Signature-Pricing Framework

• Signature-Pricing Framework is a method that encodes time-ordered market paths using iterated integrals for pricing a variety of path-dependent derivatives.
• It employs mathematical tools like truncated signatures and signature kernels to approximate continuous path-functionals accurately and efficiently.
• Advanced surrogate models and deep-learning architectures, such as Sig-LSTM, are used to rapidly calibrate and price derivatives under non-Markovian and rough volatility conditions.

The Signature-Pricing Framework denotes, in its dominant quantitative-finance usage, a family of pricing, calibration, and hedging methods that encode time-ordered information from market paths through path signatures, log-signatures, or signature kernels, and then apply linear functionals, kernel methods, neural networks, or density estimators to derivative valuation. Across the literature, the framework is used for European, American, Asian, lookback, basket, and variance-linked claims, and is especially prominent when the underlying dynamics are path-dependent, non-Markovian, or rough (Arribas, 2018, Lyons et al., 2019, Cuchiero et al., 2022, Bayer et al., 12 Jan 2025, Molla et al., 12 Nov 2025).

1. Mathematical basis: signatures, truncation, and universality

Given a continuous path $X:[0,T]\to\mathbb{R}^d$, its signature is the infinite collection of iterated integrals

$$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$

with level-$\ell$ coordinates

$$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$

In practice one truncates at level $N$, obtaining a finite feature map $S^N(X)$. Several papers use time augmentation, such as $\widehat X_t=(t,X_t)$, or richer augmentations including volatility or lead–lag coordinates, so that the representation retains sufficient information about the original path (Arribas, 2018, Bayer et al., 12 Jan 2025, Molla et al., 12 Nov 2025).

The central structural result is universality. Linear functionals on truncated signatures approximate continuous path-functionals arbitrarily well as $N\to\infty$, via Stone–Weierstrass- or Chow–Lyons-type arguments. In the derivative-pricing literature this is the basis for regarding signatures as a universal path-space feature representation. Perez Arribas formulates this through signature payoffs and a density theorem on compact sets of augmented paths (Arribas, 2018). Lyons, Nejad, and Pérez Arribas extend the same idea to model-free exotic pricing with implied expected signatures (Lyons et al., 2019). For optimal stopping, however, weak-topology continuity fails; the higher-rank signature literature replaces ordinary law-based regression by regression on measure-valued paths and adapted topologies (Horvath et al., 2023).

A related construction is the signature kernel, typically an inner product on truncated signatures or a Gaussian RBF applied to signature vectors. In American-option pricing under rough volatility, the kernel induces an RKHS of path-functionals and is used in primal–dual algorithms and distribution regression (Bayer et al., 12 Jan 2025, Shah, 10 Aug 2025, Horvath et al., 2023).

2. Signature payoffs, expected signatures, and linear pricing

The earliest signature-pricing formulations are built around the notion of a signature payoff: a payoff that is a linear functional of the signature. If $w$ collects coefficients indexed by multi-indices $I$, then

$$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$0

This construction turns path dependence into finite-dimensional linear algebra once a truncation level is fixed. The pricing principle is correspondingly simple: under a risk-neutral measure, the fair price is the inner product of the payoff coefficients with the expected signature of the underlying path (Arribas, 2018, Lyons et al., 2019).

For model-based pricing, one calibrates $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$1 by regressing a target payoff on truncated signature features and then computes

$$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$2

In the Black–Scholes experiments of Perez Arribas, truncation $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$3 on a 3-dimensional augmented path gave $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$4 features, out-of-sample $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$5 for European call, American put, Asian call, lookback call, and variance swap, pricing error on the order of $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$6 of notional, and price evaluation of approximately $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$7 ms once the expected signature had been computed (Arribas, 2018).

