{"id":"38accf6c-3f4d-4e05-ad0e-908152caf7d6","arxiv_id":"2411.13995","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A fractional Poisson process fitted to California wildfire data with λ=0.69 and β=0.8 is reported to predict the next ten fire dates about 90% more accurately than a Poisson process.","lead":"This paper models California wildfire occurrence dates using a fractional Poisson process, a counting model that permits bursts and long memory. The authors report the fitted model predicts future fire dates with about 90% lower error than a standard Poisson model, suggesting dependence between wildfire events.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dependence and 90% prediction claims rest on a two-parameter fit from one count N(200)=56 with no uncertainty and on a single unseeded forecast draw; the full interarrival data are never used.","rationale":"The reader identified exactly the load-bearing weakness: two parameters are identified from a single realized count via method of moments, with no sampling basis or uncertainty quantification. I checked whether any other step is more fragile. The fractional Poisson formulation and the simulation-based validation in §3-4 are standard and plausible as small-scale numerical checks, but the application in §5.2 discards the rich interarrival structure and reduces the entire dataset to one number, N(200)=56. The paper's own §5.3 admits the resulting difficulty for MLE but does not quantify it for MoM. The prediction comparison in §6 then compounds this by using a single unseeded simulated sequence, so the reported 90% reduction is not a probabilistic statement about the model. A re-analysis using the observed interarrival times for estimation and repeated-simulation predictive evaluation would settle whether the dependence and prediction claims survive. Since the reader already recommended rejection and this assessment keeps that recommendation, the verdict is unchanged.","tokens_in":9199,"tokens_out":7740,"duration_ms":74980,"concrete_test":"Re-fit the fPP to the actual 56 (or 146) event times/interarrival times by MLE on the Mittag-Leffler interarrival density (or by using all counts at multiple t values), rather than moment-matching a single count. Compute a 95% confidence interval for β from the observed information matrix or profile likelihood; if the interval includes 1, the dependence claim fails. Then, from the fitted model, simulate 10,000 realizations of the next 10 interarrival times and form a predictive interval for the next occurrence time, or a distribution of MSE/MAD; compare the actual next 10 occurrence times against the PP under the same repeated-simulation scheme using CRPS or a rolling-origin design. If the distribution of MSE/MAD does not show a >90% reduction in at least 95% of simulations, the prediction claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claims collapse to one underidentified statistical step and one unseeded simulation. In §5.2 the authors fix t=200, observe a single count Nβ(200)=56, and solve moment equations (4)-(5) by treating nt and n_t^2 as sample moments. With one count path, equations (4)-(5) have no sampling distribution; no standard errors, confidence interval, or profile likelihood is reported. The point estimates λ*=0.69 and β*=0.8 are therefore not evidence that β differs from 1, and the statement that β=0.8 'proves' dependence is unsupported. §5.3 even concedes MLE is impossible from a single sample value, but the data actually contain 56 observed event/interarrival times from the first 200 days (and 146 occurrence days overall), which are discarded by the count reduction. In §6, the 90% prediction-error reduction comes from comparing the actual next 10 occurrence times with one generated sequence of fPP interarrival times (Table 3). A single draw of a heavy-tailed interarrival process can land near the observed gap by luck; no averaging, seeding, predictive distribution, or uncertainty propagation is performed. Both the dependence claim and the superiority claim therefore rest on quantities that could change entirely under a different draw or a more honest estimation procedure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the fractional Poisson process (fPP) as a model for wildfire occurrences in California between June 2019 and April 2023. It introduces method-of-moments and maximum-likelihood estimation approaches for the fPP parameters, validates them in simulation, and then applies the method-of-moments to a single observed count N(200)=56, obtaining λ*=0.69 and β*=0.8. The authors interpret β*=0.8 as evidence of dependence/long memory in wildfire events. They then generate ten fPP and ten Poisson interarrival times, compare the resulting predicted occurrence times with the actual next ten occurrence times, and report a roughly 90% reduction in MSE/MAD relative to the Poisson process.","tokens_in":9550,"tokens_out":6954,"duration_ms":62697,"significance":"If the claims were reliable, the finding that California wildfire occurrences exhibit long-range dependence and can be predicted far better by a fractional Poisson process than by a homogeneous Poisson process would be practically relevant for wildfire risk planning. The paper also provides a simulation comparison of two estimation methods for the fPP, which could serve as a reference if the methods were properly validated. However, the central applied conclusions rest on two very fragile statistical steps: estimating