3.15.5 Failure Modes & Practice

Basic Prerequisites: 02 · The Signature Algebra and 04 · Rough Path Theory.

---

1. Intuition & Practical Objective

Pages 01–03 proved beautiful identities about the signature of a path. This page is about the gap between that theory and raw financial data, and the two standard bridges across it:

1. Raw data is non-geometric. The signature identities (shuffle product, geometric rough-path structure) hold for geometric (Stratonovich) iterated integrals. Financial increments are Itô-type - and on top of that, a univariate series is pathologically uninformative: the signature of a 1D path is just its endpoint. The two fixes are time augmentation and the lead-lag transform, both of which geometricise the data and make the signature informative. They are not optional pre-processing; they are what makes the theory applicable at all.
2. Signatures are hedging instruments. Because path-functionals are representable (universally) as combinations of signature terms, the ingredients of a hedge - the quantities a desk needs to replicate a path-dependent payoff - are themselves signature features. The cleanest example: the variance-swap payoff (realized variance) is, up to a factor, the Lévy area of the lead-lag path, i.e. a level-2 signature coefficient. Signatures do not merely describe paths; they price and hedge them.

Three practical claims:

3. The endpoint is not the signal. For a univariate path, $S^{11}=\frac{1}{2}(S^1{)}^2$ is determined by the endpoint - the raw signature contains no shape information. Augmenting with time, $X_t\mapsto (t,X_t)$, restores a meaningful area (the signed area under the path).
4. Lead-lag makes the area equal a variance. The lead-lag path $(X_t,X_{t+\delta })$-style construction turns the quadratic variation into the Lévy area of the transformed path. Since realized variance is a (stylised) variance-swap payoff, the lead-lag signature is a direct pricing/hedging object - the bridge used in the literature (Lyons–Ni–Zhang) for rough-vol lead-lag signatures.
5. Truncation, augmentation and the lift are coupled choices. The depth to which you truncate, how you augment, and which rough-path lift you implicitly selected together determine what your signature does and does not see (§4).

The practical objective: know how to take a raw time series, geometricise it (time augmentation or lead-lag), compute a meaningful truncated signature/log-signature, and connect specific signature features to specific hedging quantities.

---

2. Mathematical Ground Truth & Derivations

2.1 The failure: raw 1D and Itô-type data

A 1D path $X_t$ has signature with words over a single letter; every level is a power of the endpoint, $S^{1^k}=(S^1{)}^k/k!$. So the raw 1D signature carries only the endpoint - all shape is gone. For a multivariate path the situation is better but still Itô-flawed: the classical (Itô) iterated integrals fail the shuffle identity (an Itô correction at the diagonal), so the object that falls out of raw increments is not a geometric signature.

2.2 Time augmentation

For a path $X_t$ (any dimension), the time augmentation is ${\hat{X}}_t=(t,X_t)\in {\mathbb{R}}^{d+1}$. Adding time as a coordinate achieves three things at once: the path becomes strictly non-degenerate (time strictly increases), the level-2 cross terms with time acquire meaning (the area under the path), and the leading behaviour is regularised. §3 shows two same-endpoint univariate paths whose raw signatures are identical but whose augmented signatures differ exactly through the area $-\int X_{t}dt$ captured at level 2. Time augmentation is the minimal geometricisation: it requires no extra parameters.

2.3 The lead-lag transform

The lead-lag transform of a discrete path $X_0,X_1,\dots ,X_n$ builds a path in ${\mathbb{R}}^{2d}$ that moves the lead copy first and the lag copy second:

$$X_0^{lead}=X_0^{lag}=X_0,\text{then }(X_1,X_0)\to (X_1,X_1)\to (X_2,X_1)\to \cdots$$

Between each pair of sample points the lead-lag path sweeps a right triangle in the (lead,lag) plane, so its Lévy area equals

$$\boxed{{\mathrm{Area}}_{\text{lead-lag}}=\frac{1}{2}\sum _k(X_{k+1}-X_k{)}^2=\frac{1}{2}\text{realized variance}}$$

Each little triangle has legs $\Delta X_k,\Delta X_k$ (area $\frac{1}{2}\Delta X_k^2$), and the sum is exact - §3 verifies it to $10^{-14}$. Consequences: (a) the lead-lag path is geometric (the shuffle identities hold on it - the cross-covariance between lead and lag is the quadratic covariation), and (b) the variance-swap payoff $=$ realized variance $=2\times$ the lead-lag Lévy area, a level-2 signature term. This is the mathematical core of signature-based pricing of variance and of the lead-lag signature of rough volatility (Lyons–Ni–Zhang).

