{"id":"ac12a2bc-d8f0-4730-b8b3-58c84b169b56","arxiv_id":"1908.06835","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A particle-filter algorithm samples the spectral measure of GARCH(p,q) processes, enabling computation of the tail index, extremogram, and extremal index for general GARCH models.","lead":"This paper develops numerical algorithms to compute extreme-value properties of GARCH volatility models, such as the tail index and the clustering of large losses. It claims to work for nearly all GARCH model families used in finance, including models with persistent volatility and heavy-tailed or skewed innovations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4's independence between Bernoulli signs and the squared tail chain fails for asymmetric innovations, invalidating the claimed extremogram and extremal index for that broad class.","rationale":"The reader's weakest assumption identifies precisely the same load-bearing concern: the independent Bernoulli sign thinning in Section 4. This is the most consequential vulnerability because the paper's headline contribution is to compute extremal cluster functionals for all GARCH(p,q) processes including asymmetric innovations, and Section 4 is the only step that converts the squared-process tail chain, whose simulation is genuinely well motivated and empirically supported for symmetric cases, into the original-process tail chain. If the independence fails, the computed extremogram and extremal index for asymmetric innovations are not approximations with a controlled error; they are derived from a false model. The concern is not merely that δ might be misestimated but that the conditional sign probability depends on \\hat X_t^2, so no constant δ can salvage the representation. A secondary issue in the proof of Theorem 3.4, where all spectral-mean components are asserted equal to 1/(p+q), is also present but less central to the algorithmic claims; the numerical algorithms for the squared process and for κ appear sound and are cross-validated against long-run simulations for symmetric models. Because the reader already recommends a conditional acceptance contingent on fixing or restricting the asymmetric case, my read does not change the verdict.","tokens_in":28161,"tokens_out":9889,"duration_ms":97962,"concrete_test":"For the ARCH(1) model X_t = σ_t Z_t with σ_t^2 = α0 + α1 X_{t-1}^2 and a standardized skew-normal innovation, use the exact tail-chain recursion \\hat X_t = sqrt(α1) \\hat X_{t-1} Z_t (α0 asymptotically negligible) for t≥1. Simulate N=10^6 chains conditional on \\hat X_0>1 and estimate χ_X(1)=Pr(\\hat X_1>1|\\hat X_0>1) directly. Compare this with the paper's Bernoulli-thinning formula δχ_{X^2}(1), where δ is computed from Section 5.7 and χ_{X^2}(1)=Pr(α1 Z_1^2>1). If the two estimates differ beyond Monte Carlo error, the independence assumption in (4.2) is refuted for asymmetric innovations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires methods for all GARCH(p,q) processes, including asymmetric innovations. The bridge from the squared tail chain to the original GARCH tail chain in Section 4 is representation (4.2): \\hat X_t^U = I_t (\\hat X_t^2)^{1/2}, with I_t iid Bernoulli(δ) and independent of {\\hat Y_t}. This independence is the load-bearing step for all subsequent cluster functionals. For t≥1, \\hat X_t^2 is a function of Z_t^2 through the recursion \\hat Θ_t = A_t \\hat Θ_{t-1} (Section 3.3, (3.10)), and the true sign of \\hat X_t is sign(Z_t). For an asymmetric innovation distribution, sign(Z_t) and Z_t^2 are dependent, so the conditional probability of a positive sign given \\hat X_t^2 is not the constant δ; it varies with the realized magnitude of Z_t. Moreover, δ in (4.1) is the limiting probability Pr(X_t>x | |X_t|>x), conditioning on the contemporaneous |X_t| being extreme, whereas inside the tail chain Z_t^2 need not be extreme. Therefore the identities χ_{X^U}(τ)=δχ_{X^2}(τ), the cluster-size formulas, and θ_{X^U}=θ_{X^2}(1-Π_U)/δ in Section 4 are not generally valid for asymmetric innovations. The same dependence also breaks the iid-Bernoulli