Signed, Sealed, Delivered: A Rough Path to Better Volatility Models

TL;DR for operators

Options calibration has a familiar operational problem: the model that is fast enough to run every day is usually the model that assumes the market is behaving politely. The market, naturally, has other hobbies.

This paper compares two ways of calibrating implied volatility surfaces. The first is the classical route: use model-specific analytical approximations for Heston and rough Bergomi. The second is the rough-path route: represent volatility as a linear functional of the truncated signature of a primary stochastic process.¹

The useful conclusion is not “signatures beat formulas”. They do not, at least not in that lazy headline form. Analytical expansions remain the cleaner choice when the volatility model is correctly specified: they are low-dimensional, fast, interpretable, and highly accurate. The paper’s Heston and rough Bergomi formula-based calibrations are deliberately strong baselines, not straw men politely waiting to be defeated.

The signature method matters because it changes the calibration architecture. Instead of hard-coding a volatility process and then forcing market prices to confess the parameters, it uses path signatures as features that encode how a driving process has evolved through time. Volatility becomes

where $S(\widehat{X})^{\leq N}_t$ is the truncated signature of a time-augmented path. In plain desk language: the model learns volatility from a structured memory of the path, not just from today’s state.

For an options desk, the operating pattern is straightforward:

The business takeaway is therefore boring in the right way: use analytical formulas as the production benchmark, signatures as the flexible challenger, and the gap between them as a model-risk signal. Glamorous? No. Useful? Yes, which is worse for the PowerPoint industry.

The calibration problem is not pricing; it is choosing what to believe

In options markets, implied volatility is not just an output. It is the language in which prices are quoted, compared, risk-managed, and argued over at unhelpful times of day. A calibration engine translates a surface of market option prices into model parameters. Those parameters then feed pricing, hedging, scenario analysis, VaR, XVA, risk limits, stress tests, and trader confidence, which is technically not a risk factor but often behaves like one.

The awkward part is that calibration is never purely numerical. It is also a modelling commitment.

A Heston calibration says volatility is stochastic, mean-reverting, Markovian, and driven by a variance process with a specific square-root structure. A rough Bergomi calibration says volatility has fractional, rough behaviour, often better aligned with empirical short-maturity volatility smiles. A signature model says something different: perhaps the surface can be fitted by learning a functional of the path itself, without committing so early to one named parametric temple.

The paper’s comparison is valuable because it does not pretend that one method makes the others obsolete. Analytical expansions and signature methods solve different pain points:

That last row is not a footnote. It is the bill.

Analytical calibration wins when the model has already won

The paper first revisits an analytical calibration method for Heston. This is not a novelty contest. The Heston method from Alòs, De Santiago, and Vives is used as a benchmark because it is already good.

The logic is familiar. Under Heston dynamics, the stock price follows a stochastic volatility process whose variance mean-reverts. Since exact implied-volatility calibration is difficult, the method uses a second-order approximation to implied volatility. From short-maturity, long-maturity, and at-the-money information, it derives equations that recover the Heston parameters.

The benchmark calibration is extremely accurate. In the uncorrelated Heston case, the paper reports calibrated values close to the true synthetic parameters: for example, an initial volatility of $0.2$ is recovered as $0.200013$, volatility-of-volatility $0.3$ as $0.307340$, mean reversion $3$ as $2.998598$, and long-run variance $0.09$ as $0.089960$. In the correlated case, the same pattern remains: the recovered parameters are still close to the generating values.

The paper then introduces its new analytical contribution for rough Bergomi: a closed-form calibration scheme using short-time implied-volatility asymptotics and VIX-implied volatility information. The route is operationally interesting because rough Bergomi is usually attractive but computationally unpleasant. The authors estimate:

1. the Hurst parameter $H$ from short-maturity skew behaviour;
2. the volatility-of-volatility parameter $\eta$ using short-time ATM implied volatility of VIX options;
3. the correlation parameter $\rho$ from short-time ATM skew;
4. the initial volatility $\sigma_0$ from ATM implied volatility across maturities.

On synthetic rough Bergomi data, this VIX-informed procedure also calibrates closely: $\sigma_0=0.2$ is estimated as $0.199884$, $H=0.1$ as $0.100968$, $\eta=0.5$ as $0.490527$, and $\rho=-0.7$ as $-0.672485$.

That is the first important correction to the reader’s likely instinct. The analytical baselines are not weak. If the model family is right, formulas are hard to beat. They are fast, compact, and interpretable. In finance, that combination is not old-fashioned. It is usually called “the thing that survives production”.

Signatures are not magic; they are structured memory

The signature method enters because markets may not respect the model family. This is inconvenient, but hardly breaking news.

A path signature is a collection of iterated integrals. The first level captures increments. Higher levels capture ordered interactions through time. If the path moves up, down, accelerates, reverses, or carries temporal patterns that matter for volatility, the signature provides a systematic way to encode those patterns.

The paper motivates this through rough path theory. Classical integration becomes fragile when paths are too irregular. Brownian and rough-volatility paths do not always behave nicely enough for ordinary first-order approximations. Rough paths add higher-order information so the path can be treated as an enhanced object rather than a mere sequence of increments.

