Signature Volatility Model

• Signature Volatility Model is a stochastic volatility framework where instantaneous volatility is defined as a linear functional of a time-extended path's signature.
• The model leverages tensor algebra and iterated integrals to achieve tractable calibration, pricing, and hedging despite its non-Markovian nature.
• It bridges classical models and modern deep architectures by ensuring universality, explicit Fourier pricing, and robust arbitrage conditions.

Signature volatility models are stochastic-volatility models in which the instantaneous volatility is represented as a linear functional of the signature of a time-extended driving path, most commonly a Brownian motion or a more general primary continuous semimartingale. In the standard Brownian formulation, one sets $\widehat W_t=(t,W_t)$ and defines volatility through iterated integrals of $\widehat W$; in broader formulations, the same construction is applied to a primary process $\widehat X_t=(t,X_t)$ or to an augmented process that also carries the stock-price driver. The class is explicitly non-Markovian at the path level, yet it remains linearly tractable in signature coordinates, which is why the literature emphasizes universality, calibration, Fourier methods, and structural arbitrage conditions rather than a single parametric volatility factor (Jaber et al., 21 Mar 2025, Jaber et al., 2024, Cuchiero et al., 2022).

1. Mathematical formulation

The canonical signature volatility model starts from two Brownian motions $(B,W)$ with

$$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$

and the time-extended path

$\widehat W_t=(t,W_t)\in\mathbb R^2.$

Its Stratonovich signature is

$\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$

with

$\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$

Given a finite linear functional

$$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$

the instantaneous volatility is

$\sigma_t=\bigl\langle \ell,\Sig(\widehat W)_{0,t}\bigr\rangle,$

and under the risk-neutral measure the asset price satisfies

$\widehat W$0

with $\widehat W$1 and $\widehat W$2 independent of $\widehat W$3 (Jaber et al., 21 Mar 2025).

A broader formulation replaces the Brownian driver by a general primary process $\widehat W$4 and its time-extension $\widehat W$5. Then a truncated linear signature model takes the form

$\widehat W$6

or, in the martingale representation relevant for one-asset pricing,

$\widehat W$7

This makes the instantaneous volatility exactly a linear function of the signature (Cuchiero et al., 2022).

The Sig-SDE literature places the same idea inside a controlled SDE: $\widehat W$8 with a truncated signature state $\widehat W$9. In that setting, the volatility can be parameterized as a linear pre-activation of signature coordinates followed by a positive link function such as $\widehat X_t=(t,X_t)$0 or $\widehat X_t=(t,X_t)$1 (Arribas et al., 2020).

2. Tensor-algebra structure and universality

The model’s algebraic backbone is the extended tensor algebra and the shuffle product. In the Brownian formulation of Abi Jaber and Gérard, the signature lives in the extended tensor algebra $\widehat X_t=(t,X_t)$2, with coordinates indexed by words over the alphabet $\widehat X_t=(t,X_t)$3. For $\widehat X_t=(t,X_t)$4 one writes

$\widehat X_t=(t,X_t)$5

whenever the series converges, and Itô’s formula can be written directly in signature coordinates: $\widehat X_t=(t,X_t)$6 This algebraic representation is the source of the model’s tractability in pricing and hedging (Jaber et al., 2024).

The universality claims in the literature are concrete. Signature volatility models are described as a “remarkably universal” class that includes, but is not limited to, Stein-Stein, Bergomi, and Heston models, together with path-dependent variants such as delayed equations and Volterra models (Jaber et al., 2024). In the 2022 calibration framework, the class is called universal in the sense that classical models can be approximated arbitrarily well and the parameters can be learned from all sources of available data by simple methods (Cuchiero et al., 2022). In the rough-path calibration study, the linear signature expansion is justified by the Universal Approximation Theorem for rough paths, and the paper explicitly states that any continuous path-functional $\widehat X_t=(t,X_t)$7 can be uniformly approximated on compact sets by a linear functional of the truncated signature (Alòs et al., 31 Jul 2025).

Time augmentation is not merely cosmetic. One explicit statement is that augmenting $\widehat X_t=(t,X_t)$8 guarantees that the signature uniquely encodes both spatial path and its parametrization, preventing aliasing of differently timed paths (Alòs et al., 31 Jul 2025). A related fact used in joint SPX–VIX calibration is that when the primary process is a polynomial diffusion, its truncated signature is again a finite-dimensional polynomial diffusion, so conditional expectations reduce to matrix exponentials (Cuchiero et al., 2023).

3. Martingality, arbitrage, and moment explosions

A central issue is that the discounted price process is always a nonnegative local martingale, but it may fail to be a true martingale. For the finite-order Brownian signature volatility model with leading coefficient $\widehat X_t=(t,X_t)$9, the martingality criterion is

$(B,W)$0

In particular, when $(B,W)$1, one needs $(B,W)$2 and odd $(B,W)$3; equivalently,

$(B,W)$4

The proof proceeds by a Girsanov change of measure that reduces martingality to non-explosion of the “signature SDE”

$(B,W)$5

and by the asymptotic estimate

$(B,W)$6

for large $(B,W)$7 (Jaber et al., 21 Mar 2025).

