{"id":"465ac684-8309-46e2-ae38-8611f7e03fa6","arxiv_id":"2506.02631","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The ordinary MLE is asymptotically normal for locally stationary Hawkes processes, and a Bernstein-polynomial likelihood ratio test detects any non-constant continuous time-dependent reproduction rate when the degree is chosen large enough.","lead":"Locally stationary Hawkes processes with a time-varying reproduction rate can be estimated by the ordinary maximum likelihood estimator, which is shown to be consistent and asymptotically normal. The paper also builds a test for whether the reproduction rate is constant and applies it to German intraday power market order flow.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract overstates Proposition 2: consistency over C[0,1] requires choosing d above a g-dependent threshold, so no fixed-degree test is consistent over the whole space of continuous alternatives.","rationale":"I read the paper in good faith. The central parametric result, Theorem 1/Corollary 1, follows a known proof scheme (Clinet-Yoshida) with reasonable assumptions; I did not find an internal inconsistency in that argument. The main concern is the scope of Proposition 2 as advertised in the abstract. The proposition is existential in d*(g), not uniform; the test statistic depends on a fixed d chosen by the user. For any fixed d, alternatives with high-frequency components or small Lipschitz constant can be orthogonal to the Bernstein basis of degree d, making the best degree-d fit essentially constant and collapsing the gap in Lemma 13. Hence the abstract's unqualified 'consistent over the whole space of continuous functions' is an overstatement of what the fixed-degree procedure provides. The reader identified exactly this weakness. I also noted minor presentation issues (the LRT definition in Definition 1 appears to have a typo, and Proposition 2's assumption list omits Assumptions 8 and 11), but these do not change the verdict. Since the mathematical claims are likely correct under the stated qualifiers and the missing artifacts (code/data, unpublished annex for Corollary 2) are addressable, a CONDITIONAL verdict remains appropriate.","tokens_in":30322,"tokens_out":24808,"duration_ms":243703,"concrete_test":"For the linear activation Φ(x)=x, set g_m(x)=1+0.5 cos(2π m x) with m=20 and fix d=2. Analytically compute the Bernstein approximation error ∥B_2 g_m - g_m∥_{L2[0,1]} and compare it to inf_C ∥g_m - C∥_{L2[0,1]}. If the ratio exceeds 1/2, the lower bound in Lemma 13 has a non-positive coefficient, so the claimed divergence of Λ_d^T is vacuous. Simulate T=5000, N=2000 replications, and measure the rejection rate at nominal 5%: if the power is near 5%, the fixed-degree test fails for this g_m, confirming that consistency over C[0,1] requires the d*(g) qualifier and that the abstract's wording is misleading.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim that the likelihood-ratio test 'remains consistent over the whole space of continuous functions' is not supported without a qualifier. Proposition 2 asserts only that for each fixed non-constant g* there exists d*(η*,g*) such that Λ_d^T diverges for d>d*. The statistic Λ_d^T itself is computed with a degree d fixed in advance, and the practitioner has no data-driven way to know whether d exceeds d*. For any fixed d, there are continuous alternatives whose best degree-d Bernstein approximation is essentially constant in the likelihood sense—for instance high-frequency, small-amplitude oscillations with m>d—so the gap in Lemma 13 collapses and the test has no power. Thus the abstract's 'consistent over the whole space' is true only in the existential sense 'for each g there is some d', not as a property of a single implementable procedure. The theorem itself is internally consistent; the concern is that the central claim as advertised overstates what the fixed-degree test delivers. Additionally, the statement of Proposition 2 lists only Assumptions 1–3, while its proof invokes Assumptions 8 and 11; this should be corrected for verifiability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers maximum likelihood