Time-Consistent Volatility Uncertainty from Mellin Moments of Volatility Growth

Risk models used for capital and hedging decisions under model uncertainty have to be time-consistent: a position judged acceptable today should not be judged unacceptable tomorrow unless new information has arrived. For families of possible market models, this holds exactly when the family is "stable under pasting." That means you can follow one model up to some random time, switch to another, and still be inside the family. In practice, though, the family is almost always written down by hand as a fixed volatility band $[\underline{\sigma},\overline{\sigma}]$, and the band is not estimated from data. A family built only from bounds is also very conservative, because an adversary choosing the model can sit at the top of the band forever.

We propose a different kind of uncertainty set. It constrains how much volatility can grow from one step to the next, through power moments of the growth factor, and it estimates those constraints from ordinary return data with the Mellin transform. The Mellin transform is the multiplicative counterpart of the Fourier transform: it turns products of independent random variables into products of functions. The point of the construction is that volatility compounds multiplicatively, pasting composes models step by step, and the Mellin transform turns products into products. As a result, the worst-case moments of future volatility have a closed form, they are valid at every horizon at once, and estimating them from returns is a textbook scale-mixture deconvolution problem whose noise distribution is known exactly.

Core Claim

Model log-variance as multiplicative: $V_t = V_{t-1} M_t$. Here $V_t$ is the latent variance in period $t$ and $M_t>0$ is the uncertain growth factor. Let the per-step ambiguity set $\mathcal{Q}_\Psi$ contain every law of $M$ satisfying $\mathbb{E}[M^u]\le \Psi(u)$ on a grid of real exponents $u$. Let $\mathcal{P}_\Psi$ be the set of all path laws whose conditional law of $M_t$, given the past, lies in $\mathcal{Q}_\Psi$. This is a "rectangular" construction: the adversary chooses the step-$t$ law separately after each history. It is therefore stable under pasting by design, and the induced worst-case expectation $\mathcal{E}$ satisfies the tower property studied in Superhedging and Dynamic Risk Measures under Volatility Uncertainty and Constructing sublinear expectations on path space.

Our first claim is a sublinear analogue of the Mellin convolution theorem:

$$\mathcal{E}\big[(V_n/V_0)^u\big] = \sup_{P\in\mathcal{P}_\Psi}\mathbb{E}_P\Big[\prod_{t=1}^n M_t^u\Big] = \Psi(u)^n .$$

The identity follows by peeling off one step at a time. Positivity lets the nonnegative past product factor out of each conditional supremum, and the remaining factor is the constant $\Psi(u)$. The identity has a sharp consequence: for Mellin moments of compounded quantities, an adaptive adversary gains nothing over an adversary who draws the growth factors independently and identically. For these quantities, time consistency costs nothing. It also gives a worst-case tail bound for future variance that holds simultaneously at every horizon $n$:

$$\sup_{P\in\mathcal{P}_\Psi} P\big(V_n/V_0 > k\big) \le \inf_{u>0} \Psi(u)^n k^{-u}.$$

The bound comes from Markov's inequality applied inside each model. The coverage consequence is the important one. If a single per-step confidence statement places the true conditional growth laws inside $\mathcal{Q}_\Psi$, then every horizon's bound is valid at once, with no extra multiple-testing correction across horizons. This differs from the band set of Nutz and Soner in two ways. It constrains growth rather than level, so it needs no positive lower volatility bound; that lower bound is the condition whose failure breaks strict comparison in On properties of solutions to Black–Scholes–Barenblatt equations. It is also strictly smaller than any band containing the same data, because putting all mass at the upper extreme violates the moment caps.