Once martingality is established, higher moments are governed by the correlation boundary

$(B,W)$8

equivalently

$(B,W)$9

If $$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$0 then $$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$1 (Jaber et al., 21 Mar 2025).

Low-order examples illustrate the theorem sharply. In the first-order case $$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$2,

$$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$3

and $$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$4 is always a martingale, with no condition on $$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$5. In the cubic model $$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$6,

$$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$7

with $$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$8, one has martingality if and only if $$d\langle B,W\rangle_t=\rho\,dt,\qquad \rho\in[-1,1],\quad \bar\rho:=\sqrt{1-\rho^2},$$9, and moments of order $\widehat W_t=(t,W_t)\in\mathbb R^2.$0 are finite exactly when $\widehat W_t=(t,W_t)\in\mathbb R^2.$1 (Jaber et al., 21 Mar 2025).

The broader arbitrage discussion is more structural. In the 2022 signature-based model, absence of arbitrage is tied to the existence of an equivalent local martingale measure under which the drift and quadratic-variation terms in the signature representation disappear, leaving a pure stochastic integral representation (Cuchiero et al., 2022). In the weighted-tensor-algebra formulation of Xodarev, the stochastic exponential

$\widehat W_t=(t,W_t)\in\mathbb R^2.$2

is a true, uniformly integrable martingale under the exponential-integrability hypothesis H3, yielding NFLVR, while global solvability of the associated infinite-dimensional Riccati equation is separated from the question of transform non-explosion (Xodarev, 16 May 2026). This suggests that “arbitrage-free” and “transform tractable” are related but distinct structural requirements.

4. Pricing, transforms, and hedging

Pricing theory for signature volatility models is built on transform methods. In the infinite-dimensional Brownian-signature model, if a time-dependent tensor-valued Riccati equation admits a solution and the relevant exponential process is a true martingale, then one obtains the joint characteristic functional of the log-price and integrated variance. This yields Fourier inversion formulas for European and path-dependent options and supports quadratic hedging formulas through martingale representation (Jaber et al., 2024).

The pricing results extend beyond plain vanilla options. For European calls, Lewis’ formula is combined with the model characteristic function; for geometric Asian options, one inserts a time-dependent weight into the Riccati system; for quadratic hedging, one differentiates the Fourier integral under the integral sign and identifies the $\widehat W_t=(t,W_t)\in\mathbb R^2.$3 and $\widehat W_t=(t,W_t)\in\mathbb R^2.$4 processes in the martingale representation (Jaber et al., 2024). In the 2022 theory paper, a “sig-payoff” is any payoff that is linear in the signature of the price path, and its expectation reduces to expectations of finitely many signature coordinates, which are tractable when the primary process is polynomial (Cuchiero et al., 2022).

A distinct line of work develops SPX–VIX pricing. Cuchiero, Gazzani, Möller, and Svaluto-Ferro show that when volatility is parametrized by a linear functional $\widehat W_t=(t,W_t)\in\mathbb R^2.$5 of the signature of a polynomial diffusion, one can write

$\widehat W_t=(t,W_t)\in\mathbb R^2.$6

so $\widehat W_t=(t,W_t)\in\mathbb R^2.$7 is an explicit quadratic form in $\widehat W_t=(t,W_t)\in\mathbb R^2.$8, while $\widehat W_t=(t,W_t)\in\mathbb R^2.$9 and $\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$0 can both be represented as linear functionals of the signature of an augmented process. The calibration task then splits into an offline sampling stage and a standard optimization stage (Cuchiero et al., 2023).

The hedging literature uses the same transform structure. For Fourier-invertible claims, semi-explicit quadratic hedges can be derived in a signature volatility model (Jaber et al., 3 Aug 2025). Malliavin calculus has also been specialized to finite linear combinations of time-extended Brownian motion signatures, yielding explicit formulas for the Malliavin derivative, Clark–Ocone representation, integration by parts, and Greeks for path-dependent options under signature volatility models (Jaber et al., 24 Apr 2026).

5. Calibration methodologies and reported performance

Calibration exploits the linearity of volatility in signature coordinates, even though prices depend nonlinearly on the coefficients. Time-series calibration in the 2022 framework is a penalized least-squares problem on precomputed signature features, while implied-volatility-surface calibration is phrased as optimization over model prices computed from Monte Carlo samples of signature vectors (Cuchiero et al., 2022). In the Sig-SDE framework, risk-neutral calibration minimizes weighted squared pricing errors with $\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$1 regularization, and real-world calibration may proceed by Euler–Maruyama likelihood or by matching expected signature moments (Arribas et al., 2020).