structure at t=0, where I_0 is degenerate at 1 in the upper tail chain. This directly undermines the claimed scope for asymmetric innovations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops numerical algorithms for extremal properties of GARCH(p,q) processes. It rewrites the squared GARCH process as a stochastic recurrence equation, introduces a sequential importance sampling algorithm (Algorithm 1) to sample the spectral measure of the state vector, uses this to evaluate the tail index κ through a self-consistency equation, and constructs the forward tail chain of the squared process (Algorithm 2). From the squared tail chain, Section 4 proposes to obtain the upper and lower tail chains of the original GARCH process by multiplying by independent Bernoulli(δ) sign variables, leading to formulas for the extremogram, extremal index, and cluster size distribution. The paper also states new Lyapunov-exponent representations (Theorems 3.1–3.3), claims that all IGARCH processes have κ = 1 (Theorem 3.4), and reports numerical results for several GARCH models with Gaussian, symmetric t, and skew-t innovations.","tokens_in":28472,"tokens_out":31311,"duration_ms":280320,"significance":"The particle-filter machinery for the squared GARCH process is a genuine methodological contribution: it applies to unbounded innovations, appears to converge quickly in the examples, and the squared-process extremal index is checked against runs estimates from long simulations (Section 5.5). The κ evaluation via the fixed point ρ_k = 1 is also a practical new route to the tail index. However, the bridge from the squared tail chain to the original GARCH process in Section 4 contains load-bearing flaws: the sign variables are not independent of the squared tail chain when innovations are asymmetric, and the initial exceedance is mishandled in the cluster-size formulas even in the symmetric case. The proof of Theorem 3.4 is also not justified as written. These problems affect the headline claims about extremal indices and cluster functionals for GARCH processes, especially for asymmetric innovations, and the numerical entries for skew-t models in Table 1 are therefore not established.","major_comments":[{"comment":"The representation \\hat X_t^U = I_t (\\hat X_t^2)^{1/2} with I_t an iid Bernoulli(δ) sequence independent of the squared tail chain is not valid for asymmetric innovations. For t ≥ 1 the squared tail chain value \\hat X_t^2 is a function of Z_t^2 through the matrix A_t in the recursion (3.10), whereas the sign of \\hat X_t is sign(Z_t); when Z_t is skewed, sign(Z_t) and Z_t^2 are dependent, so the conditional probability of a positive sign given the squared tail chain is not the constant δ. Moreover, δ in (4.1) is the limiting probability that a contemporaneous large |X_t| is positive, which is not the conditioning that applies to the tail chain at positive lags. Consequently the identities χ_{X^U}(τ)=δ χ_{X^2}(τ), the binomial cluster-size formulas, and θ_{X^U}=θ_{X^2}(1-Π_U)/δ are unsupported for skew-t innovations, and the skew-t entries in Table 1 and the corresponding Section 6 conclusions are not established.","section":"Section 4, Eq. (4.2)"},{"comment":"Even under the Bernoulli-thinning assumption, the cluster-size derivation mishandles the initial time. The upper tail chain is conditioned on \\hat X_0^U > 1, which forces I_0 = 1, but the formulas for Π_U, π_{X^U}(j), and θ_{X^U} treat I_0 as an ordinary Bernoulli(δ) variable and allow the whole cluster to vanish. For a squared-process cluster of length k, the upper-chain cluster length is 1 + Binomial(k-1, δ), not a binomially thinned count that can be zero. The correct extremal-index relation under the thinning model is θ_{X^U} = θ_{X^2}/[δ + (1-δ)θ_{X^2}]; for example, Model B-1 in Table 1 gives 0.55 rather than the reported 0.49. Thus the cluster-size and extremal-index formulas of Section 4 are wrong even for symmetric innovations.","section":"Section 4, cluster-size formulas"},{"comment":"The proof of Theorem 3.4 asserts that 'as E(Z_t^2)=1, it follows that all E(\\hat ϑ_t^{(i)}) = 1/(p+q)'. This