{"id":"6c111ab2-dd26-424d-a414-722beff27657","arxiv_id":"2505.08623","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The grey Bergomi model adds a non-Gaussian 'grey' Brownian driver to rough Bergomi, producing new pricing and asymptotic formulas, but its joint SPX/VIX calibration remains inaccurate and can collapse back to the log-normal case.","lead":"This paper introduces a new volatility model, grey Bergomi, that replaces the Gaussian driver in rough Bergomi with a non-Gaussian process to give VIX options a steeper smile. The authors derive fast pricing formulas for both S&P 500 and VIX options, but their own calibration tests show the model still fails to fit the two markets together cleanly.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Calibration evidence for the joint-calibration claim is built on T↓0 asymptotics at T=0.094 with no finite-T check; the only non-log-normal calibration (β=0.11) is post hoc and the SPX fit is poor.","rationale":"The mathematical construction is coherent and the semi-closed and asymptotic formulae appear to follow the established template of [43,44]. The load-bearing weakness is empirical: the abstract's central claim is that gBergomi is a viable joint SPX/VIX calibration model, and Section 5 is the only evidence for that. Both calibration steps use T↓0 asymptotics at a one-month horizon, with no finite-T validation; this is a regime mismatch between the proved statements and their application. The calibration outcome is also internally fragile: the unconstrained optimum gives β=1, the post-hoc grid choice β=0.11 is not a genuine joint optimum, and the final SPX fit is poor. This is not an outside-consensus disagreement; it is an internal validation gap. A finite-T Monte Carlo check of the asymptotic formulas at T=0.094, followed by an MC-based recalibration, would settle whether the reported parameters and the β=0.11 curve are artefacts. The reader's weakest assumption identifies exactly this bridge, and I agree. The verdict should remain CONDITIONAL rather than ACCEPT or REJECT: the model construction and pricing formulae are a plausible contribution, but the central empirical claim is not established without the missing finite-T validation.","tokens_in":21590,"tokens_out":10409,"duration_ms":108967,"concrete_test":"Implement Algorithms 3.1 and 3.3 (or a hybrid/Markovian scheme) to compute VIX and SPX implied volatilities by Monte Carlo at T=0.094 for the reported parameter sets, using the same smoothed market data. Compare the MC smiles with the asymptotic formulas of Propositions 4.2 and 4.6; a discrepancy larger than the market noise (e.g. more than 10% of the observed skew/curvature) at this maturity would invalidate the calibration. Then re-run the Section 5 objective with MC prices replacing the T↓0 limits; if the joint-optimum β remains 1, or if the best joint fit still misses the SPX skew, the 'viable suggestion' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that gBergomi is a viable joint SPX/VIX calibration model. The only empirical support is Section 5, which calibrates VIX ATM level/skew/curvature and SPX skew using the small-time limits in Propositions 4.2 and 4.6. These limits are proved as T→0, but the market data used have T=0.094 (about one month); no finite-T error bounds or Monte Carlo validation are provided. If the asymptotic approximations are inaccurate at T=0.094, then the reported optimum (H*,β*,η*,ρ*)=(0.054,1,0.468,-1) and the later grid values (H,β,η,ρ)=(0.015,0.11,0.42,-1) are not reliable. The tension is visible in the paper itself: the unconstrained fit returns β*=1, the log-normal rough Bergomi case, and the claimed non-log-normal regime is reached only by fixing η and grid-searching (H,β), after which the SPX fit is explicitly acknowledged to be poor. Remark 5.1 admits the low η may be an artefact of not using 1-day VIX options. Thus the bridge between the model and the Holy Grail claim is empirically unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the 'grey Bergomi' (gBergomi) model, in which the Gaussian driver of rough