The rough-path calibration study uses a linear signature model

$\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$2

and optimizes

$\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$3

typically with $\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$4 or $\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$5, Monte Carlo precomputation of truncated signatures, and L-BFGS-B for the coefficient update (Alòs et al., 31 Jul 2025).

The following reported results illustrate the range of calibration settings.

Study	Setup	Reported result
(Alòs et al., 31 Jul 2025)	Uncorrelated Heston, SIG $\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$6 vs ASV	RMSE in IV: $\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$7 for SIG, $\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$8 for ASV; calibration time: $\Sig(\widehat W)_{0,t}=\Bigl(1,\Sig^1(\widehat W)_{0,t},\Sig^2(\widehat W)_{0,t},\dots\Bigr),$9 min vs $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$0 s
(Cuchiero et al., 2022)	S&P 500 market data, $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$1, $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$2	worst-case error $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$3 bp initially; with weighted loss and/or $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$4, reduced to $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$5 bp
(Cuchiero et al., 2022)	Time-dependent parameters, truncate $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$6	uniform absolute error $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$7 bp even at the shortest maturity, $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$8 bp at ATM
(Cuchiero et al., 2023)	VIX only, $\Sig^n(\widehat W)_{0,t}=\int_{0<u_1<\cdots<u_n<t}\circ\,d\widehat W_{u_1}\otimes\cdots\otimes\circ\,d\widehat W_{u_n}.$9, 2-dimensional OU primary process	fits six maturities of VIX options to within bid–ask spreads, with relative futures-price errors $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$0

These results are heterogeneous rather than directly comparable, because the objectives, data sets, and model classes differ. What is consistent across the cited papers is the computational strategy: expensive signature sampling is typically pushed offline or precomputed pathwise, while the online stage is a lower-dimensional optimization in linear signature coefficients (Cuchiero et al., 2023, Cuchiero et al., 2022).

6. Extensions, deep architectures, and structural developments

The model class has expanded in two directions: richer representation spaces and sharper structural theory. The deep signature approach for non-Markovian stochastic volatility reformulates the asset dynamics as a rough stochastic differential equation and then represents the rough paths via linear or non-linear combinations of time-extended Brownian motion signatures. After truncation, the pair $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$1 becomes a finite-dimensional Markov diffusion, and the pricing error in the linear-signature case satisfies a factorial-decay bound of order

$$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$2

as $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$3 (Ma et al., 21 Aug 2025).

Machine-learning applications use signatures both as a model class and as a feature extractor. Sig-SDEs are presented as a framework that can be calibrated under both $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$4 and $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$5, simulate future market scenarios, and provide theoretical guarantees for existence, uniqueness, and Euler–Maruyama convergence under global Lipschitz assumptions (Arribas et al., 2020). In hedging experiments under non-Markovian stochastic volatility, signature features in feedforward neural networks are reported to outperform LSTMs in most cases, while shallow regression based on a calibrated signature volatility model is described as more accurate and stable than direct linear signature regression across different payoffs and volatility dynamics (Jaber et al., 3 Aug 2025).

The American-option literature has used signature methods in a different modular configuration. A 2025 hybrid framework combines a rolling R/S estimator for a time-varying Hurst parameter, gradient-boosted forecasting, regime switching between rough Bergomi and Heston simulation, Random Fourier Features for signature kernels, and a least-squares Monte Carlo primal–dual scheme. The reported implementation uses signature depth three and $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$6, and the paper states that $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$7 paths are feasible on a standard CPU (Shah, 10 Aug 2025).

Structural theory has also advanced. Xodarev’s 2026 treatment formulates the signature SDE on a weighted tensor algebra $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$8, proves global existence and uniqueness of strong solutions under hypotheses H1–H3, separates NFLVR from Riccati non-explosion, characterizes market completeness through the density of truncated signature spans in $$\ell\in\Bigl(\bigoplus_{n=0}^N(\mathbb R^2)^{\otimes n}\Bigr)^*,$$9, and derives a Galtchouk–Kunita–Watanabe hedging-error decomposition with residual controlled by a weighted signature tail bound (Xodarev, 16 May 2026). A common misconception is that truncation alone guarantees a viable asset-pricing model; the martingale criteria, exponential-integrability conditions, and transform solvability results show that admissibility depends on parity, sign conditions, and integrability, not merely on finite-dimensional approximation (Jaber et al., 21 Mar 2025, Xodarev, 16 May 2026). Another misconception is that signature volatility models are intrinsically deep-learning constructions; in the core finance literature they are primarily linear functionals on signatures, with deep architectures appearing later as one implementation layer rather than the defining feature (Cuchiero et al., 2022, Arribas et al., 2020).

Taken together, the literature presents signature volatility models as a unified pathwise framework: linear in signature coordinates, compatible with non-Markovian dynamics, able to recover classical stochastic-volatility families, and now supported by detailed results on martingality, moments, Fourier pricing, Malliavin Greeks, calibration, and completeness (Jaber et al., 2024, Jaber et al., 21 Mar 2025, Xodarev, 16 May 2026).