Home › Blog › Quant Research
Arbitrage-Free Volatility Surfaces: SVI and Mixtures of Lognormals
A six-part series on how option markets price volatility, from the textbook model to the arbitrage-free surfaces used by professional desks.
- Black-Scholes and the Greeks
- Implied Volatility and the Smile
- Local Volatility and Dupire's Formula
- Stochastic Volatility: the Heston Model
- The SABR Model
- Arbitrage-Free Surfaces: SVI and Mixtures of Lognormals (you are here)
Every model in this series needs the same thing from the market: a smooth, consistent volatility surface. Local volatility differentiates it, Heston and SABR are calibrated to it, and traders read prices off it. This final part covers what makes a surface "arbitrage-free," and two parameterizations that professional systems use to build one: Jim Gatheral's SVI and the mixture of lognormals behind Bloomberg's OVDV surfaces.
The problem
The market only quotes a finite set of strikes and expirations, and those quotes are noisy, with bid-ask spreads that widen in the wings. A surface has to fill in everything else: interpolating between quotes, extrapolating beyond the last strike, and connecting one expiration to the next. A careless fit can create prices that allow riskless profit, and those errors propagate into every model built on top of it. Part 3 showed how arbitrage in the input surface turns into negative local variance.
What "arbitrage-free" means
Three families of static arbitrage need to be ruled out. Writing k = ln(K / F) for log-moneyness and w(k, T) = σimp² T for total implied variance:
| Arbitrage | Price condition | Meaning |
|---|---|---|
| Call spread | Call prices fall as the strike rises | A higher-strike call can never cost more than a lower-strike call |
| Butterfly | Call prices are convex in strike | The implied probability density is never negative |
| Calendar | Total variance w(k, T) never decreases with T at fixed k | A longer option is never cheaper than an equivalent shorter one |
The butterfly condition can be written directly in terms of the smile. Following Gatheral and Antoine Jacquier, define:
where w′ and w″ are the first and second derivatives with respect to k. The smile is free of butterfly arbitrage when g(k) ≥ 0 everywhere, because the implied density is proportional to g. There is also a constraint on the far wings. Roger Lee's moment formula shows that total variance can grow at most linearly in log-moneyness: w(k) / |k| cannot exceed 2 for very large |k|. A smile with quadratic wings, which a simple polynomial fit would produce, eventually violates this.
SVI: a smile with the right shape
Gatheral introduced the Stochastic Volatility Inspired parameterization at Merrill Lynch in 1999 and published it in 2004. It describes total variance for a single expiration with five parameters:
Each parameter has a geometric meaning. a sets the overall level, b the angle between the two wings, ρ rotates the smile to create skew, m shifts it left or right, and σ controls how rounded the bottom of the smile is. As |k| grows, the wings become straight lines with slopes b(1 + ρ) and b(1 − ρ), which is exactly the linear behavior Lee's formula allows. That is no coincidence: the form was chosen because the large-strike behavior of the Heston model has this shape.
SVI fits typical equity smiles remarkably well with only five numbers, which is why it is used widely for interpolation and as the input to local volatility. It does not guarantee absence of arbitrage on its own, however. Raw SVI parameters can produce a negative g(k) for some strikes, and separately fitted expirations can cross each other. Gatheral and Jacquier's 2014 surface SVI (SSVI) addresses this by tying all expirations together through the at-the-money total variance and a small number of shared parameters, with simple inequalities on those parameters that rule out both butterfly and calendar arbitrage.
Mixture of lognormals: arbitrage-free by construction
A different strategy is to start from a probability distribution rather than a smile shape. If the surface comes from a valid distribution at each expiration, butterfly and call spread arbitrage are impossible by construction. Damiano Brigo, Fabio Mercurio and Francesco Rapisarda developed this idea in the early 2000s: model the risk-neutral density of the price at time T as a weighted sum of lognormal densities.
Each component has its own weight pi, mean ξiF and volatility Σi. The constraints are simple: weights are non-negative and sum to at most one, and the weighted means must equal the forward, Σ piξi = 1, so the distribution prices the forward correctly. If the weights sum to less than one, the missing probability sits at zero, a point mass that represents default. That small detail is powerful for equities, because it lets the model produce the very steep low-strike skew that otherwise requires extreme parameters.
The payoff is that pricing stays inside the Black-Scholes world. A call price is just a weighted sum of Black-Scholes prices:
Retrieving an implied volatility requires only a handful of closed-form evaluations and one inversion, so the method is almost as fast as a simple parametric curve. With two lognormals the smile is already curved; with four, it can match most equity and index smiles closely. Calendar arbitrage is controlled by constraining how the parameters evolve across maturities, for example requiring each component's total variance to increase with time.
This is the approach Bloomberg adopted for its OVDV volatility surface function. A 2011 Bloomberg Markets article by David Frank and Daniele Caporaletti describes a mixture of four lognormals fitted maturity by maturity, with constraints across maturities to limit calendar arbitrage, after testing showed it performed best overall at producing arbitrage-free surfaces that fit within bid-ask spreads. The same article lists stability under small quote changes and near-Black-Scholes retrieval speed as design requirements.
Choosing a method
A good surface has to balance four goals that pull against each other: no arbitrage, stability when a few quotes change, speed of evaluation, and a close fit to market prices, ideally within the bid-ask spread. Simple curve fits are fast but give no guarantees. Full stochastic volatility models are internally consistent but slow to evaluate and often too rigid to fit every smile. Parameterizations such as SSVI and lognormal mixtures sit in between, which is why they dominate production systems.
Series recap
Across six parts, each model answered a weakness of the one before:
| Part | Model or tool | What it adds | Main weakness |
|---|---|---|---|
| 1 | Black-Scholes | Replication pricing and the Greeks | One volatility for all strikes |
| 2 | Implied volatility | A common quoting language and the risk-neutral density | Describes prices, not dynamics |
| 3 | Local volatility | Exact fit to every vanilla option | Unrealistic smile dynamics |
| 4 | Heston | Random, mean-reverting volatility with fast pricing | Weak at very short maturities |
| 5 | SABR | Transparent smile parameters and correct dynamics | One expiry at a time; low-strike arbitrage |
| 6 | SVI and lognormal mixtures | Smooth, arbitrage-free surfaces | Describe the surface, not its evolution |
No single model wins everywhere. Practitioners combine them: an arbitrage-free parameterization to build today's surface, local volatility or stochastic local volatility to price exotics consistently with it, and SABR or Heston intuition to understand how it will move. The Vol Surface tool on this site shows the raw market data that all of these methods start from.
Key takeaways
- A usable surface must avoid call spread, butterfly and calendar arbitrage, and its wings can grow at most linearly in total variance.
- SVI describes each smile with five interpretable parameters and the right wing behavior; SSVI extends it to an arbitrage-free surface.
- A mixture of lognormals starts from a valid probability distribution, so strike arbitrage cannot occur, and prices remain sums of Black-Scholes values.
- Allowing the mixture weights to sum to less than one adds a default probability that helps fit steep equity skews.
- Production systems balance arbitrage-freeness, stability, speed and fit, and usually combine several of the models in this series.