Bergomi is replaced by a generalized grey Brownian motion, adding a parameter β that makes the VIX marginals non-log-normal while retaining self-similarity and stationary increments. The authors derive the VIX representation, VIX futures bounds, Malliavin-derivative-based small-time asymptotics for both VIX and SPX implied volatilities, and numerical schemes based on truncated Cholesky and Markovian approximations. They then calibrate the model to SPX and VIX options data with maturity T=0.094 and claim that the model is a viable candidate for the joint SPX/VIX calibration problem. The mathematical framework and formulae are the main contribution; the empirical support for the headline claim is the weakest part of the manuscript.","tokens_in":21775,"tokens_out":7077,"duration_ms":73111,"significance":"If the model and its formulae are correct, this is a useful contribution to the rough-volatility literature: it provides a tractable non-log-normal generalization of rough Bergomi, with semi-closed VIX expressions, explicit VIX futures bounds, and Malliavin-based asymptotic smile formulas that reduce to known rough Bergomi results when β=1. The derivation of the VIX formula in Proposition 2.8 and the asymptotic formulas in Propositions 4.2 and 4.6 are consistent with the cited framework and appear algebraically sound. The calibration section, however, does not support the paper's central claim that gBergomi is a viable joint SPX/VIX calibration model: the unconstrained fit returns β*=1, the non-log-normal fit is obtained only after fixing parameters, and the SPX fit is acknowledged to be poor. The manuscript also contains the kernel of a finite-T validation problem, since small-time asymptotic limits are applied at a one-month maturity without checking their accuracy. These issues are fixable, but they currently prevent the empirical claim from being established.","major_comments":[{"comment":"The calibration is built on T↓0 asymptotic limits applied to market options with maturity T=0.094, but the paper provides no finite-T Monte Carlo comparison, error bound, or convergence-rate estimate for these asymptotics. The reported optima (H*,β*,η*,ρ*)=(0.054,1,0.468,-1) and the later grid values (H,β,η,ρ)=(0.015,0.11,0.42,-1) are obtained entirely from these limits, so the empirical support for the joint-calibration claim depends on an unvalidated bridge between the asymptotics and the data. Remark 5.1 itself acknowledges that the low fitted η may be an artefact of not calibrating to one-day VIX options, which reinforces the need for a finite-T check.","section":"Section 5, Propositions 4.2 and 4.6"},{"comment":"The abstract's claim that gBergomi 'breaks away from the log-Normal assumption' and is a viable joint SPX/VIX calibration model is not supported by the paper's own calibration results. The unconstrained optimization returns β*=1, which is exactly the log-normal rough Bergomi case, and the only non-log-normal calibration (β=0.11) is obtained by fixing η and grid-searching (H,β). After that procedure, Figure 8 shows a visibly poor SPX fit and the text states that the model 'does not achieve a particularly accurate fit, especially with respect to the short-dated skew'. The empirical evidence therefore does not establish the central claim; at most it shows that a generalized specification can produce a non-log-normal VIX smile when the parameters are constrained.","section":"Section 5 and Abstract"},{"comment":"The paper acknowledges that β and η play similar roles and that one can fit the SPX smile with several (β,η) combinations if the VIX smile calibration is ignored. This implies that β is not identified from SPX options alone, which weakens the paper's claim that the non-log-normal feature is a genuine and separately estimable advantage of gBergomi. The manuscript should either provide an identification analysis, for example through the joint VIX and SPX fit, or explicitly limit the claim to the VIX marginals rather than to the joint calibration problem.","section":"Remark 5.2"},{"comment":"The proof invokes 'reverse Fatou's