two parameters from a single realized count with no uncertainty quantification, and comparing forecasts on the basis of a single simulated interarrival sequence. Neither step supports the strength of the claims made in the abstract. The paper does not provide code, data, or a reproducibility statement, which further limits verification of the simulation-based results.","major_comments":[{"comment":"The parameters λ and β are estimated by solving equations (4) and (5) with nt = Nβ(200) = 56 and nt² = 56², i.e., from a single realized count. A single count has no sampling distribution, and treating nt and nt² as sample moments provides no basis for inference. No standard errors, confidence intervals, sensitivity analysis, or profile likelihood is reported, so the estimate β*=0.8 cannot be used to claim that β differs from 1, let alone that it 'proves' dependence. Moreover, the full interarrival/occurrence-time data (146 occurrence days) are available but are discarded by the count reduction; using the full data would be needed to identify the dependence parameter reliably.","section":"§5.2, Eqs. (4)–(5)"},{"comment":"The reported 90% prediction-error reduction is computed from one simulated sequence of ten fPP interarrival times and one simulated PP sequence. Because the fPP interarrival distribution is heavy-tailed, a single Monte Carlo draw can be unrepresentative; the MSE values (526.14 vs. 39562.1) are realized values with no measure of variability. A proper comparison would report the mean and spread of the prediction error over many simulated sequences (or a predictive distribution) and would propagate the uncertainty in the parameter estimates. Additionally, the 'time of occurrence' values in Table 3 are consistent with adding the simulated interarrival times to an origin at day 146—the total number of occurrence days in the full dataset—but the paper never states that the 56th occurrence happens on day 146, making the forecast construction ambiguous.","section":"§6, Tables 3–4"},{"comment":"The simulation study solves equations (4)–(5) separately for each of 1000 simulated paths (each path yields a single count), producing per-path estimates λ*_j, β*_j that are then averaged. This is not the standard method of moments, which would equate the sample mean and second moment over the 1000 independent counts to the theoretical moments in (2) and (3). The per-path procedure may produce estimates even when a single count is inconsistent with the two-moment system, and the reported bias and MSE describe that ad-hoc per-path estimator rather than a proper MoM estimator. Since the real-data application in §5.2 follows the same single-count logic, the simulation does not validate the step actually used.","section":"§3, MoM simulation"},{"comment":"The initial condition for the fPP pmf is stated as 'pβ(n, 0) = 0 if n = 0 and is zero if n ≥ 1', which gives pβ(0,0)=0. The correct condition is pβ(0,0)=1 and pβ(n,0)=0 for n≥1. As written, the pmf (1) cannot be consistent with a proper probability distribution at t=0. This is a basic definitional error that affects the likelihood and the moment equations derived later.","section":"§2.2, initial condition"}],"minor_comments":[{"comment":"The sentence 'Using the method of moments equations, (5) and (6)' should refer to equations (4) and (5); equation (6) is a bias formula.","section":"§5.2"},{"comment":"The statement 'Finding the MLE is hence not possible from a single sample value' is imprecise: the likelihood function is still defined for a single count and can in principle be maximized, though parameter identifiability from one observation is poor. The authors should explain why MLE is infeasible or unstable rather than impossible.","section":"§5.3"},{"comment":"The phrase 'MSE (as defined in Eq. (9))' is a dangling reference; equation (9) is the observed information matrix, and no MSE formula appears in Section 6.","section":"§6, Table 4"},{"comment":"The description of the ECDF comparison is unclear about how the ECDFs for the two fitted processes are constructed (e.g., by simulating many sample paths and pooling counts, or by using a theoretical distribution). Without this detail, the reported Kolmogorov-Smirnov p-values (0.1761 and 0.00034) are not reproducible, and the fact that parameters were estimated from the same data should be addressed when interpreting those p-values.","section":"§5.1–5.2, Figure 6"},{"comment":"In the likelihood expression (8), the threshold t appears without being indexed or declared as a known constant; the notation should be made explicit (e.g., L(λ, β; t)) so that the dependence on t is clear.","section":"§4, Eq. (8)"}],"recommendation":"reject","confidential_remarks":"The manuscript's central claims are not supported by the analysis. Estimating two parameters from a single count and comparing forecasts on a single Monte Carlo draw are fundamental flaws that cannot be fixed by local edits; the paper would need a substantially different statistical treatment (e.g., using the full interarrival data, reporting uncertainties, and averaging over simulations). I also note that method-of-moments and maximum-likelihood estimation for the fPP have been studied in the prior literature, so the novelty claim regarding 'introducing' these methods should be carefully reassessed in any revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nBottom line: the paper's two headline claims — that β=0.8 proves long-memory dependence in California wildfires and that fPP cuts prediction error by 90% — are not supported by the evidence, because the estimation uses a single count observation and the prediction comparison uses a single simulated draw.