This is the mechanism that matters for business readers. Signatures are not being used because “AI-ish feature engineering” sounds fashionable. They are being used because volatility is plausibly a functional of path history, and signatures provide a mathematically disciplined basis for approximating such functionals.

The model assumes that volatility can be approximated as:

where $\widehat{X}_t=(t,X_t)$ is the time-augmented primary process. The time augmentation matters because the signature needs to retain temporal structure. A path without time can forget too much about how it arrived at the same endpoint. Markets also dislike being reduced to endpoints, which is one of their few relatable qualities.

The calibration then chooses the coefficient vector $\ell$ to minimise a weighted least-squares difference between model option prices and market option prices. The paper uses inverse Vega weights, so the calibration compares errors in a way that is more aligned with implied-volatility sensitivity.

There is also a production-relevant implementation detail: signatures are computed once and reused during optimisation. That helps. But the computation is still not free. With truncation level $N=3$, the time-augmented two-dimensional signature has 15 coefficients, while associated matrix computations involve higher signature levels up to $2N+1=7$, reaching a 255-dimensional space. In other words, the elegance of the theory still invoices the GPU.

The Heston comparison: signatures learn the surface, but formulas keep the home advantage

The Heston experiment is the cleanest comparison because the data-generating model is exactly the kind of structure that analytical methods like.

The paper calibrates signature models to 20 option contracts, with maturities ${0.1, 0.6, 1.1, 1.6}$ and strikes ${90, 95, 100, 105, 110}$. It uses 800,000 Monte Carlo paths and signature truncation level $N=3$. The numerical setup is not toy-level casual: signatures are GPU-accelerated, optimisation uses L-BFGS-B, and the authors explicitly discuss numerical diagnostics such as factorial decay of signature levels.

In the uncorrelated Heston case, the signature model fits well. The minimum loss is reported as $1.05 \times 10^{-4}$. More interestingly, the learned coefficients show that the signature model has rediscovered part of the Heston structure: the coefficient on the primary process is about $1.085$, indicating strong linear dependence on the variance process, while the intercept-like coefficient is about $0.201$, close to the initial volatility $0.2$.

This is a useful result. It suggests that the signature model is not merely drawing a fancy curve through prices. In a setting where the true structure is Heston-like, it learns a representation consistent with that structure.

But the analytical expansion still usually wins. In the 20-contract uncorrelated Heston comparison, both methods produce small implied-volatility errors, mostly in the $10^{-4}$ to $10^{-5}$ range. The analytical method generally has smaller errors, though the signature method wins in several individual contracts.

The correlated Heston case is less friendly. The signature model still calibrates, but less precisely: the minimum loss rises to $1.46 \times 10^{-3}$, and most errors sit in the $10^{-4}$ to $10^{-3}$ range. The analytical method again tends to produce smaller errors. The authors suggest a plausible reason: negative correlation may be encoded in higher-order interactions that the chosen truncation level captures only partially. Increasing the truncation from $N=3$ to $N=4$ gives only marginal improvement, while increasing cost.

This is the second important correction. Signature models are flexible, but flexibility is not the same as free expressiveness. If the dependence structure lives in higher-order interactions, truncation becomes a modelling decision. Too low, and the model misses structure. Too high, and the computation starts acting like it has tenure.

The rough Bergomi comparison: flexibility starts paying rent

The rough Bergomi setting is where the comparison becomes more interesting. Here, the market is generated from a non-Markovian rough volatility model, and the signature method uses fractional Brownian motion as the primary process.

The analytical benchmark is no longer the Heston expansion. The paper uses the VIX-informed rough Bergomi calibration scheme introduced earlier. This is an important design choice: the signature model is compared against a relevant rough-volatility baseline, not against an obviously mismatched Markovian model.

The signature implementation also evolves. Direct use of fractional Brownian motion as the primary process works but is computationally expensive because the model must implicitly learn variance positivity. A geometric transformation improves numerical behaviour, reducing runtime from roughly three hours to 39 minutes with comparable loss around $9 \times 10^{-4}$. A shifted exponential transformation improves runtime further to about 17–19 minutes and achieves a lower loss of $3.5 \times 10^{-4}$.

That transformation is not just a computational trick. It encodes a practical lesson: in flexible models, the choice of representation can matter as much as the optimiser. A badly chosen primary process asks the model to learn structural constraints indirectly. A better transformation puts useful geometry into the input before learning begins. Very rudely, the math refuses to be replaced by brute force.

In the rough Bergomi comparison, both methods achieve high accuracy. The errors are typically of order $10^{-4}$. But the signature method outperforms the VIX-informed analytical approximation in several contracts. From the table, it wins in 7 of the 20 reported strike-maturity combinations, including several away-from-the-money cases.

The paper’s interpretation is cautious and sensible. The advantage may come from the non-Markovian nature of fractional Brownian motion. Since signatures encode temporal interactions, they may represent rough path-dependent effects more naturally than a fixed parametric formula, even a well-designed rough Bergomi formula.

