{"id":"f72d861c-4b30-45b5-97d1-b4f8d02b4d58","arxiv_id":"2608.03925","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A gamma time-changed fractional Brownian motion is a semimartingale, and the resulting five-parameter fractional Variance Gamma model fits S&P 500 return moments with H≈0.45.","lead":"The authors build a fractional Brownian motion process run on a random 'trading time' clock, making it usable in standard no-arbitrage option pricing. Applied to S&P 500 returns, it estimates a Hurst exponent near 0.45, slightly below the random-walk benchmark.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 4.3's true-martingale proof relies on the false identity A(t)=E[X(t)|F(t−)]; the claimed martingale property is therefore unproven as written.","rationale":"The reader's weakest assumption concerns non-adaptivity of the compensator to the observable filtration. That is a real limitation of implementation, but it does not necessarily invalidate the existence of an EMM: an EMM constructed on the enlarged filtration restricts to an equivalent measure on the observable filtration, and finite variation already makes X a semimartingale in its own filtration. The more load-bearing issue is an internal proof error: Appendix A.3 equates the compensator with the conditional expectation E[X(t)|F(t−)], which is false for pure-jump processes. This directly affects Proposition 4.3, which is used to identify the drift and to justify the martingale decomposition underlying the no-arbitrage and Girsanov results. Because the paper's central contribution is the semimartingale-plus-compensator construction, an unsupported true-martingale claim is a genuine correctness risk, though likely repairable. I therefore keep the reader's CONDITIONAL verdict unchanged, while noting that the specific concern differs from the one emphasized by the reader.","tokens_in":29422,"tokens_out":19944,"duration_ms":236773,"concrete_test":"Independently re-prove Proposition 4.3 without invoking A(t)=E[X(t)|F(t−)]. Directly bound E[A(t)^2] using the explicit form a_X(s)=∫ m_B(g;γ(s))ψ_γ(g)dg and the Gaussian structure of fBm, e.g., by showing ∫_0^t E[a_X(s)^2]ds<∞. If this bound cannot be established without additional assumptions (such as square-integrability of the stochastic integral defining m_B), then the true-martingale claim requires new hypotheses and the current proof is incomplete; if it can, the proposition is salvageable and the central semimartingale claim remains intact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends not only on X being a semimartingale (which follows from finite variation) but on Proposition 4.3, which asserts that the explicitly constructed A_X is the compensator and that X−A_X is a true P-martingale. The proof in Appendix A.3 uses: 'Since A is the predictable compensator of X, we know that A(t)=E[X(t)|F(t−)]' and then bounds E[A(t)^2] by Jensen. This identity is false for pure-jump processes: for a Poisson process N with compensator λt, E[N(t)|F(t−)]=N(t−), not λt. The compensator is a dual predictable projection, not the optional projection. Hence the square-integrability argument establishing that X−A_X is a true martingale does not go through. If only a local martingale is obtained, the Girsanov results in §4.3, the no-arbitrage sufficient conditions in Proposition 5.3, and the endogenous-drift decomposition in Eq. (25) are not fully supported as stated. This is a sharper, internal proof gap than the non-adaptivity of A_X to the observable filtration, which the paper acknowledges; the present issue is an incorrect step in a central proof.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a time-changed fractional Brownian motion X(t)=B_H(γ(t)), where γ is an independent gamma activity time, and claims that this process is a semimartingale while retaining fBm-like roughness, long memory, and anomalous diffusion. It develops a marked-point-process representation, derives an explicit compensator A_X, states a Girsanov theorem, and uses these to formulate no-arbitrage drift conditions and an EMM existence criterion. On this basis it constructs the fractional Variance Gamma (fVG) model S(t)=S(0)exp{Ξ(t)+θγ(t)+σB_H(γ(t))}, obtains closed-form unconditional moment conditions, proposes a two-step feasible GMM estimator, and illustrates the procedure on ten years of daily S&P 500 returns, reporting H≈0.45.","tokens_in":29759,"tokens_out":11586,"duration_ms":133737,"significance":"If the technical gaps are repaired, the paper would be a useful contribution: it offers a way to keep fractional features in a tradable semimartingale price process, extends