Our second claim is that $\Psi$ can be estimated from returns without observing volatility. Suppose returns are $X_j=\sqrt{V}Z_j$ with standard Gaussian $Z_j$, and volatility is roughly constant within a block of $m$ returns. Then the ratio of consecutive block realized variances is

$$Y_t = M_t W_t, \qquad W_t \sim F(m,m),$$

where $F(m,m)$ is the Fisher distribution, a law known in closed form. Because the Mellin transform factorizes over the product, the ratio $\widehat{\mathbb{E}}[Y^u]/\mathbb{E}[W^u]$ estimates $\mathbb{E}[M^u]$ for $|u|<m/2$, which is the strip where the Fisher moments exist. This is exactly the known-noise scale mixture of Statistical Inference for Scale Mixture Models via Mellin Transform Approach. That paper's main restriction, that the noise law must be known, is satisfied naturally here because the Gaussian structure supplies it. The asymptotic normality established for empirical Mellin estimators in Statistical Skorohod embedding problem: Optimality and asymptotic normality, which recovers a random time scale from Brownian observations, supplies the confidence band that turns the point estimate into $\Psi$. The block size $m$ is a real design knob. Larger blocks widen the usable range of exponents $u$ and shrink the noise, but they strain the assumption that volatility is constant within a block.

Research Program

The theoretical work has three parts.

First, we prove the factorization identity and the all-horizon coverage statement. We also characterize precisely when pasting is costly. For functionals that are not products, such as worst-case quantiles of summed returns, the adaptive adversary can beat the best i.i.d. adversary. We will compute this gap by dynamic programming. Each step of that program is a finite moment-constrained linear program over the law of $M$.

Second, we derive the estimator's joint behaviour on a grid of exponents. Consecutive ratios share blocks, so the ratios are 1-dependent, and the joint central limit theorem has to account for this. We will build a simultaneous band over the $u$-grid and state the rate. Under the regularity used by Belomestny, Morozova and Panov, this should be the parametric rate, because evaluating at real $u$ requires no inversion step.

Third, we examine the edges of what the data identify. Exponents beyond the strip edge $m/2$ cannot be identified from block ratios alone. This is the Mellin strip limitation flagged in the file. We will report it rather than extrapolate past it.

Full Mellin inversion along a vertical line in the complex plane is used only as a diagnostic, to reconstruct the density of $M$ and check that the estimated envelope is plausible. The sublinear operator itself is defined only at real exponents, because a supremum of complex-valued moments has no meaning. This is a deliberate narrowing of the "sublinear Mellin transform" gap in the file.

The software deliverable has two components. One is an estimator of $\Psi$ from return series, with confidence bands. The other is a dynamic-programming solver that computes worst-case variance quantiles and tail probabilities under $\mathcal{P}_\Psi$, together with the same quantities under competing uncertainty sets.

Baselines

The do-nothing reference is the plug-in i.i.d. model. It fits the law of $M$ by the Mellin point estimate, adds no ambiguity, and propagates the fitted law forward. It is time-consistent but should under-cover whenever the growth law drifts. Beating it means reaching nominal coverage at long horizons where it falls short.

The strongest existing competitor on this question is the volatility-band set of Superhedging and Dynamic Risk Measures under Volatility Uncertainty. In that set, volatility may follow any adapted path inside $[\underline{\sigma},\overline{\sigma}]$, and worst-case values solve the Black–Scholes–Barenblatt equation; we use a grid solver in one dimension. It is the strongest competitor because it is the canonical pasting-stable set and is already time-consistent. We will calibrate the band in two ways: from rolling extreme quantiles of block realized variance, and from the support of the Mellin-reconstructed density. Both versions are matched to the same coverage target and fit on the same estimation windows as our method. Beating it means the same out-of-sample coverage with a strictly smaller worst-case variance quantile, which translates into lower capital, at horizons of 5 to 60 blocks. We found no reported result on Lacuna for calibrated coverage of band-type volatility sets on any real dataset. On the estimation side, the strongest reported comparison is "60 to 1,000 times more accurate" for Mellin over log-Fourier deconvolution under two-point noise (Statistical Inference for Scale Mixture Models via Mellin Transform Approach).

The time-inconsistent competitor is a non-rectangular ambiguity set: all laws of the whole $n$-step growth path whose $n$-step Mellin moments lie within the confidence band of a horizon-specific estimate. This set can be less conservative than ours. We will measure it on the same windows for both conservatism and preference reversals. A preference reversal is a pair of positions ranked one way at time 0 and the opposite way after every possible step-1 outcome.