step is not justified: the preceding sentence only establishes identical marginal distributions within the first q block and within the last p block, and the argument that the two blocks have equal means because E(Z_t^2)=1 is not supplied. Since the theorem is used to assert κ = 1 for IGARCH models C and D and to compute η and γ for those models in Table 1, the proof needs to be completed (or the result cited) before the IGARCH claims can be accepted.","section":"Appendix A, proof of Theorem 3.4"}],"minor_comments":[{"comment":"The displayed definition of the runs estimator contains 'Pr(X^2_t < 0 | X^2_0 > 1)', which should read 'Pr(X^2_t < u | X^2_0 > u)' or similar, since the empirical formula that follows uses the threshold u.","section":"Section 5.5, runs estimator"},{"comment":"The sentence 'min(θ_{X^U},θ_{X^U}) ≥ θ_{X^2}' should read 'min(θ_{X^U},θ_{X^L}) ≥ θ_{X^2}'.","section":"Section 6.2"},{"comment":"Immediately after (4.2), the extremogram display writes 'Pr(I_t \\hat X^2_τ > 1)'; the index should be I_τ for consistency with the lag τ being considered.","section":"Section 4, Eq. (4.2)"},{"comment":"The expression for \\tilde ρ_k mixes a Monte Carlo average over particles with an integral with respect to F_Z; please clarify the exact estimator, for instance by stating that it averages over independent draws of A and Z.","section":"Section 3.2, Eq. (3.7)"},{"comment":"The reference 'Kallemberg (1983)' should be 'Kallenberg (1983)'.","section":"References"},{"comment":"Model E is described as p = 2, q = 0, but the model definition requires q ∈ N+ and the parameter list includes α1 and α2; please clarify the parametrization.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The central issue is Section 4: the asymmetric-innovation claim requires a substantially different derivation than Bernoulli thinning of the squared tail chain, and the symmetric-case cluster formulas also need correction at the initial time. If the authors can fix or properly restrict these claims, the squared-process methodology and the κ-evaluation algorithm are valuable enough to warrant publication. The proof of Theorem 3.4 should be checked carefully; if the result is already known, a citation would be simpler than an incomplete proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core algorithmic contribution is real: a particle filter that samples the spectral measure of the squared GARCH(p,q) process with max(p,q)>=2 and unbounded innovations, plus a method to generate the forward tail chain and estimate kappa. That fills a genuine gap: the Basrak-Segers MCMC algorithm is unusable for GARCH(p,q), and Janssen's bounded-support method is slow and restrictive. The numerical checks against long simulations for the squared process are convincing, and the authors are honest about the numerical instabilities they are solving. Second, the paper's headline scope is too big. The Section 4 bridge from the squared chain to the original chain assumes iid Bernoulli(delta) signs independent of the squared chain. That holds for symmetric innovations, but not for asymmetric ones: the sign of X_t is sign(Z_t), and Z_t^2 drives the squared tail chain through the recursion, so signs and magnitudes are dependent. For skew-t innovations, theta_{X^U} and theta_{X^L} in Table 1 are not reliable. This is not a minor technicality; it is load-bearing for the claimed asymmetric coverage.\n\nThe other soft spot is Theorem 3.4. The proof that kappa=1 for IGARCH hinges on all spectral components having mean 1/(p+q), which is asserted without support and is not generally true, and the \"argument is simply reversed\" is not a valid converse. The result may be true, but this proof does not establish it. Theorem 3.2 also has an interchange of limit and expectation that is not justified; that one looks fixable.