estimation for a single observed path of a multivariate locally stationary Hawkes process with intensity λ^T_{k,t}(η,g)=Φ_k[μ_k(t/T,η)+Σ_l ∫_0^t g(t/T)φ_{kl}(t-s,η)dN^T_{l,s}]. In the fully parametric case, Theorem 1 and Corollary 1 claim √T-consistency and asymptotic normality of the Ogata MLE, extending earlier results of Ogata, Kwan, and Clinet-Yoshida. In the semi-nonparametric case, the authors propose a likelihood-ratio test based on Bernstein polynomials of fixed degree d to test whether the reproduction rate g is constant. Theorem 2 claims a χ²(d) null limit, and Proposition 2 claims that the test is consistent for every non-constant continuous g provided d exceeds a threshold d*(η*,g*). The paper also reports simulations and an application to German intraday power market order flow.","tokens_in":30558,"tokens_out":10676,"duration_ms":110212,"significance":"If the central results are correct, the paper makes a useful contribution by showing that the classical Ogata MLE remains √T-consistent and asymptotically normal in a locally stationary, multivariate, nonlinear Hawkes model, and by providing a simple misspecification-robust test for time dependence of the reproduction rate. The proof strategy follows the well-established Clinet-Yoshida framework with a locally stationary ergodic approximation, and the numerical/empirical sections give constructive evidence. However, the advertised nonparametric consistency claim is more delicate than the abstract suggests: it is not uniform over all continuous alternatives for a fixed implemented degree d, and the proof of Proposition 2 relies on assumptions that are not listed in its statement. These issues affect the paper's headline claims and need to be corrected before publication.","major_comments":[{"comment":"The abstract states that the likelihood-ratio test \"remains consistent over the whole space of continuous functions of [0,1]\". Proposition 2, however, only proves that for each fixed non-constant g* there exists d*(η*,g*) such that Λ_d^T diverges for every d>d*. The statistic Λ_d^T is computed with a degree d fixed before seeing the data, and no data-driven rule is given to ensure d exceeds d*. Thus for any fixed d there are continuous alternatives—for instance high-frequency, small-amplitude oscillations whose best degree-d Bernstein approximation is nearly constant—for which the likelihood-ratio gap in Lemma 13 collapses and the test has vanishing power. This is not a uniform consistency statement over C[0,1]. The abstract and Section 1.2 should be reworded to state exactly what is proved: for every non-constant continuous g*, there exists a sufficiently large degree d such that the test is consistent, and the practical limitation of the fixed-degree choice should be acknowledged, as the paper itself partly does in Remark 6.","section":"Abstract and §1.2/Proposition 2"},{"comment":"Proposition 2 is stated under Assumptions 1 to 3 only, but its proof invokes Assumption 11 (strict lower bound on Φ'_k) in Lemma 12 and uses Assumption 8 for the positivity of the denominator in the lower bound. Without Assumption 11, the key bound (28) may vanish and the strict positive gap K that drives Λ_d^T→∞ is not guaranteed. The statement of Proposition 2 should therefore include the assumptions actually used, for example Assumptions 6, 8, and 11, or the proof should be modified to avoid them. In addition, Assumption 11 is misprinted: it says \"for any k=1...p\" but the index k runs over the K coordinates of the process, not the dimension p of η.","section":"Proposition 2 and Lemma 12"},{"comment":"The proof of Theorem 2 contains inconsistent counting of the null restrictions. The null hypothesis in (11) is ϖ_0=⋯=ϖ_d, which imposes d restrictions on the d+1 Bernstein coefficients. In §6.2, however, the constrained set is written as Ξ_d^0={ϖ_0=⋯=ϖ_{d-1}}, and later the null constraints are stated as ϑ_{p+1}=⋯=ϑ_{p+d-1}=0, which is d−1 restrictions. The final degrees of freedom are then computed as k=p+d+1−(p+1)=d. These index shifts need to be aligned so that the chi-square limit has the correct number of degrees of freedom; as written, the proof does not rigorously