lemma' to conclude that condition (iv), limsup_{T↓0} E[(sqrt(V_T/V_0)-1)^2]=0, follows from almost-sure continuity of V_t. Reverse Fatou requires a uniform integrability or domination argument, and none is supplied. Since condition (iv) is one of the hypotheses imported from [44, Proposition 5.1] that justify the SPX skew asymptotics used in the calibration, this is a load-bearing gap in the proof as written; please add the missing L^2 domination estimate or replace the argument with a direct bound.","section":"Section 4.2, proof of Proposition 4.6"}],"minor_comments":[{"comment":"Both panels of Figure 2 use the legend entry 'ggBergomi' where 'gBergomi' is intended; this typo appears twice.","section":"Figure 2"},{"comment":"The trapezoidal sum in step (iv) uses Q^2_{T,τ_j} but the grid {τ_j}_{j=0,...,N} is introduced only implicitly in step (i); the notation should be aligned so that the index range and the quadrature weights are unambiguous.","section":"Algorithm 3.1, step (iv)"},{"comment":"The correspondence between the generalized grey Brownian motion parameter α and the Hurst parameter H (α=2H) is stated only after the Riemann-Liouville replacement; stating it explicitly at the point of the replacement would improve readability.","section":"Section 2.2, around (2.9)"},{"comment":"The empirical normality tests are applied to VIX log-returns, while the model implication discussed in the introduction concerns the squared VIX being close to log-normal. The connection between these two objects should be stated explicitly so that the reader can see why the tests are evidence against the rBergomi VIX marginal assumption.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and I see no citation-practice problem: the two main templates, [43] and [44], are published papers, not circular references. The main issue is that the abstract's joint-calibration claim exceeds what Section 5 demonstrates, and the finite-T validation gap is real. I recommend major revision rather than rejection because the model and the asymptotic formulae are a potentially useful contribution and the empirical gap appears fixable with additional numerical validation and a more cautious framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the model is a real extension—replacing the fBm driver with a generalised grey Brownian motion effectively randomises the vol-of-vol and breaks log-normality of VIX while keeping a tractable rough structure. That part is genuinely new and worth engaging with. But the paper's own calibration results undermine the 'viable suggestion for the joint SPX/VIX calibration' claim, so I'd read it as a model-formula paper with a weak empirical tail.\n\nWhat it does well: the grey Bergomi specification in (2.9) is coherent, the VIX futures bounds and Malliavin derivative computations check out, and the small-time asymptotic formulas for VIX and SPX smiles are useful. The derivations followed the template of [43,44], and the authors are transparent about some limitations—they report β*=1 from the first calibration and admit the SPX fit is poor. That honesty counts.\n\nThe soft spots are empirical, not mathematical. The headline claim rests on Section 5, but the calibration applies T→0 asymptotics at T=0.094 with no finite-T error bounds or Monte Carlo validation. One month is not obviously small for these limits, and the paper doesn't show that the asymptotics are accurate there. The unconstrained fit returns β=1, returning to log-normal rough Bergomi; the non-log-normal β=0.11 is reached only via post hoc grid search, and the SPX fit remains visibly poor. So the joint-calibration claim is not supported by the evidence presented. The stress-test note is right about that.\n\nMinor points: the VIX data smoothing uses an arctan parametric form which could influence the curvature target, and the SPX data being from Yahoo Finance close to the 2024 US election adds noise. These are minor next to the missing finite-T validation.