The ablation that isolates our key ingredient replaces Mellin deconvolution with naive empirical power moments of the raw ratios $Y_t$. Because $\mathbb{E}[W^u]>1$ for $u$ away from zero, this inflates $\Psi$. A second ablation replaces Mellin with log-Fourier deconvolution of $\log Y_t$. A last ablation keeps the deconvolved $\Psi$ but drops the confidence band, which isolates the contribution of the band.

All methods use identical windows, identical block size, and an identical coverage target. Every tunable choice, meaning band quantiles, ambiguity radius, and $u$-grid, gets the same number of validation evaluations.

Why not just…?

Why not just set the volatility band from realized variance and run the Barenblatt solver? This is the most serious alternative. It is simple, it is pasting-stable, and with high-frequency data realized variance needs no deconvolution at all. Our argument is not that it is wrong but that it is wasteful. A band lets the adversary hold the maximum level forever, so its worst-case variance quantile at horizon $n$ is the level bound and is insensitive to how rarely the extremes actually occur. A growth-moment set charges the adversary for extremes through the moment caps. The head-to-head coverage-matched capital comparison settles this. If the band achieves equal coverage at similar capital on the crypto data, the moment construction has no practical value there, and we will report that result.

Why not just take empirical power moments of realized-variance ratios, with no deconvolution? With hourly data aggregated into daily blocks, each ratio carries Fisher-distributed noise with sizeable spread. Raw moments therefore overstate volatility growth at every nonzero $u$, by the known factor $\mathbb{E}[W^u]$. Dividing by that factor is the entire Mellin step, so the deconvolution costs essentially nothing to apply. The ablation will show whether the bias matters at realistic block sizes. If $m$ is large enough that $\mathbb{E}[W^u]\approx 1$ over the useful range of $u$, the deconvolution is unnecessary, and we will say so.

Why not just take logarithms and use Fourier deconvolution? In the simulation study of Statistical Inference for Scale Mixture Models via Mellin Transform Approach, the log-Fourier route "often fail[ed] to converge" under two-point noise, while the Mellin estimator stayed stable at the $1/n$ rate. Regime-switching volatility makes the growth factor itself close to two-point. Here the discreteness sits in the target $M$ rather than the noise, so the cited evidence is suggestive rather than decisive. The log-Fourier ablation tests the question directly.

Why not just use a deep second-order backward stochastic solver? The Deep 2BSDE method of Machine Learning Approximation Algorithms for High-Dimensional Fully Nonlinear Partial Differential Equations and Second-order Backward Stochastic Differential Equations solves Barenblatt equations in up to 100 dimensions. It still takes the volatility band as given, so it answers how to compute under a set rather than which set to use. It is complementary: we use it only to check that our conclusions about band sets are unchanged on multi-asset baskets.

Proposed Experiments

The first experiment is a simulation built on the scale-mixture design of Belomestny, Morozova and Panov. The growth factor $M$ is drawn as two-point with values 1 and 2, or as Gamma(2,2), Beta(2,2), or Zeta. Returns are generated as Gaussian blocks with $m\in\{6,24,96\}$, over $n$ from 100 to 1000 blocks. We measure the estimation error of $\Psi$ on the $u$-grid, the realized coverage of the variance tail bound at horizons 1 to 60, and the gap between adaptive and i.i.d. adversaries. The factorization identity predicts that this gap is exactly zero for product moments and positive for quantiles of summed returns.

The second experiment is a stress test. Growth laws are drawn non-identically from a GARCH-like recursion, so that only the conditional-membership assumption can fail. This shows how gracefully coverage degrades when the truth is not i.i.d.

The real-data study uses the hourly log-returns of the 452 USDT pairs in CTBench: Cryptocurrency Time Series Generation Benchmark, aggregated into daily blocks. Each set is estimated on rolling windows. We evaluate coverage of future block realized variance against the worst-case quantile, the capital implied at fixed coverage, and the time-inconsistency rate of each set.

The decisive checks are three. First, the dynamic-programming solver must reproduce $\Psi(u)^n$ to numerical precision. Second, at matched coverage, the growth-moment set must require less capital than both calibrated band sets. Third, the non-rectangular set must show nonzero preference reversals while ours shows none. If the second check fails on real data, the contribution reduces to the structural identity and the estimator, and the claim that this set is practically better than a volatility band is withdrawn.