establish χ²(d).","section":"§6.2, proof of Theorem 2"},{"comment":"Corollary 2, which accounts for nuisance parameters on the boundary of the parameter space, is not proved in the manuscript; it is deferred to the self-cited unpublished annex [Lotz, 2024]. Since the boundary issue is explicitly identified as important in Section 2.3 and is relevant to the empirical application, the paper should either include a proof of Corollary 2 or clearly label it as a result from an external unpublished manuscript. The statement also contains a typo: \"for any sufficiently large n\" should presumably read \"for sufficiently large T\", and the statistic is written as Λ_T rather than Λ_d^T.","section":"Corollary 2"}],"minor_comments":[{"comment":"The sentence \"These may be regarded as a standard approach within which falls our proof for Theorems and 1 and 2\" contains a typo and should read \"Theorems 1 and 2\".","section":"§1.3"},{"comment":"The first sentence says \"an empirical illustration of Theorem 2 and Proposition 3\", but the relevant result for the simulation of the time-dependence test is Proposition 2, not the consistency result Proposition 3 in Section 6.1.","section":"§3.1"},{"comment":"The notation Ξ_d^0 is used in Definition 1 before it is formally defined; it would help to define it explicitly at the point of the null hypothesis (11).","section":"§2.3, Definition 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central theorems appear defensible, but the abstract overstates the uniformity of the consistency claim, and the proof of Proposition 2 needs its assumption list corrected. Corollary 2 relies on an unpublished annex by one of the authors and should be either proved or clearly flagged as external. The companion paper [Deschatre et al., 2025] is also self-cited without being available; the authors should make the dependence on this companion work explicit. These issues are fixable within the scope of the manuscript, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [colleague],\n\nYou should know two things. First, the paper shows that the standard Ogata MLE, with no local weighting, is sqrt(T)-consistent and asymptotically normal for a wide class of locally stationary Hawkes processes with time-varying reproduction rate g(t/T). That is a real practical shortcut: practitioners don't need a bespoke local MLE for this class. Second, the authors build a likelihood-ratio test for time-invariance of g, and the test's consistency is real but narrower than the abstract claims. For each non-constant continuous g there is a degree threshold d*(g), and the test works only if the user's Bernstein degree d exceeds it. A fixed-degree test is therefore not consistent over all of C[0,1] as advertised; it is consistent for each g, but not uniformly in g. The abstract says \"consistent over the whole space of continuous functions\" without disclosing the g-dependent d condition. That is an overstatement, not a fatal flaw.\n\nWhat's actually new: Theorem 1 extends Kwan's univariate linear MLE result to multivariate, nonlinear, locally stationary Hawkes processes with Lipschitz activation functions and separable kernels. Proposition 2 (the LRT) is a genuine extension even though the proof scheme is transposed from Clinet-Yoshida and Kwan. The paper is also honest about that in Section 1.3. The German intraday power market application is a nice real-data demonstration.\n\nSoft spots, in rough order of severity:\n\n1. Proposition 2's statement and the abstract. The d* threshold depends on g*, and there is no data-driven way to know if your chosen d exceeds it. Remark 6 gives a bound for Lipschitz g only in a simplified experiment with eta known. This should be stated upfront.\n\n2. Corollary 2, the boundary correction for null nuisance parameters, is deferred to an unpublished annex [Lotz, 2024] by one of the authors. A main result shouldn't rest on an inaccessible self-citation. Either include the proof or drop the corollary.\n\n3. The proof of Theorem 2 has indexing slips: the null hypothesis is written as varpi_0=...=varpi_{d-1} where it should involve d+1 coefficients, and later \"vartheta_{p+1}=...