Sandwiched Volterra Volatility Model

• Sandwiched Volterra Volatility (SVV) is a stochastic volatility framework that employs a Gaussian Volterra process with a singular drift to confine volatility between two deterministic barriers.
• The model integrates precise mathematical tools like Malliavin calculus and Markovian approximations to derive option pricing, implied volatility asymptotics, and quadratic hedging.
• SVV’s design ensures bounded volatility to prevent explosion or collapse, enabling explicit asset price representations and robust risk management in complex financial settings.

Sandwiched Volterra Volatility (SVV) model is a stochastic volatility framework in which the asset price is driven by Brownian noise while the volatility factor is driven by a Gaussian Volterra process and constrained by a singular drift to remain between two prescribed deterministic barriers. In the one-factor formulation, the log-price and volatility satisfy

$S(t)=e^{X(t)},$

$$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$

$$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$

with $B_1,B_2$ independent Brownian motions, $\rho\in(-1,1)$, and $\mathcal K$ a Volterra kernel (Nunno et al., 2023). The class was introduced as a stochastic volatility model driven by an arbitrary Hölder continuous Gaussian Volterra process whose volatility is “sandwiched” between two arbitrary Hölder continuous functions chosen in advance, and it has since been developed for option pricing, implied-volatility asymptotics, and quadratic hedging (Nunno et al., 2022, Nunno et al., 2022, Nunno et al., 2023).

1. Model specification and Volterra structure

The defining feature of SVV is that the instantaneous volatility is not postulated as an unconstrained diffusion or rough process, but as a process $Y$ obtained by adding a singular drift to a Gaussian Volterra noise. In the multi-asset version, for each asset $i=1,\dots,d$,

$$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$

$$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$

where $$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$0 and $$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$1 are correlated Brownian motions and $$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$2 is the Gaussian Volterra noise (Nunno et al., 2022).

The kernel assumptions are standard Volterra regularity conditions. The kernel is square integrable,

$$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$3

and satisfies a Hölder-type increment condition: there exists $$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$4 such that for every $$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$5,

$$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$6

or, in the formulation used for hedging,

$$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$7

These conditions imply that the Gaussian Volterra process

$$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$8

has a modification with Hölder continuous paths of any order $$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$9 (Nunno et al., 2023, Nunno et al., 2022).

This architecture places SVV within the broader Volterra volatility literature, but with an additional confinement mechanism. General volatility-modulated Lévy-driven Volterra processes are of the form

$$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$0

or, in stationary form,

$$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$1

with the kernel $$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$2 encoding memory and the factor $$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$3 encoding stochastic volatility or intermittency (Barndorff-Nielsen et al., 2012, Benth et al., 2016). SVV replaces the free volatility factor by a volatility state $$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$4 that is itself generated by a Volterra kernel and constrained by singular drift.

2. Sandwiching drift and boundary confinement

The term “sandwiched” refers to the requirement that volatility remain between two deterministic Hölder-continuous bounds,

$$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$5

with $$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$6 for all $$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$7 (Nunno et al., 2022). The drift is designed to explode near the boundaries. In the option-pricing formulation, the standing assumptions are

$$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$8

$$Y(t)=y_0+\int_0^t b(s,Y(s))\,ds+\int_0^t \mathcal K(t,s)\,dB_1(s),$$9

with

$B_1,B_2$0

In the one-factor power-law analysis, the analogous exponents satisfy

$B_1,B_2$1

ensuring that the singular repulsion is strong enough relative to the kernel roughness parameter $B_1,B_2$2 (Nunno et al., 2022, Nunno et al., 2023).

Under these assumptions, the solution remains trapped in the corridor almost surely: $B_1,B_2$3 A sharper estimate states that for any $B_1,B_2$4,

$B_1,B_2$5

where $B_1,B_2$6 is the Hölder modulus of the Volterra process (Nunno et al., 2023). In the hedging paper, the corresponding theorem yields existence of a unique strong solution together with explicit pathwise lower and upper margins from the barriers, and finite moments of inverse distances to both boundaries (Nunno et al., 2022).

This confinement is not a cosmetic addition. Because $B_1,B_2$7 is bounded above and below away from zero, the asset price admits the explicit stochastic exponential representation

$B_1,B_2$8

and the model satisfies

$B_1,B_2$9

(Nunno et al., 2022). A common misconception is therefore to treat SVV as merely another rough-volatility parametrization; the singular sandwiching drift is the structural device that prevents volatility from hitting zero or exploding.

3. Position within the Volterra stochastic calculus literature

SVV sits at the intersection of two earlier Volterra traditions: stochastic integration with respect to volatility-modulated Volterra processes, and infinite-dimensional representations of Volterra fields.

For Lévy-driven Volterra processes, the core object is

$\rho\in(-1,1)$0

and a Malliavin-calculus-based integration theory defines anticipative integrals $\rho\in(-1,1)$1 even when $\rho\in(-1,1)$2 is singular at the diagonal and the integrand is not adapted (Barndorff-Nielsen et al., 2012). In the Brownian case, the operator

$\rho\in(-1,1)$3

is central, and the integral is written as a Skorohod term plus a Malliavin-derivative correction (Barndorff-Nielsen et al., 2012). This is directly relevant to nested Volterra structures, because the same paper emphasizes that the ability to integrate random functionals of $\rho\in(-1,1)$4 against $\rho\in(-1,1)$5 is especially important for models where volatility is itself driven by a Volterra process.