the VG model with a Hurst exponent, and provides closed-form moment conditions plus a simulation-based estimation and valuation strategy. Theorem 4.2 is plausible and rests on a bounded-variation argument for the composed process, and the moment formulas (Propositions 4.1 and 4.5, Corollary 4.6) are concrete and potentially useful. The paper is also honest that the conditional law and the pricing-performance evaluation are not yet fully developed. However, several load-bearing parts of the theoretical apparatus are not correctly established as written, and the acknowledged non-adaptivity of the compensator to the observable price filtration leaves the central claim of compatibility with the classical arbitrage-free pricing framework incomplete.","major_comments":[{"comment":"The compensator density is defined as ψ_X(t,x)=∫ φ(x;γ(t),g)ψ_γ(g)dg, with m_B(g;γ(t))=E[B_H(γ(t)+g)−B_H(γ(t))|F(t−)]=∫_0^{γ(t)} K_H(γ(t)+g,u)dB(u). For a jump of the gamma activity time at t, the pre-jump level is γ(t−), and the jump size is B_H(γ(t−)+g)−B_H(γ(t−)). The conditional distribution given F(t−) must be based on γ(t−), not on γ(t), and the integral for m_B should run to γ(t−) with kernel K_H(γ(t−)+g,u). As written, the integrand is not F(t−)-measurable and the displayed object is not a predictable compensator density. This affects Proposition 4.3 and everything downstream that uses ψ_X. Please correct the notation and derivation, or explain why γ(t) is intended to mean γ(t−) throughout.","section":"Section 4.2, Eq. (14) and definitions of m_B, σ_B"},{"comment":"The proof that X−A is a true P-martingale uses the assertion 'Since A is the predictable compensator of X, A(t)=E[X(t)|F(t−)]'. This identity is false for pure-jump processes: for a Poisson process N with compensator βt, E[N(t)|F(t−)]=N(t−), not βt. The compensator is a dual predictable projection, not an optional projection. Consequently the Jensen bound on E[A(t)^2] does not establish square-integrability of A, and the argument elevating the local martingale to a true martingale is invalid. The true-martingale property is load-bearing for Corollary 4.4, Proposition 5.2/5.3, and the decomposition in Eq. (25). The statement may be salvageable by instead proving E[∫ x^2 ν_X(ds,dx)]<∞, using the gamma Lévy density and the Hölder continuity of fBm paths, but that proof is not what is in the manuscript.","section":"Appendix A.3 / Proposition 4.3"},{"comment":"The paper acknowledges that A_X is not adapted to the natural filtration F^X generated by X, because it depends on the latent activity time γ and on conditional moments m_B(g;γ(s)). Yet the EMM construction and drift conditions are then stated for this A_X. If the filtration is enlarged to include γ, the discounted stock price is a semimartingale relative to F, but admissible trading strategies in the classical no-arbitrage framework are predictable with respect to the observable price filtration; the drift in Eq. (22) cannot be evaluated from observed prices, and Eq. (25) is not operational. The central claim of compatibility with classical observable-market arbitrage-free pricing therefore remains incomplete. The manuscript should either compute the compensator under F^X, or explicitly present the framework as a latent-factor/partially observed model and state the resulting pricing and","section":"Section 4.3, final paragraph; Propositions 5.2–5.3"},{"comment":"Condition 4 states the arbitrage-free drift restriction as r(t)=ξ(t)+∫(e^x−1)λ(t,x)ψ_X(t,x)dx. Proposition 5.2 and Eq. (22) correctly use ψ_Y, the compensator density of the innovation Y. The fVG specification used in the paper is Y=W=θγ+σX, whose jump compensator includes the θγ component and is not ψ_X. As stated, the sufficient condition in Proposition 5.3 is not the correct condition for the fVG model unless Y is taken to be exactly X. Please replace ψ_X by ψ_Y throughout the proposition and either derive ψ_W for the W specification or clarify that the proposition applies only to the generic pure-jump innovation Y of Section 5.1.","section":"Proposition 5.3, condition 4"}],"minor_comments":[{"comment":"The gamma argument in the raw moment formula contains an undefined symbol: Γ(h/v + m − 2k(1−H)) should almost certainly be Γ(h/v + n − 2k(1−H)). Please correct.","section":"Proposition 4.5, Eq. (16)"},{"comment":"The jump measure is written as Σ_{s: ΔY(s)=1} δ_{(s,1)}, but in the example the process is N. Replace ΔY with ΔN.","section":"Example 3.1"},{"comment":"The claim that deviations from H=0.5 are 'significant and economically meaningful' is not supported by any standard errors, confidence