\n\nWhat's genuinely new and useful: applying the fPP to wildfire occurrence data is a fresh application, and the paper gives a clear derivation of the fPP count pmf, plus simulation results suggesting the proposed estimators behave well when the model is exactly true. The authors are also honest about MLE being infeasible from a single count, and the writing is clear.\n\nThe soft spots are substantial. Section 5.2 fits the fPP by taking Nβ(200)=56 and solving equations (4)-(5) for λ and β. That's not method of moments; it's two equations with one data point. There's no sampling distribution, no standard error, no basis for claiming β differs from 1. The full interarrival series (146 occurrence days) is never used for estimation, so the testable information is thrown away. Section 6 then compares predictions using a single generated sequence of interarrival times. The fPP's heavy tails produce a few huge gaps, which happen to land near the actual cluster of later events, but that's one random draw. The 90% reduction in MSE/MAD is computed from that one draw; without averaging or prediction intervals it's meaningless. The ECDF comparison and KS test in Section 5.2 are also hard to interpret because the parameters were fit to a count, not to the interarrival data used in the test.\n\nThere are smaller issues: the initial condition for the fPP pmf is misstated (should be pβ(0,0)=1), and the paper cites its own earlier work once without relevance to the analysis.\n\nWho is this for? A reader curious about fPP applications might skim the simulation section. It's not a reliable contribution to wildfire modeling or to fPP estimation.\n\nRecommendation: I would not send this to peer review in its current form. The load-bearing statistical steps are invalid, and a serious referee would need the estimation redone on the actual interarrival data. If the authors do that, the idea might be worth a look. For now, this is a desk-reject.\n\nBest,\n[Your name]","headline":"The fPP application to wildfires is new, but the real-data analysis is unsound: two parameters estimated from one count, and a 90% prediction improvement from a single simulated draw.","tokens_in":10011,"tokens_out":6233,"would_cite":false,"duration_ms":52159,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that California wildfire occurrences are dependent events with fractional parameter β ≈ 0.8, and that a fractional Poisson process predicts the next fire dates about 90% more accurately than the classical Poisson process.","keywords":["wildfire modelling","fractional Poisson process","long memory","method of moments","maximum likelihood","Mittag-Leffler distribution","California wildfires","forecasting"],"falsifier":"Refit the full set of 146 recorded interarrival times with a maximum-likelihood estimator for a Mittag-Leffler waiting-time distribution, and compute a confidence interval for $\\beta$; if the interval covers 1, the dependence claim collapses. Alternatively, repeat the ten-step forecast on many independently simulated draws from the fitted fractional process and compare average error to the Poisson model; if the 90 percent reduction does not persist, the prediction claim fails.","tokens_in":9031,"feed_emoji":"🔥","tokens_out":6742,"duration_ms":63350,"temperature":0.7,"pith_summary":"The paper argues that wildfire occurrences in California between June 2019 and April 2023 should be modeled by a fractional Poisson process rather than a classical Poisson process. The fractional model allows past fires to influence the timing of future fires, and the authors estimate the fractional parameter at 0.8, which they read as evidence of long-memory dependence. On the same data, their fitted fractional model predicts the next ten fire dates with roughly a 90 percent smaller mean-squared error than the Poisson model. If correct, the result would replace the standard independence assumption with a concrete dependence parameter and improve operational forecasts.","feed_headline":"California wildfires are dependent, not Poisson-random","feed_subtitle":"A fractional Poisson fit gives dependence 0.8 and 90% lower forecast error.","key_machinery":"The central object is the fractional Poisson process (fPP), a counting process whose waiting times between events follow a Mittag-Leffler distribution; the fractional order $\\beta \\in (0,1]$ controls deviation from Poisson behavior, with $\\beta = 1$ recovering the exponential interarrival times of the classical Poisson process. The argument is carried by two moment identities: the mean count $E[N_\\beta(t)] = q t^\\beta$ and a variance formula involving the $\\beta$ function, which the method-of-moments procedure equates to an observed count and its square to solve for $\\lambda$ and $\\beta$. Once estimated, the model generates future interarrival times through a stochastic representation for Mittag-Leffler variables, enabling the date-by-date forecast comparison.","core_discovery":"The central claim is that a two-parameter fractional Poisson process, with rate $\\lambda^* = 0.69$ and fractional order $\\beta^* = 0.8$, describes California wildfire occurrence better than a one-parameter Poisson process with estimated rate 0.28. Because $\\beta$ is below 1, the fitted model says wildfire interarrival times