)}(k+1)+4ax^{(1)}(k)+(a-3)x^{(1)}(k-1)-6b=0$, which the paper rewrites as a first-order linear difference equation in $x^{(1)}(k+1)-wx^{(1)}(k)$, with $w=\\frac{\\sqrt{3a^2+9}-2a}{a+3}$. The time response (Eq. 14) is the exact solution of this difference equation, not a sampled exponential, and that exactness is what makes the model unbiased for homogeneous exponential sequences and free of the parameter-dependent error of the continuous GMSC(1,1).","core_discovery":"The paper's central discovery is that replacing the trapezoidal background value of GM(1,1) with the Simpson-rule value, and crucially using the exact solution of the resulting difference equation as the time response, yields a discrete model that is unbiased for homogeneous exponential sequences. In Section 3.4, for $X^{(0)}(k)=rq^k$, the least-squares estimates of $a$ and $b$ reduce to $3(1-q^2)/(1+4q+q^2)$ and $3rq(1+q)/(1+4q+q^2)$, which forces $w=q$; substituting into the restored-value formula gives $\\hat{x}^{(0)}(k+1)=x^{(0)}(k+1)$, so simulation and prediction coincide with the true series. The paper also shows that the continuous Simpson-based model GMSC(1,1) does not share this property: its exponential time response satisfies the discretized equation only when $|a|$ is small, and its errors on exact exponential data grow with $|a|$.","pith_inferences":["The unbiasedness theorem is exact only for pure geometric sequences; for real series the reported advantage is empirical, so a natural next test is whether the model keeps its edge on noisy or trended data.","Because Simpson's rule uses three consecutive accumulated values, the model effectively incorporates second-order information; this suggests a family of higher-order quadrature grey models, and one could test whether Newton-Cotes formulas of higher order preserve or improve the unbiasedness property.","The paper's suggestion to extend the idea to GM(1,n) and GMC(1,n) could be made concrete by checking whether the exact-solution construction removes parameter-dependence in multi-variable grey models as it does here."],"forward_implications":["For exact homogeneous exponential data, GMSD(1,1) achieves simulation and prediction errors at the $10^{-13}$ percent level, effectively machine precision, while GMSC(1,1) errors grow with $|a|$.","In the electricity consumption example, GMSD(1,1) has the lowest prediction MAPE (3.2669%) among the four models, ahead of GMSC(1,1) (3.5137%), GM(1,1) (5.0428%), and DGM(1,1) (5.1811%).","In the Lanzhou GDP application, GMSD(1,1) has the lowest prediction MAPE (7.6118%) and overall MAPE (5.0454%); in the freightage application, the corresponding values are 6.3579% and 4.3325%.","Because the time response is the exact solution of the model's own difference equation, the model's accuracy does not degrade for large values of the system parameter $a$, a limitation of the continuous Simpson model.","The paper's closing remark proposes that the same discretization idea can be carried over to other grey models such as GM(1,n) and GMC(1,n)."],"supporting_citations":[{"why":"Defines the classical GM(1,1) whitening equation and least-squares parameter estimation that the new model modifies.","marker":"[1]"},{"why":"Introduces the discrete grey model DGM(1,1), which GMSD(1,1) is compared against and whose discretization strategy it extends.","marker":"[23]"},{"why":"Constructs the continuous Simpson-formula model GMSC(1,1), the direct predecessor whose parameter-dependent inconsistency GMSD(1,1) is designed to fix.","marker":"[32]"},{"why":"Supplies the China electricity consumption data used as the first real-data validation of GMSD(1,1).","marker":"[36]"}],"fun_headline_variants":["Simpson-rule grey model hits exponential sequences exactly","Discrete GM(1,1) with Simpson rule: exact for exponentials","Grey model with Simpson background: perfect for exp growth","Simpson-based GM(1,1) unbiased for exponential sequences"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exactness result assumes the data are a pure geometric sequence $rq^k$, and the real-data advantage is demonstrated on only three data sets, so the general claim of superior forecasting rests on those examples.","fun_headline_variants_meta":{"raw":{"variants":["Simpson-rule grey model hits exponential sequences exactly","Discrete GM(1,1) with Simpson rule: exact for exponentials","Grey model with Simpson background: perfect for exp growth","Simpson-based GM(1,1) unbiased for exponential sequences"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000435,"raw_usage":{"total_tokens":2219,"prompt_tokens":952,"completion_tokens":1267,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1194}},"tokens_in":568,"tokens_out":1267,"duration_ms":10424,"temperature":1.0,"reasoning_tokens":1194,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:15:49.118614+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the sequence $x^{(0)}(k)=rq^k$ with $q=2.25$ and $r=0.05$ as in Table 2, fit GMSD(1,1), and verify that the restored values equal $rq^k$ to machine precision. Then repeat the comparison on a non-exponential series, such as a linear or quadratic trend or a geometric series with additive noise, and check whether GMSD(1,1) still has the lowest rolling MAPE; if its advantage does not persist, the practical superiority claim is limited to near-exponential data.","supporting_citations":[{"cited_title":"Deng, Control problems of grey systems, Systems & Contro l Letters 1 (5) (1982) 288–294","cited_arxiv_id":null,"evidence_quote":"Defines the classical GM(1,1) whitening equation and least-squares parameter estimation that the new model modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the discrete grey model DGM(1,1), which GMSD(1,1) is compared against and whose discretization strategy it extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Constructs the continuous Simpson-formula model GMSC(1,1), the direct predecessor whose parameter-dependent inconsistency GMSD(1,1) is designed to fix."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the China electricity consumption data used as the first real-data validation of GMSD(1,1)."}],"review_version":1}