intervals, or specification test. Given the instability of H across p and across restricted specifications (H ranges from 0.35 to 0.48), please present the empirical section as descriptive and qualify the abstract's 'estimated H≈0.45' accordingly.","section":"Section 7, Table 1"},{"comment":"The phrase 'retains the defining properties of fBm' is too strong: the time-changed process is not self-similar and is not Gaussian. It would be more precise to say it preserves selected distributional features such as roughness, moment-scaling behavior, and long-range dependence in the relevant parameter regime.","section":"Abstract and Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of q-fin.MF and the main idea is attractive, but the compensator/martingale arguments need substantive repair before the central claim is supported. In particular, the γ(t) versus γ(t−) issue in Eq. (14) and the false optional-projection identity in the proof of Proposition 4.3 should be addressed head-on. The empirical section should be framed as an illustration only; it is not yet a formal test of H≠0.5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What is genuinely new here is the packaging: taking the already-known time-changed fBm B_H(gamma), proving it is a semimartingale via bounded variation, and then building the whole apparatus—compensator, Girsanov, EMM drift conditions, marked-point representation, and a GMM estimation recipe—around it. That financial wrapper is absent from the statistical papers the authors cite, so the extension is real. The moment formulas in Section 4 are clean and potentially useful, and the paper is honest about what it does not do: no option price validation, no standard errors on the GMM estimates, and a Hurst exponent that wanders between 0.35 and 0.48 depending on specification. That level of candor earns credit.\n\nThe soft spot is more serious than the paper lets on. In the proof of Proposition 4.3 (Appendix A.3), the authors assert that because A is the predictable compensator of X, A(t) = E[X(t)|F(t-)] on pure-jump processes. That is false for any process with predictable jump times, and even for a Poisson process the identity fails—the left side is deterministic, the right side is N(t-) plus an infinitesimal. The square-integrability bound on E[A^2] therefore does not go through, and X-A is only established as a local martingale, not a true martingale. The Girsanov corollary, the sufficient conditions in Proposition 5.3, and the drift decomposition in (25) all lean on the true-martingale status. This is a load-bearing gap, but it looks repairable: the finite-variation structure may allow a direct L^2 bound, or one can cite existing results on compensators of subordinated processes.\n\nSeparately, Section 4.3 concedes that the compensator is not adapted to the natural filtration of X, so the EMM conditions are not directly implementable from observed prices. That is a known limitation of latent-activity-time models, but it undercuts the claim that the framework is operationally compatible with classical observable-market pricing. The empirical exercise is illustrative only, which the authors themselves state.\n\nBottom line: the semimartingale construction is sound, the model is economically motivated, and the authors flag the right caveats. But the central martingale proof has a concrete error, and the non-adaptivity issue needs a cleaner resolution before the arbitrage-free machinery is fully trustworthy. Worth sending to referees—they can push on the proof and the operational gap. I would not cite it as-is, but I would put it on a reading list for fractional-model people.","headline":"The fVG construction is plausible and the semimartingale claim is likely true, but a central proof uses a false identity for pure-jump compensators, leaving the true-martingale result unproven as written.","tokens_in":30218,"tokens_out":2543,"would_cite":false,"duration_ms":30588,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G22","60G51","60G55","91G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that evaluating fractional Brownian motion at a gamma-distributed activity clock makes it a semimartingale, opening classical no-arbitrage option pricing to fractional dynamics.","keywords":["fractional Brownian motion","time-changed process","gamma activity time","semimartingale","variance gamma model","option pricing","GMM estimation","Hurst exponent"],"falsifier":"Simulate the fVG model with known parameters and run the proposed GMM on the observed price path alone: if the moment conditions leave a flat ridge in (H, v) or the estimates