For model-free pricing and hedging, the key object becomes the implied expected signature. Lyons–Nejad–Pérez Arribas formulate calibration as a linear inverse problem: approximate traded exotic payoffs by signature payoffs, solve a regularized least-squares system for the unknown expected signature coordinates, and then price a new exotic by another inner product. In their Section 5.2 example, calibration on $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$8 vanillas, $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$9 up-and-out barriers, $\ell$0 up-and-in barriers, and $\ell$1 variance swaps with $\ell$2 and $\ell$3 paths yielded $\ell$4 and mean-squared-errors down to $\ell$5, while calibration degraded sharply if the traded set was too small or contained only vanilla options (Lyons et al., 2019). In the related nonparametric framework for pricing and hedging exotic derivatives, the same linearization supports both pricing from observed exotic prices and $\ell$6-optimal signature trading strategies for hedging (Lyons et al., 2019).

3. Signature-based asset models, volatility models, and calibration

A second major branch treats the asset itself as a linear functional of the signature of a primary process. In Cuchiero et al., a signature-based model of order $\ell$7 is

$\ell$8

where $\ell$9 is the time-extended primary process. The stochastic-integral representation of $$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$0 yields explicit local-martingale conditions, and hence no-arbitrage conditions, in terms of vanishing drift and bracket terms. This paper also defines sig-payoffs and shows that their prices reduce to finite sums involving polynomial expressions in $$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$1 and unconditional moments of the signature of the primary process (Cuchiero et al., 2022).

The tractability of this representation is especially visible in calibration. For time-series calibration, Cuchiero et al. use a single linear regression on precomputed signature features. For implied-volatility-surface calibration, the model price is evaluated by Monte Carlo after the path signatures of the primary driver have been computed offline, so each optimization step reduces to repeated dot-products. Reported results include Heston-generated surfaces with $$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$2, $$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$3 parameters, and $$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$4, where absolute implied-volatility errors were below a few basis-points in $$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$5–$$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$6 minutes; on real S&P 500 data dated March 17, 2021, the method recovered a full $$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$7-point smile within approximately $$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$8–$$X^{i_1,\dots,i_\ell}_{0,T} = \int_{0<t_1<\cdots<t_\ell<T} dX^{i_1}_{t_1}\cdots dX^{i_\ell}_{t_\ell}.$$9 bps in similar time, and slice-wise calibration achieved sub-$N$0 bps error on each maturity (Cuchiero et al., 2022).

A closely related development is the signature volatility model, where the volatility itself is modeled as a linear functional of the time-extended signature of Brownian motion. This framework contains explicit or approximate signature representations of Stein–Stein, Bergomi, and Heston-type dynamics, and derives a joint characteristic functional of log-price and integrated variance from an infinite-dimensional Riccati equation on the extended tensor algebra. European and path-dependent options are then priced by Fourier inversion, while quadratic hedging is obtained from the same characteristic representation (Jaber et al., 2024). Reported numerical performance uses truncation levels $N$1–$N$2, $N$3 time steps, and $N$4–$N$5 Fourier nodes, with characteristic-function evaluations on the order of $N$6–$N$7 ms on a modern CPU and implied-volatility errors of order $N$8–$N$9 against Monte Carlo on OU, Heston, and m-GBM examples (Jaber et al., 2024).

4. Learning architectures and surrogate pricing

Recent work replaces explicit linear pricing functionals by learned maps from signature features to continuation values, BSDE controls, densities, or prices.

Approach	Signature input	Output
Deep Signature / Log-Signature FBSDE (Feng et al., 2021, Sun et al., 2024)	Segment-level truncated signature or log-signature of time-augmented paths	$S^N(X)$0, $S^N(X)$1, or reflected BSDE values
Deep signature under non-Markovian volatility (Ma et al., 21 Aug 2025)	Signatures of time-extended Brownian motion	Volatility approximation and option price
Signature-conditioned MDN (Molla et al., 12 Nov 2025)	$S^N(X)$2	Conditional terminal density and basket-option price