\n\nOn the positive side, the paper is well organized, the simulations are extensive, and the numerical evaluation of kappa via particle filtering plus the squared tail chain are worth having even if the asymmetric extension needs rewriting. The paper deserves serious peer review, but with major revision: prove or repair the sign-thinning representation (or restrict claims to symmetric innovations), fix the IGARCH proof, and tighten Theorem 3.2. I would not desk reject it.","headline":"A genuinely useful particle-filter method for the squared GARCH tail chain, but the asymmetric-innovation extension rests on an invalid independence assumption and the IGARCH proof has a gap.","tokens_in":29047,"tokens_out":4928,"would_cite":true,"duration_ms":52837,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G70","62M10","60H25"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a particle-filter algorithm can generate the tail chain of any GARCH(p,q) process, making the extremal index, extremogram, and cluster-size distribution numerically available for all such models.","keywords":["cluster of extremes","extremal index","fixed point distributions","GARCH process","multivariate regular variation","particle filtering","stochastic recurrence equations","tail chain"],"falsifier":"For a GARCH(2,2) with skew-$t$ innovations, simulate a long stationary series, estimate at a high threshold (e.g., the 0.9999 quantile) the lag-one upper-tail extremogram $\\chi_{X^U}(1)$ by the threshold method, and compare it with the paper's value $\\delta \\chi_{X^2}(1)$; a systematic gap beyond Monte Carlo error would show the Bernoulli-thinning representation of Section 4 does not hold for asymmetric innovations.","tokens_in":27915,"feed_emoji":"📈","tokens_out":15102,"duration_ms":131488,"temperature":0.7,"pith_summary":"GARCH(p,q) models dominate volatility modelling, but for $\\max(p,q) \\ge 2$ the quantities that matter for risk — the extremal index, the extremogram, and the cluster-size distribution — had no numerical evaluation scheme that works for real-world innovation distributions. The paper claims to supply one: a particle-filter method that samples the spectral measure of the squared GARCH process, generates its forward tail chain, and reads off cluster functionals, for every strictly stationary GARCH(p,q), including integrated GARCH and processes with unbounded or asymmetric innovations. It also proves new identities tying the top Lyapunov exponent $\\gamma$, the tail index $\\kappa$, and the largest eigenvalue $\\lambda$ of the recurrence matrix, and shows $\\kappa = 1$ for all IGARCH(p,q). If these claims are right, extreme-event clustering in GARCH can be quantified routinely without special-case formulas or impossibly long simulations.","feed_headline":"New algorithm computes extreme-event clustering for every GARCH(p,q)","feed_subtitle":"Spectral-measure sampling yields extremal index, extremogram, and cluster sizes for integrated and asymmetric GARCH.","key_machinery":"The load-bearing device is a sequential importance-sampling particle filter for the spectral measure $H_{\\hat\\Theta_0}$ of the squared GARCH process, defined on the $(p+q)$-dimensional unit simplex. The Markov chain $\\tilde\\Theta_s = A_s\\tilde\\Theta_{s-1}/\\|A_s\\tilde\\Theta_{s-1}\\|$, weighted by $\\|A_s\\tilde\\Theta_{s-1}\\|^\\kappa$, has $H_{\\hat\\Theta_0}$ as its invariant distribution because the spectral measure satisfies the fixed-point equation (2.14) of the paper. Once $\\hat\\Theta_0$ is sampled, the tail chain follows by matrix multiplication $\\hat\\Theta_t = A_t\\cdots A_1\\hat\\Theta_0$, with $\\hat X_t^2 = \\hat R_0 \\hat\\vartheta_t^{(1)}$ and $\\hat R_0$ an independent Pareto variable with tail index $\\kappa$. The supporting identities are Theorem 3.2, $\\kappa$ solves $E[(\\lambda e^\\eta)^\\kappa] = 1$ with $\\lambda$ the largest eigenvalue of $A_t$ and $\\eta = -(1/\\kappa)\\ln E(\\lambda^\\kappa)$, and Theorem 3.4, $\\kappa = 1$ for IGARCH(p,q).","core_discovery":"The central claim is that one object controls the extremal behaviour of a GARCH(p,q) process — the spectral measure $H_{\\hat\\Theta_0}$ of the squared process in its stochastic recurrence equation representation — and that this object can be sampled. Algorithm 1 runs a Markov chain $\\tilde\\Theta_s = A_s\\tilde\\Theta_{s-1} / \\|A_s\\tilde\\Theta_{s-1}\\|$ with weights $\\|A_s\\tilde\\Theta_{s-1}\\|^\\kappa$; its invariant distribution is $H_{\\hat\\Theta_0}$, and $\\kappa$ is found by solving $E\\|A\\hat\\Theta_0\\|^\\kappa = 1$. Algorithm 2 then propagates $\\hat\\Theta_t = A_t\\cdots A_1\\hat\\Theta_0$ and forms the squared tail chain with $\\hat X_t^2 = \\hat R_0 \\hat\\vartheta_t^{(1)}$. The paper proves the numerically stable identities $\\gamma = E(\\ln \\lambda) + \\eta$ and $E[(\\lambda e^\\eta)^\\kappa] = 1$, with $\\eta = -(1/\\kappa)\\ln E(\\lambda^\\kappa)$, and proves $\\kappa = 1$ for every IGARCH(p,q) process. From the squared tail chain, upper and lower tail chains of the original process are obtained by Bernoulli($\\delta$) sign thinning, yielding extremograms and extremal indices for symmetric and asymmetric innovations.","pith_inferences":["Inference: the Bernoulli-thinning step in Section 4 is asserted for asymmetric innovations; a direct simulation check of whether the sign of an extreme at lag $t$ is independent of the squared tail chain would either confirm or refute the computed extremal indices for skew-$t$ cases.","Inference: the paper's own observation that the $\\eta$-based route to $\\kappa$ is reliable only when $|\\phi - 1| > 0.05$, where $\\phi = \\sum\\alpha_i + \\sum\\beta_j$, means the particle filter, not the eigenvalue identity, is the load-bearing numerical component for near-integrated GARCH.","Inference: the Table 1 pattern — $\\kappa < 1$ for $\\phi > 1$ and $\\kappa > 1$ for $\\phi < 1$ — is illustrated for five models; proving monotonicity of $\\kappa$ in $\\phi$ for $\\max(p,q) \\ge 2$ would be a natural companion theorem.","Inference: the same spectral-measure sampler should transfer to other heavy-tailed stochastic recurrence equations that lack an eigenvalue shortcut, replacing MCMC schemes restricted to bounded innovations."],"forward_implications":["For any GARCH(p,q) with Gaussian, Student-$t$, or skew-$t$ innovations, the extremal index, extremogram, and cluster-size distribution can be computed numerically, including for IGARCH models.","The tail index $\\kappa$ can be evaluated without bounded-support assumptions and without the cubic slowdown of the previous rejection-based method, so high-order GARCH models become feasible.","Because every cluster functional is a functional of the forward tail chain, quantities such as mean cluster length, lag-$\\tau$ exceedance probabilities, and total cluster excess are obtained from the same simulated chains.","The same algorithms apply to the wider class of stochastic recurrence equations $Y_t = A_tY_{t-1} + B_t$ satisfying Kesten's conditions, giving strict-stationarity checks and extremal analysis for those processes as well.","The proof that $\\kappa = 1$ for all IGARCH(p,q) sharpens the moment boundary: $E|X_t|^{2-\\epsilon} < \\infty$ for every $\\epsilon > 0$ while the variance is infinite."],"supporting_citations":[{"why":"Establishes multivariate regular variation of GARCH processes and the product-form evolution of the angular tail vector.","marker":"Basrak et al. (2002)"},{"why":"Gives the fixed-point equations for the spectral measure that Algorithm 1 inverts.","marker":"Basrak and Segers (2009)"},{"why":"Erratum showing their earlier MCMC sampler for the spectral measure is not valid for general GARCH, motivating the new particle filter.","marker":"Basrak and Segers (2011)"},{"why":"Provides the random-matrix renewal theory that guarantees existence and uniqueness of the tail index kappa.","marker":"Kesten (1973)"},{"why":"Evaluates kappa for GARCH(1,1) through E[(alpha_1 Z^2 + beta_1)^kappa] = 1, the benchmark special case of the paper's identity.","marker":"Mikosch and Stărică (2000)"},{"why":"Develops the tail-chain and sign-thinning approach for ARCH(1), whose structure is generalized here.","marker":"de Haan et al. (1989)"},{"why":"Handles asymmetric innovations for GARCH(1,1), providing the comparison case for the Bernoulli(delta) construction.","marker":"Ehlert et al. (2015)"},{"why":"Previous method for kappa and the spectral measure, valid only for bounded innovations; the new algorithm improves on