do not converge to the true values as the sample grows, the latent activity time cannot be recovered from prices, and the Proposition 5.2 drift condition cannot be implemented in the natural filtration — the compatibility claim fails at the observable-market level.","tokens_in":29341,"feed_emoji":"⏱️","tokens_out":13295,"duration_ms":117260,"temperature":0.7,"pith_summary":"Fractional Brownian motion (fBm) has the features stock returns seem to show — long-range dependence, rough paths, anomalous diffusion — but it is not a semimartingale, and under continuous trading that creates arbitrage, ruling it out of the standard option-pricing framework. The paper claims the obstacle disappears if fBm is evaluated not at calendar time but at a gamma-distributed stochastic clock called activity time: the composition X = B_H(γ) is a pure-jump semimartingale, while retaining fBm's defining statistical properties. On this construction the paper builds the fractional Variance Gamma (fVG) model, derives the compensator that acts as the process's predictable drift, obtains a Girsanov change of measure for risk-neutral pricing, and proposes a GMM estimation procedure. Fitting the model to S&P 500 index returns yields an estimated Hurst exponent of about 0.45, mildly below the Brownian value 0.5, which the authors read as mildly sublinear scaling of return moments. If the semimartingale claim is right, fractional persistence and roughness become compatible with classical arbitrage-free valuation without altering trading restrictions or the definition of arbitrage.","feed_headline":"A gamma clock makes fractional Brownian motion a semimartingale","feed_subtitle":"The fractional Variance Gamma model keeps long-memory and rough paths inside arbitrage-free pricing.","key_machinery":"The load-bearing object is the time-changed process X = B_H(γ), fBm run on a gamma 'activity-time' clock. Because the gamma process is an infinite-activity pure-jump subordinator of bounded variation and fBm has continuous paths, the composition is a pure-jump process of bounded variation — the property that makes it a semimartingale. The argument then runs through the marked point process compensator ν_X(dt, dx) = ψ_X(t,x) dx dt, whose density is a convolution of the conditional Gaussian density of fBm increments with the gamma Lévy density; integrating x against ψ_X gives the explicit predictable compensator A_X that plays the role of the drift, and a Girsanov kernel λ(t,x) yields the equi","core_discovery":"The central discovery is Theorem 4.2: X(t) = B_H(γ(t)), fBm evaluated at an independent gamma process, is a semimartingale. Because γ is a pure-jump process of bounded variation and B_H is continuous, X is itself a pure-jump bounded-variation process, hence a special semimartingale with a unique predictable compensator A_X. The paper computes A_X as a Gaussian–Gamma mixture, derives a Girsanov density yielding equivalent martingale measures, and builds the fVG stock-price model, whose log returns split into risk-free accrual, jump-risk premium, dependence-induced endogenous drift, and martingale innovation. Fitting the model to S&P 500 returns by GMM gives H ≈ 0.45, read as mildly sublinear","pith_inferences":["The construction is portable: any continuous Gaussian process run on a finite-variation subordinator becomes a pure-jump semimartingale with the same compensator machinery, so the result plausibly extends to multifractional or fractional Ornstein–Uhlenbeck drivers, not just fBm.","The paper's own Section 4.3 concedes that the compensator A_X is not adapted to the natural filtration of observed prices, because the activity time is latent; closing that gap would require a filtering or nonparametric method that recovers γ (or its conditional law) from the path of X — without it, the EMM drift condition is not directly implementable from market data.","A testable reading of H ≈ 0.45: compute separate tail-index estimates on the same S&P 500 returns and check whether correcting for heavy tails genuinely moves the Hurst estimate toward the Brownian benchmark — currently that attribution rests on in-sample model comparison.","Because the paper leaves option pricing to future research, the nearest decisive test is whether the fVG's implied volatility surface — smile persistence across maturity generated by H < 1/2 and mixture kurtosis — is observed; a mismatch would localize which of the five parameters must absorb the misspecification."],"forward_implications":["fBm's long-range dependence (H > 1/2), roughness (small H), and anomalous diffusion can be embedded in a stock price process