3.15.6 Advanced Extensions

Basic Prerequisites: 03 · Log-Signature & Uniqueness and 05 · Failure Modes & Practice. The statistical framing is Statistics & Inference.

1. Intuition & Practical Objective

Pages 01–05 built signatures for single paths. Almost every applied problem, though, is about a distribution of paths or a similarity between paths:

“Given the history of spot and vol, what is the expected payoff / hedge ratio?” - an expectation over paths.
“Which of two historical episodes is more like today?” - a similarity between paths.
“Train a model on thousands of intraday paths.” - a feature map into a vector space.

The signature answers all three through two upgrades, and this page is about them:

The expected signature $E [S (X)]$ - the path-level analogue of the moment generating function. By the Chevyrev–Lyons uniqueness theorem, the expected signature determines the law of the path (under mild conditions): it is a universal, faithful characterisation of a distribution of paths, not just a descriptor of one. It is also the object that makes signature-based pricing and hedging rigorous - prices are expectations of path-functionals, and those functionals are signature-features.
The signature kernel $K (X, Y) = ⟨ S (X), S (Y)⟩ = \sum_{w} S^{w} (X) S^{w} (Y)$ - a positive-definite inner product on paths that respects the ordered geometry of the whole path. It lets any kernel method (SVM, Gaussian-process regression, kernel ridge) operate directly on paths without hand-engineered features, and it is computable recursively in linear time (the signature kernel, Kiraly–Oberhauser). This is the machine-learning face of the folder.

A third application closes the loop to finance:

Lead-lag signatures for rough volatility. Because vol is rougher than Brownian ( $H \approx 0.1$ , Gatheral–Jaisson–Rosenbaum 2018), and because the lead-lag transform turns quadratic covariation into a signature (Lévy) area (§05), the lead-lag signature of spot-and-vol is exactly the right feature to learn the spot/vol dependence that prices vol products. Lyons–Ni–Zhang make this concrete: signature features of lead-lagged (log-price, vol) paths drive prediction of implied/realised vol. §3 shows the lead-lag Lévy area of a rough 2D path is a stable, computable object - the building block of those features.

The practical objective: know the expected signature and its uniqueness property, know the signature kernel and that it is computable and positive-definite, and see how lead-lag + signature features turn a rough-vol path into ML-ready inputs.

2. Mathematical Ground Truth & Derivations

2.1 The expected signature and law-uniqueness

For a random path $X$ (a stochastic process over $[0, T]$ ), the expected signature is

E [S (X)] = (E [S^{w} (X)])_{w}, E [S^{w} (X)] = \int S^{w} (x) d μ (x),

where $μ$ is the law of $X$ . For a 1D Brownian motion, because $S^{1^{k}} = B_{T}^{k} / k!$ , the expected signature is determined by the Gaussian moments: odd levels vanish and $E [S^{11}] = T /2$ , $E [S^{1111}] = T^{2} /8$ - the exact values the §3 Monte Carlo reproduces. The central result (Chevyrev–Lyons 2016) is that, under mild moment conditions, the expected signature characterises the law: $E [S (X)] = E [S (Y)] \Rightarrow X ≜ Y$ (the laws are equal). It is the path-space analogue of “the moment generating function determines the distribution” - and it is why the expected signature, not ad hoc summary statistics, is the right description of a path ensemble for pricing and hedging.

2.2 The signature kernel

For two paths $X, Y$ , the signature kernel is the inner product of their truncated-or-full signatures:

K (X, Y) = ⟨ S (X), S (Y)⟩ = w \sum S^{w} (X) S^{w} (Y)

The sum runs over all words $w$ (in practice truncated at some level $N$ ). It is symmetric and positive semi-definite (it is a Gram-matrix of the signature feature map), so it is a valid kernel; §3 verifies the Cauchy–Schwarz inequality numerically. Because the signature is reparametrisation-invariant and complete, $K$ is a natural similarity between paths - two paths are close iff their ordered geometry (increments and areas and higher shape) is close. Computationally the kernel avoids materialising the (huge) truncated signature explicitly: it is evaluated recursively in $O (T_{1} T_{2})$ time (Kiraly–Oberhauser 2019). This is the practical enabler for kernel methods on thousands of paths.

2.3 Signatures, expected signatures, and pricing/hedging

A (path-dependent) payoff is a functional $F (X)$ ; universality says $F \approx \sum_{w} α_{w} S^{w}$ , so its risk-neutral price is $E [F (X)] \approx \sum_{w} α_{w} E [S^{w}] = ⟨ α, E [S]⟩$ - a linear function of the expected signature. This is the systematic content of signature-based pricing: price vectors are linear functionals of the expected signature, and hedging weights are the sensitivities of those functionals. The §3 Monte Carlo, by estimating $E [S]$ for Brownian motion, is precisely the numerical ingredient of such a pricing scheme.