In the white-noise framework, Brownian-driven Volterra integration is extended from $\rho\in(-1,1)$6-based calculus to the space $\rho\in(-1,1)$7 of Potthoff–Timpel distributions, and volatility can be inserted either by pointwise multiplication or by the Wick product,

$\rho\in(-1,1)$8

This provides a mathematically consistent formulation when the volatility is generalized and pointwise multiplication is not defined (Barndorff-Nielsen et al., 2013). Although that paper does not use the SVV name, it supplies analytic machinery for Volterra kernels with a stochastic factor “sandwiched” inside the integrand.

A different branch of the literature represents Volterra processes as boundary traces of hyperbolic SPDEs. For volatility modulated Volterra processes and Lévy semistationary processes,

$\rho\in(-1,1)$9

where $\mathcal K$0 solves a hyperbolic SPDE driven by the same noise and volatility structure (Benth et al., 2016). This suggests an infinite-dimensional state-space interpretation of layered Volterra volatility systems, especially when one seeks joint simulation of current values and forward curves.

4. Pricing, martingale measures, and Malliavin-based valuation

The pricing theory developed for SVV begins from the discounted asset

$\mathcal K$1

which satisfies

$\mathcal K$2

(Nunno et al., 2022). A central result is a full description of strict local martingale densities $\mathcal K$3 such that $\mathcal K$4 are local martingales. Choosing the თავისუფર directions $\mathcal K$5, Novikov’s condition applies because $\mathcal K$6 is bounded away from zero, so an equivalent martingale measure exists; allowing nontrivial $\mathcal K$7 yields many equivalent local martingale measures, hence market incompleteness (Nunno et al., 2022).

The same paper proves Malliavin differentiability of both volatility and price. For volatility,

$\mathcal K$8

and for the log-price $\mathcal K$9, an explicit Malliavin derivative formula is also obtained (Nunno et al., 2022).

These derivatives are used to price discontinuous claims by Malliavin integration by parts. The discontinuity is transferred from the payoff to a Malliavin weight, and explicit weighted formulas are derived in the one-dimensional case and for basket payoffs (Nunno et al., 2022). This is one of the model’s practical differentiators: the sandwiching bounds ensure that the weights involving $Y$0 remain integrable.

5. Short-maturity implied-volatility skew and power laws

A major theoretical development is the proof that SVV can reproduce the empirically observed short-maturity power law for the at-the-money implied-volatility skew. The target asymptotic is

$Y$1

for some $Y$2 (Nunno et al., 2023).

The analysis proceeds through Malliavin calculus. The first derivative of the volatility is

$Y$3

and the paper proves the key regularity statement that for any $Y$4 and $Y$5,

$Y$6

An explicit formula is then obtained for $Y$7, together with a generalized Malliavin product rule for singular random terms (Nunno et al., 2023).

The short-time skew constant is controlled by the leading small-time behavior of $Y$8. Under the kernel conditions

$Y$9

with

$i=1,\dots,d$0

and the small-time integral asymptotic

$i=1,\dots,d$1

together with $i=1,\dots,d$2 and $i=1,\dots,d$3, the main theorem yields

$i=1,\dots,d$4

so that the SVV model reproduces the power law

$i=1,\dots,d$5

The mechanism is transparent: $i=1,\dots,d$6 higher-order terms, and a kernel singularity of order $i=1,\dots,d$7 induces the same rough-volatility skew exponent (Nunno et al., 2023).

6. Markovian approximation, quadratic hedging, and computation

Because SVV is typically non-Markovian, direct computation of conditional expectations needed for hedging is difficult. The proposed remedy is a finite-dimensional Markovian approximation obtained by replacing $i=1,\dots,d$8 with degenerate kernels

$i=1,\dots,d$9

which yields

$$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$0

The approximated state vector

$$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$1

is finite-dimensional and Markovian (Nunno et al., 2022).

Two approximation families are emphasized. For Hölder convolution kernels, Bernstein polynomials give

$$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$2

For the rough fractional kernel

$$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$3

finite sums of exponentials yield

$$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$4

These kernel errors propagate to volatility and prices. In particular, for the discounted prices,

$$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$5

with the right-hand side measured either in $$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$6 or, for convolution kernels, in $$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$7 (Nunno et al., 2022).

Quadratic hedging is formulated as

$$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$8

The optimal strategy is given by the non-anticipating derivative. For a partition $$S_i(t)=S_i(0)+\int_0^t \mu_i(s)S_i(s)\,ds+\int_0^t Y_i(s)S_i(s)\,dB_i^S(s),$$9,

$$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$0

and

$$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$1

For a European payoff $$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$2,

$$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$3

The Markovian approximation makes these conditional expectations state-dependent functions of $$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$4, enabling nested Monte Carlo and least-squares Monte Carlo implementations (Nunno et al., 2022).

The convergence theory extends to hedging. If $$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$5, then for general payoffs $$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$6, with $$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$7 globally Lipschitz and $$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$8 of bounded variation,

$$Y_i(t)=Y_i(0)+\int_0^t b_i(s,Y_i(s))\,ds+\int_0^t \mathcal K_i(t,s)\,dB_i^Y(s),$$9

while for globally Lipschitz payoffs the exponent $$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$00 is replaced by $$X(t)=x_0+rt-\frac12\int_0^t Y^2(s)\,ds +\int_0^t Y(s)\bigl(\rho\,dB_1(s)+\sqrt{1-\rho^2}\,dB_2(s)\bigr),$$01 (Nunno et al., 2022). This establishes a direct chain from kernel approximation error to hedging error.

In this sense, SVV is simultaneously a model class and a methodological program: a confined Volterra volatility dynamics, a Malliavin-calculus-based pricing theory, a rough-kernel explanation of power-law skews, and a Markovian approximation scheme that restores computational tractability without discarding the underlying non-Markovian structure.