it.","marker":"Janssen (2010)"},{"why":"Supplies the SRE representation of GARCH and the top-Lyapunov-exponent stationarity condition used throughout.","marker":"Francq and Zakoïan (2010)"},{"why":"Derives the extremal index and spectral measure for GARCH(1,1), used as a validation benchmark.","marker":"Laurini and Tawn (2012)"}],"fun_headline_variants":["Sampling method yields extremal index for all GARCH(p,q)","Extremal index now computable for every GARCH(p,q)","New algorithm unlocks GARCH tail clustering for all cases","Spectral sampling solves GARCH extreme-event clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is the Section 4 representation that the sign of each extreme observation is an independent Bernoulli($\\delta$) variable independent of the squared tail chain; for asymmetric innovations this independence fails because the sign of $Z_t$ and $Z_t^2$ are correlated, so the computed extremal index and extremogram for asymmetric cases stand or fall with this thinning.","fun_headline_variants_meta":{"raw":{"variants":["Sampling method yields extremal index for all GARCH(p,q)","Extremal index now computable for every GARCH(p,q)","New algorithm unlocks GARCH tail clustering for all cases","Spectral sampling solves GARCH extreme-event clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1384,"prompt_tokens":1134,"completion_tokens":250,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":750,"completion_tokens_details":{"reasoning_tokens":182}},"tokens_in":750,"tokens_out":250,"duration_ms":2917,"temperature":1.0,"reasoning_tokens":182,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:37:36.100416+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a GARCH(2,2) with skew-$t$ innovations, simulate a long stationary series, estimate at a high threshold (e.g., the 0.9999 quantile) the lag-one upper-tail extremogram $\\chi_{X^U}(1)$ by the threshold method, and compare it with the paper's value $\\delta \\chi_{X^2}(1)$; a systematic gap beyond Monte Carlo error would show the Bernoulli-thinning representation of Section 4 does not hold for asymmetric innovations.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes multivariate regular variation of GARCH processes and the product-form evolution of the angular tail vector."},{"cited_title":"and Segers, J","cited_arxiv_id":null,"evidence_quote":"Gives the fixed-point equations for the spectral measure that Algorithm 1 inverts."},{"cited_title":"and Segers, J","cited_arxiv_id":null,"evidence_quote":"Erratum showing their earlier MCMC sampler for the spectral measure is not valid for general GARCH, motivating the new particle filter."},{"cited_title":"(1973) Random difference equations and renewal theory for products of random matrices","cited_arxiv_id":null,"evidence_quote":"Provides the random-matrix renewal theory that guarantees existence and uniqueness of the tail index kappa."},{"cited_title":"and St a ric a , C","cited_arxiv_id":null,"evidence_quote":"Evaluates kappa for GARCH(1,1) through E[(alpha_1 Z^2 + beta_1)^kappa] = 1, the benchmark special case of the paper's identity."},{"cited_title":"I., Rootz\\'en, H","cited_arxiv_id":null,"evidence_quote":"Develops the tail-chain and sign-thinning approach for ARCH(1), whose structure is generalized here."},{"cited_title":"and Schlather, M","cited_arxiv_id":null,"evidence_quote":"Handles asymmetric innovations for GARCH(1,1), providing the comparison case for the Bernoulli(delta) construction."},{"cited_title":"(2010) On Some Connections between Light Tails, Regular Variation and Extremes\\/","cited_arxiv_id":null,"evidence_quote":"Previous method for kappa and the spectral measure, valid only for bounded innovations; the new algorithm improves on it."},{"cited_title":"and Zako \\\"i an, J.-M","cited_arxiv_id":null,"evidence_quote":"Supplies the SRE representation of GARCH and the top-Lyapunov-exponent stationarity condition used throughout."},{"cited_title":"and Tawn, J","cited_arxiv_id":null,"evidence_quote":"Derives the extremal index and spectral measure for GARCH(1,1), used as a validation benchmark."}],"review_version":1}