\n\nThe citation pattern is fine—reliance on [43,44] is expected, and those are legitimate published templates. The mathematical structure is sound on its own terms, even though the empirical delivery doesn't match the abstract.\n\nWho is this for? Researchers in rough volatility who want a non-log-normal alternative with closed-form or asymptotic pricing. They'll find the model and formulas useful, and may want to fix the calibration with proper finite-T pricing. I would send this to a serious referee—it deserves referee time—but with a clear request to temper the claims and add validation. If I worked on rough volatility, I'd cite the model construction and VIX asymptotics, though not the calibration results.","headline":"Genuine new model and useful asymptotics, but the calibration evidence does not support the 'Holy Grail' claim; the paper is stronger as a model-formula contribution than as an empirical demonstration.","tokens_in":22402,"tokens_out":2350,"would_cite":true,"duration_ms":26121,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G15","60G22","60H07","91G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The grey Bergomi model replaces rough Bergomi's fractional Brownian motion with generalised grey Brownian motion, relaxing the log-normal VIX assumption and supplying closed-form small-time smile formulas for joint SPX/VIX calibration.","keywords":["Rough volatility","grey Brownian motion","Mittag-Leffler function","M-Wright distribution","VIX options","SPX options","Malliavin calculus","asymptotic implied volatility"],"falsifier":"Simulate the calibrated gBergomi model, for example with $(H,\\beta,\\eta,\\rho)=(0.015,0.11,0.42,-1)$, at $T=0.094$ using the paper's own VIX and spot algorithms, compute VIX and SPX implied volatilities by Monte Carlo, and compare them with the limits from Propositions 4.2 and 4.6; if the gaps exceed the market bid-ask spread or the fitted skew and curvature differences, the reported parameters and the joint-calibration conclusion are artefacts of applying $T\\downarrow0$ limits at one-month maturity.","tokens_in":21285,"feed_emoji":"📈","tokens_out":10870,"duration_ms":93309,"temperature":0.7,"pith_summary":"This paper introduces the grey Bergomi (gBergomi) model, a tractable extension of rough Bergomi in which the fractional Brownian motion driving the variance is replaced by a generalised grey Brownian motion. Because conditioning on an independent M-Wright variable turns the grey noise into a fractional Brownian motion, the variance becomes a randomised volatility-of-volatility process and the log-normal constraint of rough Bergomi is relaxed, so VIX marginals can carry skew and heavier tails. The paper derives semi-closed expressions for the VIX and forward variance, small-time asymptotic formulas for the VIX and SPX implied-volatility smiles, and a calibration exercise on SPX and VIX options data. A reader should care because the joint calibration of SPX and VIX options is a long-standing open problem, and the model is designed to keep rough volatility's tractability while allowing upward-sloping VIX smiles.","feed_headline":"Rough Bergomi goes grey to fix the VIX smile","feed_subtitle":"Grey noise makes VIX marginals non-log-normal and yields closed-form asymptotics for SPX and VIX options.","key_machinery":"The central object is the generalised grey Brownian motion $B^{\\beta,\\alpha}$, a self-similar process with stationary increments whose finite-dimensional characteristic function is $E_\\beta(-\\tfrac12 u^\\top\\Sigma_\\alpha u)$, with $E_\\beta$ the Mittag-Leffler function and $\\Sigma_\\alpha$ the covariance matrix of a fractional Brownian motion with Hurst parameter $\\alpha/2$. Its load-bearing property is Lemma 2.3, $B^{\\beta,\\alpha}_t \\stackrel{d}{=} \\sqrt{Y_\\beta}B^{\\alpha/2}_t$ with $Y_\\beta$ an independent one-sided M-Wright random variable: this representation turns the grey model into a randomised volatility-of-volatility rough Bergomi model, makes the Malliavin derivatives computable from those of the underlying fBm, and converts VIX squared into the explicit integral $\\mathrm{VIX}_T^2 = \\int_T^{T+\\Delta}\\frac{\\xi_0(s)}{E_\\beta(bs^{2H})}\\zeta_T(s)E_\\beta(b(s-T)^{2H})ds$ with $\\zeta_T(s)=\\sum_{k\\ge0}\\frac{(\\eta c)^k}{k!