In the Deep Signature FBSDE Algorithm, signature or log-signature blocks are fed into an RNN, typically an LSTM, to approximate $S^N(X)$3 on a coarse grid while retaining fine path information. The reported examples include a best-ask price under no-good-deal bounds for a GBM European call, where Sig-LSTM ran $S^N(X)$4 faster than a vanilla Euler-DNN, and a Black–Scholes lookback option, where Sig-LSTM with $S^N(X)$5 segments and $S^N(X)$6 fine steps attained $S^N(X)$7 error versus closed form while running $S^N(X)$8 faster than a vanilla LSTM on $S^N(X)$9 time-steps (Feng et al., 2021). The high-dimensional extension uses feedforward networks on truncated signatures and proves convergence of both forward and backward BSDE solvers; in a pure-quadratic path functional with $\widehat X_t=(t,X_t)$0, the forward method produced $\widehat X_t=(t,X_t)$1 versus exact $\widehat X_t=(t,X_t)$2, and the backward method $\widehat X_t=(t,X_t)$3, both with errors around $\widehat X_t=(t,X_t)$4 (Sun et al., 2024).

In the non-Markovian stochastic-volatility setting, Ma–Wu–Li reformulate the volatility dynamics as a rough SDE represented through linear or nonlinear combinations of time-extended Brownian motion signatures. Their deep signature architecture uses a 1D convolutional layer with $\widehat X_t=(t,X_t)$5 channels and kernel size $\widehat X_t=(t,X_t)$6, a two-layer LSTM with $\widehat X_t=(t,X_t)$7 hidden units per layer, and a 3-layer MLP with $\widehat X_t=(t,X_t)$8 neurons each. For rough Heston and rough Bergomi, the nonlinear signature network outperformed the linear signature network for approximating $\widehat X_t=(t,X_t)$9, and by $N\to\infty$0 the option-price errors for both methods were already below $N\to\infty$1 in OTM, ATM, and ITM cases (Ma et al., 21 Aug 2025).

The most explicit surrogate-density formulation appears in the generative basket-option paper. There, a Mixture Density Network with $N\to\infty$2 hidden layers of sizes $N\to\infty$3 and LeakyReLU activations maps truncated signatures of time-varying inputs, a scaled Cholesky factor $N\to\infty$4, maturity $N\to\infty$5, and optionally basket weights $N\to\infty$6, to mixture weights, means, and scales of a Gaussian mixture density

$N\to\infty$7

Training uses AdamW, LogSumExp-stabilized negative log-likelihood, and Monte Carlo targets under GBM with time-varying volatility or local volatility. The reported configuration uses $N\to\infty$8 Gaussians, signature level $N\to\infty$9, approximately $w$0 million training samples, batch size $w$1, and approximately $w$2 minutes per epoch. Across maturities from $w$3 month to $w$4 year and correlations in $w$5, the model achieved typical $w$6, pricing errors typically within a few basis points versus $w$7-path Monte Carlo, and inference latency of approximately $w$8 ms in the “train-once, price-anywhere” regime (Molla et al., 12 Nov 2025).

5. American options, optimal stopping, and rough volatility

For Bermudan and American options, the framework bifurcates into primal continuation-value regression and dual martingale regression. Bayer, Pelizzari, and collaborators use truncated signatures of volatility or state paths as non-Markovian features in both components. In the primal method, continuation values are regressed on signature features, generalizing Longstaff–Schwartz. In the dual method, the martingale integrand is parameterized in an RKHS induced by the signature kernel, and the upper bound is obtained from the Andersen–Broadie–Rogers dual representation (Bayer et al., 12 Jan 2025).

The rough-volatility implementation compares linear signature, deep signature, and signature-kernel methods under rough Bergomi and rough Heston. In the rough Bergomi case with $w$9 and strike $I$0, the reported lower and upper bounds were $I$1 for linear signature, $I$2 for deep signature, and $I$3 for signature-kernel, with corresponding duality gaps of approximately $I$4, $I$5, and $I$6. The study also reports that signature-kernel methods are most stable in low-data regimes but expensive in offline PDE solves, while deep-signature methods often yield the sharpest dual bounds (Bayer et al., 12 Jan 2025).