=vartheta_{p+d-1}=0\" appears without clear meaning. These are fixable, but they currently undermine verifiability.\n\n4. Proposition 2's statement lists only Assumptions 1-3, but the proof invokes Assumptions 8 (positive baseline) and 11 (Phi' > epsilon). The statement should list them.\n\n5. No code or data. The simulations are described but not shipped; the power market data is proprietary, so that part is fine, but reproducible simulation code would help.\n\nThe central CLT argument looks sound conditional on the KCD framework; the proofs are structured and the assumptions are stated. The LRT's consistency gap is real but addressable by rephrasing the claim. This paper deserves a serious referee: it extends a useful estimator to a practically relevant class. I would recommend sending it to review, with the expectation that the authors clarify the d-dependence and fix the statement of Proposition 2.\n\nWho is it for? Statisticians working on point processes, and quantitative researchers in market microstructure or energy markets who want a quick test for time-varying endogeneity.\n\nMy recommendation: engage with it. It needs revision, but the core is valuable.","headline":"Useful extension of Ogata MLE to locally stationary Hawkes processes, but the time-invariance test's consistency over C[0,1] is overstated: the Bernstein degree must exceed a g-dependent threshold that the practitioner doesn't know.","tokens_in":31105,"tokens_out":3093,"would_cite":true,"duration_ms":28279,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F03","62F12","60G55"],"pacs":[],"model":"deepseek-v4-flash","headline":"On a single observed path of a locally stationary Hawkes process, the ordinary Ogata maximum likelihood estimator remains $\\sqrt{T}$-consistent and asymptotically normal, and a Bernstein-polynomial likelihood-ratio test detects any…","keywords":["Hawkes processes","locally stationary process","maximum likelihood estimation","likelihood ratio test","Bernstein polynomials","time-varying reproduction rate","endogeneity","intraday power market"],"falsifier":"Simulate a known locally stationary Hawkes process with a non-constant Lipschitz reproduction rate $g^*$, compute the Remark 6 threshold $d \\ge C_{\\mu,\\Phi} K^{-1} \\inf_{C>0} \\|g^*-C\\|^4_{L^2[0,1]}$, fix $d$ above it, and record $\\Lambda^d_T/T$ for growing $T$; Proposition 2 predicts a positive limit and a rejection probability tending to 1, so a flat likelihood ratio or bounded rejection rate across many long paths would refute the consistency claim.","tokens_in":30115,"feed_emoji":"📈","tokens_out":12975,"duration_ms":110979,"temperature":0.7,"pith_summary":"The paper asks whether the standard maximum likelihood estimator developed for stationary Hawkes processes can be used, without modification, on a Hawkes process whose baseline intensity and reproduction rate vary with rescaled time $t/T$. Its answer is yes in the parametric setting: from a single trajectory over $[0,T]$, the MLE is $\\sqrt{T}$-consistent and asymptotically Gaussian, with asymptotic variance given by the integrated Fisher information of the associated stationary processes. It then builds a likelihood-ratio test of the hypothesis that the reproduction rate $g$ is constant on $[0,1]$, proves a $\\chi^2(d)$ limit for $\\Lambda^d_T$ under the null, and shows the test is consistent against every non-constant continuous $g$ provided the Bernstein degree $d$ exceeds a $g$-dependent threshold. An application to German intraday power-market order flow finds statistically significant session-dependent fluctuations in the endogeneity rate.","feed_headline":"One-path MLE stays √T-consistent for time-changing Hawkes processes","feed_subtitle":"The classic Ogata MLE remains valid on one non-stationary path; a polynomial test catches drifting reproduction rates.","key_machinery":"The argument runs through a family of stationary Hawkes processes $N^{x,\\infty}$, indexed by rescaled time $x\\in[0,1]$, with intensities $\\lambda^{x,\\infty}_{k,t}(\\eta,g)=\\Phi_k[\\mu(x,\\eta)+\\sum_l \\int_{-\\infty}^t g(x)\\varphi_{kl}(t-s,\\eta)\\,dN^{x,\\infty}_{l,s}]$, all embedded in the same Poisson noise. Ergodicity of these stationary