2.4 Signature-based hedging

The universality of the signature (linear combinations of signature terms approximate any continuous path-functional) makes the signature the natural basis for hedging: a payoff $F(X)$ is approximated by ${\sum }_w{\alpha }_w{S}^w(X)$, and a delta/vega-style hedge is built from the sensitivity of those basis terms. In the Malliavin / functional-Itô framework (Dupire; Cont–Fournié) the hedging weights are signature-driven. The concrete, fully-computable instance in this folder is the variance swap: its payoff is a signature feature (the lead-lag area), so a variance-swap hedge is a signature-level-2 position, exactly as the log-contract replication of Volatility Modeling suggests. Signatures give a systematic coordinate system for this replication rather than an ad hoc set of contracts.

---

3. Computational Implementation - geometricisation (lead-lag & time augmentation) and the variance-swap signature

We (i) build the lead-lag transform of a seeded random-walk path and verify the Lévy-area-equals-half-realized-variance identity to machine precision, (ii) show time augmentation makes two same-endpoint univariate paths distinguishable (identical raw 1D signature, different augmented areas), and (iii) state the variance-swap payoff as a signature feature. Stdlib only, deterministic (random.seed).

Reading the output.

• Lead-lag turns a variance into an area, exactly. The lead-lag path’s Lévy area equals half the realized variance to $1.4\times 10^{-14}$. This is not an approximation - each lead-lag triangle contributes exactly $\frac{1}{2}\Delta X_k^2$. The consequence is the workhorse identity of signature pricing: the variance-swap payoff (realized variance) is a level-2 signature coefficient of the lead-lag path.
• Time augmentation restores the shape that 1D throws away. The two paths have identical raw 1D signatures ($S^1=2.0$, $S^{11}=2.0$ - the endpoint in disguise) but their time-augmented areas differ: $-0.667$ vs $+0.667$, i.e. the signed area under the two different curves. Augmentation is the minimal fix that makes univariate signatures carry information.
• The hedge is a signature position. Because realized variance $=2\times (\text{lead-lag area})$, a variance-swap hedge is literally a trade in a level-2 signature feature of the lead-lag path. This is the concrete, checkable seed of signature-based hedging: the building blocks of path-dependent hedges are signature coefficients, not hand-chosen contracts.

---

4. Failure Modes & First-Principles Breakdowns

1. Feeding raw 1D signatures to a model. A univariate signature is the endpoint in disguise ($S^{11}=\frac{1}{2}(S^1{)}^2$); the model learns nothing about the path’s shape. Always time-augment or lead-lag univariate data first - this is the single most common mistake in applied signature work (Chevyrev–Kormilitzin stress it explicitly).
2. Using Itô integrals and expecting geometric identities. The shuffle product and the geometric rough-path structure hold only for Stratonovich-type (piecewise-linear / geometric) integrals. Raw increments give an Itô object that fails the identities; the lead-lag and time-augmented constructions are exactly the fixes that restore the geometric convention.
3. Reading reparametrisation invariance as “time doesn’t matter.” It matters enormously - that is the point of lead-lag. The raw signature is speed-blind, which is a feature for geometry and a bug when the timing/order of events is the signal (order flow, lead-lag of vol). Lead-lag deliberately re-encodes time so that the order of lead/lag coordinates carries the cross-covariance information.
4. Truncating without a scale. The signature’s high levels carry small, fast oscillations; truncating at low level literally discards the short-time structure - exactly what rough vol depends on (§06). The truncation level and the time step are jointly decided, not separately.
5. Treating the lead-lag area as “free.” The identity §3 is exact for a given sampling; it is a chosen lift of the raw series, and its value is convention-dependent (Itô vs Stratonovich, choice of time grid). What is robust is the relation between the area and realized variance, not the absolute area on any one realisation.

---

5. References

• Chevyrev, I. & Kormilitzin, A. (2016), A Primer on the Signature Method in Machine Learning, arXiv:1603.03788
• Lyons, T. J., Ni, H., Zhang, H. (2019), Machine Learning Models of Financial Time Series, arXiv:1905.11666
• Chevyrev, I. & Oberhauser, H. (2022), Signature moments to characterize laws of stochastic processes (J. Mach. Learn. Res.)
• Dupire, B. (2009) / Cont, R. & Fournié, D.-A. (2010), Functional Itô calculus
• Gatheral, Jaisson & Rosenbaum (2018), Volatility is rough
• Friz & Victoir (2010), Multidimensional Stochastic Processes as Rough Paths

---

6. Connected Graph Bridges

• Back: 04 · Rough Path Theory
• Forward: 06 · Advanced Extensions · Index Hub
• Cross-links: Rough Volatility · Volatility Modeling (realized variance, variance swaps) · Numerical Methods · 03 Monte Carlo (simulation of lifted paths) · Deep Learning for Sequences