are Mittag-Leffler distributed rather than exponential, so events are not independent: past occurrences statistically affect later ones. The paper supports this by solving the process's first two theoretical moment equations against a single observed count of 56 fires in the first 200 days, and then reports a goodness-of-fit test that favors the fractional process. It further claims a 90 percent reduction in prediction error over the Poisson model when forecasting ten future fire dates. The paper frames both findings as evidence that wildfire events carry long memory.","pith_inferences":["A fairer test of the 90 percent claim would average prediction error over many simulated forecast origins, since the reported comparison uses a single random draw of ten future interarrival times.","The single-count calibration provides no confidence interval for $\\beta$, so a bootstrap over resampled interarrival times could show whether 0.8 is meaningfully different from 1.","The estimated $\\beta < 1$ may partly absorb seasonal clustering rather than genuine long memory; fitting a seasonal Poisson or self-exciting process would separate those mechanisms.","If the moment equations are ill-identified from one count, a Bayesian fit using the observed 146 waiting times would give a more direct estimate of $\\beta$ and a built-in uncertainty band."],"forward_implications":["Wildfire occurrence in California should not be treated as independent events; any count or forecast model that assumes Poisson arrivals will misstate the clustering.","Because the fitted fractional parameter is 0.8, the dependence is strong enough to matter for practical prediction: the fitted model places the next ten fire dates much closer to observed dates than the Poisson model.","The method-of-moments and maximum-likelihood estimators introduced for the two fPP parameters can be used on other disaster-event timelines, not just wildfires.","The long-memory finding would justify developing early-warning systems that use recent fire activity as an input, rather than treating each fire as a fresh random event."],"supporting_citations":[{"why":"Defines the fractional Poisson process and gives the mean-variance formulas that the method-of-moments equations are built on.","marker":"[17]"},{"why":"Provides the earlier estimation approach that this paper offers an alternative to, and sets the context for parameter estimation.","marker":"[7]"},{"why":"Supplies the random-variable representation used to simulate future interarrival times for both processes.","marker":"[13]"},{"why":"Establishes the long-range dependence property that connects $\\beta < 1$ to dependence in wildfire events.","marker":"[18]"},{"why":"Shows fPP increments have the long-range dependence that makes $\\beta < 1$ evidence of event dependence.","marker":"[20]"}],"fun_headline_variants":["Wildfires show memory: fractional Poisson beats standard model","Fractional Poisson fits California fires, cuts error 90%","Wildfire timing isn't random: memory effect found, error down 90%","California fires: dependence 0.8, prediction error slashed 90%","New model: wildfires are dependent, with 90% better forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire analysis rests on treating one observed number—56 fires within the first 200 days, and its square—as two sample moments that can uniquely determine both $\\lambda$ and $\\beta$, with no sampling distribution or uncertainty attached.","fun_headline_variants_meta":{"raw":{"variants":["Wildfires show memory: fractional Poisson beats standard model","Fractional Poisson fits California fires, cuts error 90%","Wildfire timing isn't random: memory effect found, error down 90%","California fires: dependence 0.8, prediction error slashed 90%","New model: wildfires are dependent, with 90% better forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1215,"prompt_tokens":821,"completion_tokens":394,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":301}},"tokens_in":437,"tokens_out":394,"duration_ms":4270,"temperature":1.0,"reasoning_tokens":301,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:39:51.245873+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Refit the full set of 146 recorded interarrival times with a maximum-likelihood estimator for a Mittag-Leffler waiting-time distribution, and compute a confidence interval for $\\beta$; if the interval covers 1, the dependence claim collapses. Alternatively, repeat the ten-step forecast on many independently simulated draws from the fitted fractional process and compare average error to the Poisson model; if the 90 percent reduction does not persist, the prediction claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the fractional Poisson process and gives the mean-variance formulas that the method-of-moments equations are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the earlier estimation approach that this paper offers an alternative to, and sets the context for parameter estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the random-variable representation used to simulate future interarrival times for both processes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the long-range dependence property that connects $\\beta < 1$ to dependence in wildfire events."},{"cited_title":"Maheshwari and P","cited_arxiv_id":null,"evidence_quote":"Shows fPP increments have the long-range dependence that makes $\\beta < 1$ evidence of event dependence."}],"review_version":1}