2.4 Lead-lag signatures for rough volatility

Rough vol: $ln σ$ is a fractional process with $H \approx 0.1$ (Gatheral–Jaisson–Rosenbaum 2018), so the spot-and-vol joint path is genuinely rough and its level-2 iterated integrals require the rough-path lift. The lead-lag transform (§05) makes the quadratic covariation between spot and vol visible as a signature area, and the truncated signature (or kernel) of the lead-lagged $(X, σ)$ path becomes a rich, geometry-respecting feature set. Lyons–Ni–Zhang (2019) use exactly this: signature features of lead-lagged financial series predict vol, and the lead-lag area is the signature-level realisation of the spot/vol covariance. §3 computes the lead-lag Lévy area of a Brownian 2D path (the rough-vol building block) and shows it is stable under refinement.

3. Computational Implementation - expected signature, signature kernel, and the rough-vol lead-lag area

We (i) estimate the expected signature of 1D Brownian motion by seeded Monte Carlo and compare to the exact analytic values, (ii) evaluate the signature kernel between two paths (truncated level 2) and verify positive semi-definiteness via Cauchy–Schwarz, and (iii) compute the lead-lag Lévy area of a 2D Brownian path (the rough-vol feature building block) and check its stability under refinement. Stdlib only, seeded (random.seed), deterministic.

Reading the output.

The expected signature is the law’s moment-generating object. The Monte Carlo estimates converge to the analytic Gaussian values: $E [S^{11}] = 0.49055$ vs $0.5$ and $E [S^{1111}] = 0.11719$ vs $0.125$ (MC error ~ $2%$ at 4000 paths and a coarse 200-step grid), odd levels ~ $0$ . Since these moments characterise the law (Chevyrev–Lyons), estimating $E [S]$ is estimating the law - the seed of signature-based pricing, where a price is a linear functional of the expected signature.
The signature kernel is a valid inner product on paths. $K (X, Y) = 170.5$ with $K (X, X) = 56.75$ , $K (Y, Y) = 556.75$ ; Cauchy–Schwarz holds ( $K_{X Y}^{2} = 29070.25 \leq 31595.56$ ). The kernel is the similarity an SVM/GPR would feed on - computed here by explicit truncated sums, but in production by the linear-time recursive algorithm.
The rough-vol lead-lag area is a stable building block. The lead-lag Lévy area of the 2D Brownian path stays bounded and converges under refinement ( $+ 0.244 \to + 0.233$ ). This is the same construction §05 used for realized variance, now as the spot/vol covariance feature that lead-lag signature methods feed to a rough-vol predictor (Lyons–Ni–Zhang).

4. Failure Modes & First-Principles Breakdowns

Estimating expected signatures with too few paths / too low truncation. The signature levels decay only factorially, so high-level moments are small but slow to estimate; with a coarse grid and a few thousand paths the estimator error ( $\sim 2%$ in §3) compounds at higher level. Budget the grid and the path count against the level you actually need.
Believing the expected signature is the same as the mean path. It is a tensor-valued moment - it captures the whole ordered geometry (areas, volumes) of the ensemble, not just an average trajectory. Reading it as a mean path loses the covariance content that is the whole point.
Using the kernel without checking positive semi-definiteness / truncation convergence. $K$ is PSD in the limit; a truncated kernel is PSD too (Gram of a finite feature map) but its values are truncation-dependent. Two implementations truncating at different levels give different “similarities” - fix the level by a stability study, not by convenience.
Ignoring reparametrisation invariance when it matters. The signature (and kernel) is speed-blind; if two paths differ only by how fast they were traversed, they are kernel-identical. When timing matters, lead-lag/time augmentation (§05) must be applied before the kernel - otherwise the “similarity” silently ignores the timing signal.
Applying raw-signature features to univariate series. Same failure as §05: a univariate signature is the endpoint in disguise, so the kernel/features learn nothing about shape. Always augment (time, lead-lag) univariate inputs first.
Jumping to rough-vol lead-lag without the lift. Rough paths need the geometric lift; the lead-lag area is a chosen lift whose value is convention-dependent. Computing it from raw increments without fixing the geometric convention gives an Itô-flawed, convention-arbitrary number (§04).

5. References

Chevyrev, I. & Lyons, T. (2016), Characteristic functions of measures on geometric rough paths, Annals of Probability 44(6)
Chevyrev, I. & Kormilitzin, A. (2016), A Primer on the Signature Method in Machine Learning, arXiv:1603.03788
Kiraly, F. & Oberhauser, H. (2019), Kernels for sequentially ordered data, JMLR
Lyons, T. J., Ni, H., Zhang, H. (2019), Machine Learning Models of Financial Time Series, arXiv:1905.11666
Gatheral, J., Jaisson, T., Rosenbaum, M. (2018), Volatility is rough, Quantitative Finance 18(6)
Friz & Victoir (2010), Multidimensional Stochastic Processes as Rough Paths

6. Connected Graph Bridges

Back: 05 · Failure Modes & Practice · 03 · Log-Signature & Uniqueness · Index Hub
Forward / applications: Deep Learning for Sequences · RNNs & LSTMs · Feature Engineering & Labeling · Signal Processing & Kalman
Sibling / theory: Rough Volatility · Statistics & Inference · Probability & Measure Theory · Quantitative Risk

Kwant Atlas

Explorer