}\\frac{\\Gamma(1+k/2)}{\\Gamma(1+\\beta k/2)}(V^T_s)^k$. This is what carries the paper's semi-closed pricing formulas and the small-time smile asymptotics.","core_discovery":"The central claim is that the system (2.9), with $dS_t/S_t = \\sqrt{V_t}(\\rho dB_t + \\sqrt{1-\\rho^2}dW_t)$ and $V_t = \\xi_0(t)\\,\\mathcal{E}^{\\lozenge}(\\eta\\sqrt{Y_\\beta}B^H_t)$, where $B^H$ is a Riemann-Liouville fractional Brownian motion and $Y_\\beta$ an independent M-Wright variable, preserves self-similarity and stationary increments while letting VIX marginals depart from log-normality. By Lemma 2.3 the grey noise admits the representation $B^{\\beta,\\alpha}_t \\stackrel{d}{=} \\sqrt{Y_\\beta}B^{\\alpha/2}_t$, so the model is exactly rough Bergomi with volatility of volatility randomised by $Y_\\beta$. The paper proves the stock price is a true martingale for $\\rho\\le 0$, derives explicit VIX and VIX-futures expressions and bounds, computes the Malliavin derivatives that drive the asymptotics, and obtains small-time at-the-money limits for the VIX level, skew and curvature (Proposition 4.2) and for the SPX skew (Proposition 4.6). It then calibrates to market data and reports that the ATM VIX moments alone yield $\\beta^*=1$, while a full-smile grid search gives $(H,\\beta,\\eta)=(0.015,0.11,2)$ for the VIX and $\\eta=0.42$, $\\rho=-1$ for the joint fit, which the authors describe as not particularly accurate, with the short-dated SPX skew the main miss.","pith_inferences":["Because $\\beta$ and $\\eta$ both scale the effective volatility of volatility, ATM level-skew-curvature data cannot separate them; a natural fix is to calibrate $\\beta$ from the full VIX smile or from tail statistics such as those in Appendix B.","The representation $B^{\\beta,\\alpha}_t=\\sqrt{Y_\\beta}B^{\\alpha/2}_t$ suggests a latent-factor reading: $Y_\\beta$ is a static volatility-of-volatility shock, so the model should be testable by comparing option-implied moments of VIX with the moment structure $\\Gamma(1+\\beta k/2)$ inherited from the M-Wright law.","The short-maturity SSR limit $H+\\tfrac32$ derived in Section 2.4 offers an out-of-sample check: if observed skew-stickiness at short maturities contradicts the calibrated $H$, the model cannot simultaneously match level and dynamics, pointing to the same tension seen in multifactor rough models."],"forward_implications":["The VIX smile under gBergomi can slope upward, because VIX squared is no longer a near-log-normal conditional expectation and the M-Wright mixing injects skewness and excess kurtosis into VIX marginals.","The small-time limits in Propositions 4.2 and 4.6 give a parameter recipe: VIX ATM level, skew and curvature fix $(H,\\beta,\\eta)$ through $J_1,J_2,J_3$, and the SPX ATM skew then fixes $\\rho$, allowing calibration without Monte Carlo.","Since the discounted stock price is a true martingale for $\\rho\\le 0$ and its running supremum has finite expectation, European option prices under (2.9) are arbitrage-free and numerically well behaved.","The paper's own calibration shows two regimes: ATM VIX moments alone return $\\beta^*=1$ (the log-normal case), while a full VIX-smile search gives a strongly non-Gaussian $\\beta=0.11$; the joint SPX/VIX fit is not yet accurate, with the short-dated SPX skew the main miss.","The VIX-futures upper and lower bounds in Proposition 2.10, together with the Markovian approximation in Appendix A, provide practical pricing shortcuts for longer maturities."],"supporting_citations":[{"why":"Introduces the rough Bergomi model whose log-normal VIX limitation this paper extends.","marker":"[11]"},{"why":"Supplies the VIX futures pricing framework and the truncated-Cholesky VIX simulation algorithm reused in Section 3.","marker":"[43]"},{"why":"Provides the Malliavin-calculus asymptotic machinery and small-time smile formulas adapted in Propositions 4.2 and 4.6.","marker":"[44]"},{"why":"States Proposition 3, the representation $B^{\\beta,\\alpha}_t=\\sqrt{Y_\\beta}B^{\\alpha/2}_t$ that becomes Lemma 2.3.","marker":"[50]"},{"why":"Is the Malliavin calculus in finance reference behind the Clark-Ocone derivatives and implied-volatility asymptotics.","marker":"[6]"},{"why":"Motivates stochastic volatility of volatility as a way to reconcile SPX and VIX options, the modelling idea gBergomi makes concrete.","marker":"[29]"},{"why":"Documents that two-factor volatility models largely solve joint SPX/VIX calibration, motivating the additional factor in gBergomi.","marker":"[54]"},{"why":"The sum-of-lognormals approximation that makes rough Bergomi's VIX nearly log-normal, the assumption gBergomi is designed to break.","marker":"[27]"}],"fun_headline_variants":["Rough Bergomi with randomised vol-of-vol fits VIX smiles","Grey noise breaks log-normality in rough Bergomi VIX","Joint SPX/VIX calibration via grey Brownian motion","Non-log-normal VIX asymptotics from grey rough Bergomi"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the small-time asymptotic formulas (Propositions 4.2 and 4.6, derived as $T\\downarrow0$) are accurate enough at the one-month market maturity $T=0.094$ used in the calibration, since the paper provides no finite-maturity Monte Carlo validation of those formulas.","fun_headline_variants_meta":{"raw":{"variants":["Rough Bergomi with randomised vol-of-vol fits VIX smiles","Grey noise breaks log-normality in rough Bergomi VIX","Joint SPX/VIX calibration via grey Brownian motion","Non-log-normal VIX asymptotics from grey rough Bergomi"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000809,"raw_usage":{"total_tokens":3563,"prompt_tokens":974,"completion_tokens":2589,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":2527}},"tokens_in":590,"tokens_out":2589,"duration_ms":17946,"temperature":1.0,"reasoning_tokens":2527,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:50:43.235683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the calibrated gBergomi model, for example with $(H,\\beta,\\eta,\\rho)=(0.015,0.11,0.42,-1)$, at $T=0.094$ using the paper's own VIX and spot algorithms, compute VIX and SPX implied volatilities by Monte Carlo, and compare them with the limits from Propositions 4.2 and 4.6; if the gaps exceed the market bid-ask spread or the fitted skew and curvature differences, the reported parameters and the joint-calibration conclusion are artefacts of applying $T\\downarrow0$ limits at one-month maturity.","supporting_citations":[{"cited_title":"Bayer, P","cited_arxiv_id":null,"evidence_quote":"Introduces the rough Bergomi model whose log-normal VIX limitation this paper extends."},{"cited_title":"Jacquier, C","cited_arxiv_id":null,"evidence_quote":"Supplies the VIX futures pricing framework and the truncated-Cholesky VIX simulation algorithm reused in Section 3."},{"cited_title":"Jacquier, A","cited_arxiv_id":null,"evidence_quote":"Provides the Malliavin-calculus asymptotic machinery and small-time smile formulas adapted in Propositions 4.2 and 4.6."},{"cited_title":"Mura and G","cited_arxiv_id":null,"evidence_quote":"States Proposition 3, the representation $B^{\\beta,\\alpha}_t=\\sqrt{Y_\\beta}B^{\\alpha/2}_t$ that becomes Lemma 2.3."},{"cited_title":"Al`os and D","cited_arxiv_id":null,"evidence_quote":"Is the Malliavin calculus in finance reference behind the Clark-Ocone derivatives and implied-volatility asymptotics."},{"cited_title":"Fouque and Y","cited_arxiv_id":null,"evidence_quote":"Motivates stochastic volatility of volatility as a way to reconcile SPX and VIX options, the modelling idea gBergomi makes concrete."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents that two-factor volatility models largely solve joint SPX/VIX calibration, motivating the additional factor in gBergomi."},{"cited_title":"Fenton, The sum of log-Normal probability distributions in scatter transmission systems , IEEE Transactions on Communications, 8 (1960), pp","cited_arxiv_id":null,"evidence_quote":"The sum-of-lognormals approximation that makes rough Bergomi's VIX nearly log-normal, the assumption gBergomi is designed to break."}],"review_version":1}