that still satisfies the fundamental theorem of asset pricing — no restricted trading strategies, transaction costs, or redefined arbitrage needed.","Under the fVG model, log returns decompose into risk-free accrual, a jump-risk premium, an endogenous drift from the dependence structure, and a martingale innovation; persistence therefore affects expected returns, with H > 1/2 implying positively autocorrelated returns and H < 1/2 anti-persistent dynamics.","Risk-neutral option prices can be computed as expectations under the statistical measure using the simulated conditional distribution of X together with the Girsanov likelihood factor; if the market price of jump risk is independent of jump size, the pricing measure is fixed by the estimated physical dynamics alone.","Estimated on daily S&P 500 returns, the full fVG specification gives H ≈ 0.45, implying that the term structure of return moments scales mildly sublinearly; the authors argue that apparent long memory in simpler models may partly reflect unmodeled heavy tails.","The fVG and time-changed fractional geometric Brownian motion specifications are equivalent up to the jump-amplitude transformation x ↦ ln(1+x), so the valuation framework covers both formulations."],"supporting_citations":[{"why":"Supplies the bounded-variation property of gamma processes (Proposition 2) used to show B_H(γ) is a bounded-variation pure-jump process, the step on which Theorem 4.2 rests.","marker":"Yor (2007)"},{"why":"Provides the semimartingale criteria (Theorems 7 and 35) and the special-semimartingale decomposition used to prove X is a semimartingale with unique predictable compensator, plus the Itô formula behind the no-arbitrage drift condition.","marker":"Protter (2005)"},{"why":"Establishes the fundamental theorem of asset pricing requiring the semimartingale property — the obstacle that motivates the whole construction.","marker":"Delbaen & Schachermayer (1994)"},{"why":"Constructs explicit arbitrage strategies for the fractional Black–Scholes model, defining the failure mode the time-changed process must avoid.","marker":"Rogers (1997)"},{"why":"Shows the fractional BSM model admits arbitrage under continuous trading and supplies the restricted-strategy alternatives the paper contrasts with.","marker":"Cheridito (2003)"},{"why":"The Variance Gamma model and its gamma activity-time construction that the fVG extends; also the source of the skewness term θγ(t).","marker":"Madan et al. (1998)"},{"why":"Theorem 3.24 supplies the Girsanov theorem for pure jump processes used to build equivalent martingale measures for X.","marker":"Jacod and Shiryaev (2003)"},{"why":"The circulant-embedding method used to simulate fBm sample paths, on which the paper's simulation-based conditional distribution and pricing procedure depend.","marker":"Wood & Chan (1994)"}],"fun_headline_variants":["Gamma clock turns fractional Brownian motion into a semimartingale","fVG model: option pricing with rough paths and long memory","Time-changed fBm: a semimartingale for arbitrage-free pricing","Fractional Variance Gamma fits S&P 500 with Hurst 0.45"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the no-arbitrage and risk-neutral results hold in the enlarged filtration where the gamma activity clock is observable; from observed prices alone the compensator driving the drift condition cannot be computed, because the activity time is latent — a gap the paper states in Section 4.3 and leaves open.","fun_headline_variants_meta":{"raw":{"variants":["Gamma clock turns fractional Brownian motion into a semimartingale","fVG model: option pricing with rough paths and long memory","Time-changed fBm: a semimartingale for arbitrage-free pricing","Fractional Variance Gamma fits S&P 500 with Hurst 0.45"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1262,"prompt_tokens":703,"completion_tokens":559,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":478}},"tokens_in":447,"tokens_out":559,"duration_ms":5970,"temperature":1.0,"reasoning_tokens":478,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:36:57.927917+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the fVG model with known parameters and run the proposed GMM on the observed price path alone: if the moment conditions leave a flat ridge in (H, v) or the estimates do not converge to the true values as the sample grows, the latent activity time cannot be recovered from prices, and the Proposition 5.2 drift condition cannot be implemented in the natural filtration — the compatibility claim fails at the observable-market level.","supporting_citations":[],"review_version":1}