This does not mean the signature method has “solved” rough volatility calibration. It means it becomes more competitive precisely where model structure is harder to trust.

What each experiment actually proves

A useful way to read the paper is to separate main evidence, robustness checks, and implementation details. Otherwise, the reader risks treating every table as if it carries the same inferential weight. It does not.

This distinction matters because the business reader should not ask, “Which model won?” The better question is: “Under what modelling assumptions does each method earn its keep?”

The practical architecture is hybrid, not heroic

For a trading desk or volatility analytics vendor, the paper points toward a layered architecture.

The first layer should still be analytical. A fast Heston expansion or rough Bergomi calibration gives a clean daily benchmark. It is explainable, cheap, and stable enough to be useful in production. When the surface is well described by the assumed model, using a more flexible method may simply add compute cost and governance questions. There are easier ways to annoy risk management.

The second layer should be signature-based. Use it as a challenger model, especially around regimes where the surface shows rough, path-dependent, or non-Markovian behaviour. If the signature model consistently fits regions where the analytical model struggles, that gap is information. It may indicate model misspecification, structural change, or a need for richer dynamics.

The third layer is diagnostic. Compare the surfaces, not just the global loss. Which maturities and strikes drive the difference? Does the signature model improve short-maturity skew? Does it help in wings? Does it fail under correlation? The paper’s contract-level error tables are useful precisely because global calibration loss can hide where the model is earning or losing accuracy.

A practical implementation might look like this:

Market option surface
        |
        v
Fast analytical calibration
(Heston / rough Bergomi / desk baseline)
        |
        +--> production benchmark parameters
        |
        v
Signature challenger calibration
(primary process selected by regime hypothesis)
        |
        +--> surface residual comparison
        |
        v
Model-risk interpretation
Where does flexibility help, and is the improvement worth the cost?

The ROI is not necessarily “better prices”. It may be better diagnosis: earlier detection of model misspecification, richer challenger-model governance, and more disciplined decisions about when to move beyond a standard volatility family.

The hidden cost is not just runtime; it is governance

The paper is admirably direct about computation. Analytical approximations are essentially instantaneous once derived. Signature models remain practical, but they are heavier. The paper reports that full signature calibration can take about 15 minutes with 100,000 paths, and that increasing to 800,000 paths improves accuracy at the cost of longer runtimes, roughly 45–90 minutes depending on model. In the rough Bergomi section, representation choices move runtime dramatically, from around three hours to 39 minutes and then to 17–19 minutes.

That is manageable for research, challenger models, and overnight calibration. It is less obviously attractive for every intraday production workflow.

But runtime is only the visible cost. Governance is the subtler one. Analytical models have parameters with familiar interpretations: mean reversion, long-run variance, volatility of volatility, correlation. Signature coefficients are mathematically meaningful but less desk-native. A risk manager can understand a Heston $\rho$ moving from $-0.5$ to $-0.7$. A 15-dimensional coefficient vector over iterated integrals needs more translation before it becomes a risk narrative.

That does not make signature methods unsuitable. It means they need tooling: coefficient diagnostics, truncation sensitivity, stability checks, surface residual dashboards, and clear rules for when the challenger model is allowed to influence production marks.

Flexibility without governance is just a more expensive way to be surprised.

Where the paper’s claims stop

The paper’s strongest claims are numerical and methodological. It shows that:

• analytical expansions are excellent when the model is correctly specified;
• a VIX-informed rough Bergomi calibration can recover synthetic rough Bergomi parameters accurately;
• signature-based volatility models can fit Heston and rough Bergomi implied-volatility surfaces competitively;
• signatures appear especially useful in non-Markovian rough settings;
• computation is practical but materially heavier than closed-form calibration.

The paper does not show that signature models generate live trading alpha. It does not establish superiority on a broad panel of real-market surfaces across crises, liquidity regimes, index products, single names, rates, FX, or commodities. It does not remove the need to choose a primary process, truncation level, Monte Carlo budget, interpolation scheme, or optimisation procedure.

Those are not flaws. They are boundaries. The correct business interpretation is model-risk reduction and calibration flexibility, not a new oracle for volatility trading. We have enough fake oracles. Some of them even come with dashboards.

The real contribution is a better calibration conversation

The paper’s best contribution is not that it adds another model to the shelf. Quant finance has never suffered from a shortage of shelves.

Its contribution is a cleaner comparison between two calibration philosophies. Analytical expansions say: if the model is right, exploit structure ruthlessly. Signature methods say: if the structure is uncertain, learn a disciplined functional of the path.

Both views are correct. The mistake is treating them as enemies.

For operators, the answer is hybrid. Use analytical calibration as the fast, interpretable backbone. Use signature methods as the flexible challenger where roughness, path dependence, or regime uncertainty make the backbone creak. Track where they disagree. That disagreement may be more valuable than either model alone.

The market will not tell you which model is true. It will, however, charge you for pretending too confidently.

Cognaptus: Automate the Present, Incubate the Future.