Shah extends this line to time-varying roughness. The framework first estimates a rolling Hurst parameter from the previous $I$7 trading days, then predicts the future Hurst path with $I$8 separate XGBoost regressors, each using $I$9 trees, learning rate $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$00, and maximum depth $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$01. The average predicted roughness $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$02 drives a regime switch: rough Bergomi if $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$03, Heston otherwise. Signature-kernel evaluations are accelerated by Random Fourier Features, reducing full Gram-matrix complexity from $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$04 to $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$05. On $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$06-day American puts dated August 31, 2023, AAPL had average MAE $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$07 and MSE $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$08, leading to a Heston regime, while META had average MAE $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$09 and MSE $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$10, leading to a rough Bergomi regime. The Deep Kernel (RFF) bounds were $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$11 for AAPL, containing the market premium $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$12, and $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$13 for META, narrowly containing the market premium $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$14 (Shah, 10 Aug 2025).

A common misconception is that weakly continuous distribution regression on path laws suffices for American options. The higher-rank signature literature states that this fails because American-option value functions are in general discontinuous with respect to the weak topology. The proposed remedy is a regression framework based on higher rank signatures of measure-valued paths, kernel mean embeddings, and Aldous’s extended weak topology, with the main computational core given by families of two-dimensional hyperbolic Goursat PDEs (Horvath et al., 2023).

6. Limitations, trade-offs, and divergent usages of the term

Several limitations recur across the finance literature. The first is the curse of dimensionality in the truncation level: feature dimension grows like $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$15 or $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$16, depending on the formulation (Arribas, 2018, Sun et al., 2024). The second is numerical stability: high-order iterated integrals can be very small or noisy, which motivates centering, rescaling, regularization, log-signature variants, orthonormalized bases, or learned embeddings (Arribas, 2018, Sun et al., 2024). The third is that model-free implied-signature methods require a rich enough family of traded payoffs to stabilize inversion; calibration can degrade sharply if only vanillas are available (Lyons et al., 2019). In rough-volatility American pricing, additional limitations include noisy Hurst estimates, misclassification near the $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$17 threshold, shallow Heston calibration to ATM volatilities only, limited simulation budgets, and RFF sampling noise (Shah, 10 Aug 2025).

The term is also used in distinct, non-equivalent senses outside rough-path finance. This suggests that “signature-pricing framework” is polysemous across arXiv. In Bayesian marketing science, it denotes a Bayesian Hierarchical Conjoint Analysis workflow for estimating dollar-denominated willingness-to-pay,

$$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$18

with individual random utility, hierarchical priors, and NUTS estimation. In the iPhone case study, the setup used $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$19 simulated respondents, $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$20 binary choice tasks per respondent, $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$21 chains with $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$22 post-warmup draws each, $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$23, and posterior revenues for a simulated “Pro-bundle” peaking at $$S(X)_{0,T} = \Bigl( 1,\; \int_0^T dX_t,\; \int_{0<t_1<t_2<T} dX_{t_1}\otimes dX_{t_2},\; \dots \Bigr),$$24 (Pillai et al., 14 Sep 2025). In privacy-preserving transport systems, the term refers to a group-signature-based electronic toll pricing system in which users anonymously sign hashed location-time tuples, the server publishes homomorphically encrypted fees, and disputes are resolved by opening group signatures and recomputing encrypted sums; the protocol is organized into Setup, Driving, Toll Calculation, and Dispute Resolving phases (Chen et al., 2011).

Within quantitative finance, however, the unifying idea remains stable: signatures provide a universal, order-sensitive representation of paths, and pricing becomes a problem of learning or calibrating linear, kernel, neural, or density-based functionals on that representation. The resulting methods are valued for universality, reuse of precomputed path features, tractable calibration, and the ability to handle path dependence and non-Markovianity in settings where PDE state augmentation is impractical (Lyons et al., 2019, Cuchiero et al., 2022, Molla et al., 12 Nov 2025).