processes, together with a Riemann-sum scheme borrowed from the univariate theory, converts time-averages along the single non-stationary path into integrals over $x$ of stationary expectations, so the score becomes a Gaussian martingale with covariance $I(\\vartheta^*)$. The time-invariance test relies on the Bernstein basis $B_{k,d}(x)=\\binom{d}{k}x^k(1-x)^{d-k}$, under which '$g$ is constant' is exactly the statement that all $d+1$ Bernstein coefficients are equal.","core_discovery":"The paper's central claim is that the conventional Ogata MLE, applied to the locally stationary intensity $\\lambda^T_{k,t}(\\eta,g)=\\Phi_k[\\mu_k(t/T,\\eta)+\\sum_l \\int_0^t g(t/T)\\varphi_{kl}(t-s,\\eta)\\,dN^T_{l,s}]$ as if the process were stationary, is a legitimate estimator: Theorem 1 and Corollary 1 give $\\sqrt{T}(\\hat\\vartheta_T-\\vartheta^*)\\to N(0,I(\\vartheta^*)^{-1})$ for a single path when the true parameter is interior, with $I(\\vartheta^*)$ the $x$-averaged stationary Fisher information. For the semi-nonparametric question, Theorem 2 shows the likelihood-ratio statistic $\\Lambda^d_T$ for $g$ constant converges to $\\chi^2(d)$ under the null, and Proposition 2 shows $\\Lambda^d_T$ diverges under any non-constant continuous $g$, provided $d$ exceeds a threshold $d^*(\\eta^*,g^*)$; Corollary 2 adjusts the degrees of freedom when nuisance interaction parameters sit on the boundary of the parameter space.","pith_inferences":["The fixed-degree condition means a practitioner must choose $d$ before seeing the data, and the paper gives no data-driven way to know whether $d>d^*$; running the test over several degrees and checking that the MLE stays inside the parameter space is the practical safeguard its own simulations point to.","Because the threshold $d^*$ is governed by how far $g^*$ is from the constant functions (Remark 6 makes this explicit for Lipschitz $g$), the test is hardest precisely for slowly varying, nearly constant reproduction rates; power should increase with the $L^2$ distance from $g$ to the constants.","If the convergence of $\\hat g^d_T$ could be extended to $d=d_T\\to\\infty$, the same likelihood-ratio construction would give a fully nonparametric test and an estimator of the integrated endogeneity $\\int_0^1 g(x)\\,dx$; the paper identifies this as future work.","For intraday power markets, rejection of constant $g$ in 20 of 21 sessions implies that time-dependent baseline models alone may miss an endogenous, participation-driven component of order flow; a direct follow-up would link the estimated $g(t/T)$ to time-to-delivery or market fundamentals."],"forward_implications":["A practitioner can reuse existing stationary-Hawkes MLE code on locally stationary data and still obtain $\\sqrt{T}$ rates and Gaussian confidence intervals, without estimating local windows.","The likelihood-ratio test gives an asymptotically valid level-$\\alpha$ test of constant reproduction rate, with $\\chi^2(d)$ critical values, and a conservative boundary-corrected version when some interactions are zero.","Under the alternative, the test has power tending to 1 against every non-constant continuous reproduction rate, as long as the fixed polynomial degree is chosen above the threshold $d^*$; the required degree does not grow with $T$.","The reproduction-rate estimate $\\hat g^d_T$ tracks smooth time variations of $g$ in simulations, suggesting the same polynomial machinery may support fully nonparametric estimation if the degree is allowed to grow with $T$."],"supporting_citations":[{"why":"Baseline stationary-point-process MLE whose asymptotic normality the paper extends to locally stationary Hawkes processes.","marker":"[Ogata, 1978]"},{"why":"Univariate proof scheme for consistency and weak convergence of the MLE with time-dependent baseline and constant reproduction rate; the paper transposes this scheme directly.","marker":"[Kwan, 2023, Chapter 3]"},{"why":"Supplies the sufficient-conditions framework (ergodicity, moment bounds, master theorem) that Theorems 1 and 2 verify in the locally stationary setting.","marker":"[Clinet and Yoshida, 2017]"},{"why":"Imbedding representation and stability of nonlinear Hawkes processes, used to construct the stationary copies $N^{x,\\infty}$ and obtain moment bounds.","marker":"[Brémaud and Massoulié, 1996]"},{"why":"Cluster representation defining the reproduction rate as mean offspring number, the parameter whose constancy is tested.","marker":"[Hawkes and Oakes, 1974]"},{"why":"Sets up the locally stationary Hawkes formalism with kernel depending on $t/T$ that the paper estimates.","marker":"[Roueff et al., 2016]"},{"why":"Provides the log-likelihood expression and the mixing/ergodicity background used throughout the proofs.","marker":"[Daley and Vere-Jones, 2002]"},{"why":"Supplies the M-estimator well-separatedness theorem and the standard likelihood-ratio expansion used for Theorem 2.","marker":"[Van den Vaart, 1998]"}],"fun_headline_variants":["Classic MLE stays consistent for non-stationary Hawkes processes","Single-path MLE works for time-varying Hawkes processes","Time invariance test catches drifting reproduction in Hawkes","Non-stationary Hawkes: MLE remains asymptotically normal","One path enough: MLE for locally stationary Hawkes is valid"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the process stays subcritical at every rescaled time (each event triggers fewer than one descendant on average) and that, for the nonparametric test, the practitioner's fixed polynomial degree $d$ already exceeds an unknown threshold depending on the true reproduction rate---if $d$ is too small, the best degree-$d$ polynomial fit to the true rate can be constant and the likelihood-ratio gap collapses.","fun_headline_variants_meta":{"raw":{"variants":["Classic MLE stays consistent for non-stationary Hawkes processes","Single-path MLE works for time-varying Hawkes processes","Time invariance test catches drifting reproduction in Hawkes","Non-stationary Hawkes: MLE remains asymptotically normal","One path enough: MLE for locally stationary Hawkes is valid"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000562,"raw_usage":{"total_tokens":2684,"prompt_tokens":977,"completion_tokens":1707,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":1621}},"tokens_in":593,"tokens_out":1707,"duration_ms":11954,"temperature":1.0,"reasoning_tokens":1621,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:20:43.979280+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a known locally stationary Hawkes process with a non-constant Lipschitz reproduction rate $g^*$, compute the Remark 6 threshold $d \\ge C_{\\mu,\\Phi} K^{-1} \\inf_{C>0} \\|g^*-C\\|^4_{L^2[0,1]}$, fix $d$ above it, and record $\\Lambda^d_T/T$ for growing $T$; Proposition 2 predicts a positive limit and a rejection probability tending to 1, so a flat likelihood ratio or bounded rejection rate across many long paths would refute the consistency claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline stationary-point-process MLE whose asymptotic normality the paper extends to locally stationary Hawkes processes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Univariate proof scheme for consistency and weak convergence of the MLE with time-dependent baseline and constant reproduction rate; the paper transposes this scheme directly."},{"cited_title":"and Yoshida, N","cited_arxiv_id":null,"evidence_quote":"Supplies the sufficient-conditions framework (ergodicity, moment bounds, master theorem) that Theorems 1 and 2 verify in the locally stationary setting."},{"cited_title":"and Oakes, D","cited_arxiv_id":null,"evidence_quote":"Cluster representation defining the reproduction rate as mean offspring number, the parameter whose constancy is tested."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sets up the locally stationary Hawkes formalism with kernel depending on $t/T$ that the paper estimates."},{"cited_title":"An Introduction to the Theory of Point Processes: elementary theory and methods","cited_arxiv_id":null,"evidence_quote":"Provides the log-likelihood expression and the mixing/ergodicity background used throughout the proofs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the M-estimator well-separatedness theorem and the standard likelihood-ratio expansion used for Theorem 2."}],"review_version":1}