How SVI Volatility-Surface Algorithms Fit Option Smiles: Total Variance, Log-Moneyness, Five Parameters, Static-Arbitrage Tests and Calibration Failure

Reader question: An option market quotes implied volatilities only at a finite set of strikes and expiries. How can a pricing system turn those scattered observations into a smooth surface without inventing impossible option prices between the quoted points?

One widely used answer is SVI — Stochastic Volatility Inspired. For each maturity, SVI fits a five-parameter curve to total implied variance as a function of log-moneyness. The curve is flexible enough to reproduce skew and smile shapes, has controlled asymptotic wings, and can be embedded in an arbitrage-aware surface construction.

But “the optimizer found a low least-squares error” is not enough. An SVI slice can fit market quotes beautifully and still create negative risk-neutral densities or cross-maturity calendar arbitrage. The real algorithm is therefore fit + constraints + static-arbitrage tests + out-of-sample diagnostics.

What this page owns — and what it does not

This page owns the transformation:

option quotes → implied volatility → total variance → SVI parameters → smooth smile/surface + arbitrage diagnostics.

It does not replace Black–Scholes pricing, which maps volatility to option value; implied-volatility inversion, which extracts each market volatility point; or Heston calibration, which fits a dynamic stochastic-volatility model.

This is public derivatives mathematics, not a trading recommendation.

Why fit total variance rather than volatility directly?

For maturity T and implied volatility σ_imp, define total implied variance:

w = σ_imp²T.

SVI is naturally written in total-variance space. This is useful because static-arbitrage conditions across maturity are easier to express in total variance, and asymptotic wing behaviour has a clean relationship with log-moneyness.

Use log-moneyness, not raw strike

For strike K and forward F_T:

k = ln(K/F_T).

Then:

• k = 0 is at-the-money-forward;
• k < 0 is below the forward;
• k > 0 is above the forward.

Using forward moneyness separates much of the carry and interest-rate structure from the smile shape. A surface built from inconsistent forwards can therefore generate fake skew even if the option prices themselves are correct.

The raw SVI equation

The common raw parameterization is:

w(k)=a+b[ρ(k−m)+√((k−m)²+σ²)]

with five parameters:

• a — vertical level;
• b — overall wing slope scale;
• ρ — asymmetry/skew parameter;
• m — horizontal location;
• σ — curvature/width around the centre.

Usual basic admissibility includes b ≥ 0, |ρ| < 1, σ > 0, and non-negative minimum total variance. Those conditions make the curve sensible but do not by themselves guarantee absence of butterfly arbitrage.

What each parameter does

a shifts the whole total-variance slice up or down.

b controls how rapidly the wings rise.

ρ tilts the smile. Negative ρ commonly corresponds to the steeper downside skew seen in equity-index options.

m moves the centre of curvature away from zero log-moneyness.

σ controls how rounded or sharp the central transition is.

The parameters interact strongly. A change in a can be partly offset by b and σ; a change in m can interact with ρ. This creates calibration-identifiability problems when strikes are sparse.

Asymptotic wings

For large positive and negative k, raw SVI approaches linear total-variance wings. Their slopes are governed by approximately:

right wing slope = b(1+ρ)

left wing slope = b(1−ρ).

This feature is one reason SVI is more stable than fitting high-order polynomials to volatility. Polynomials can explode or oscillate outside the observed strikes; SVI has controlled linear-in-log-moneyness asymptotics.

Lee’s moment constraints matter in the far wings

Implied-volatility theory constrains how fast total variance may grow in the extreme wings if option prices come from a valid probability distribution. Gatheral’s SVI framework is closely connected to Lee’s moment formula.

A calibration that produces unrealistically steep asymptotic slopes may fit the quoted strike range yet imply impossible or pathological tail behaviour outside it.

Step 1: clean the market quotes

Before fitting, a surface engine should reject or flag:

• option prices below intrinsic value;
• crossed or inverted bid/ask markets;
• stale zero-volume quotes where policy excludes them;
• incorrect contract multipliers or strike scales;
• bad forward or discount inputs;
• duplicate strike records with inconsistent prices.

SVI cannot repair arbitrary bad market data merely by smoothing it.

Step 2: invert prices to implied volatility

Each cleaned market price is converted to implied volatility using a root solver such as the one described in the implied-volatility solver article.

The solver should preserve bid and ask volatilities where available, because a fitted mid-volatility that sits outside the executable bid/ask range is a data-quality warning.

Step 3: transform to total variance

For each quote:

w_i=σ_i²T.

Then compute:

k_i=ln(K_i/F_T).

The calibration dataset becomes points (k_i,w_i).

Step 4: minimize a weighted error

A simple objective is:

L(θ)=Σ ω_i[w_SVI(k_i;θ)−w_mkt,i]².

The weights ω_i can reflect bid/ask width, vega, liquidity or a deliberate balance between ATM fit and wing fit.

Equal weights are transparent but can let a few noisy far-wing quotes distort the central smile. Pure vega weights can underweight wings so heavily that extrapolation becomes unstable. Weight design is therefore part of the model specification.

Price errors and volatility errors are not interchangeable

A one-volatility-point error on a high-vega ATM option can correspond to a large price error, while the same volatility error in a very low-vega deep-out-of-the-money option can have little cash impact.

A good validation pack reports both volatility-space and price-space residuals.

Butterfly arbitrage

For a fixed maturity, European call prices must be convex in strike. The second strike derivative is connected to the risk-neutral density. If the fitted smile creates a negative density, a butterfly portfolio can imply static arbitrage.

Gatheral and Jacquier derive explicit SVI conditions and diagnostics for arbitrage-free construction. In practical implementations, one commonly evaluates the SVI-specific density condition across a dense log-moneyness grid rather than assuming a low calibration error guarantees convexity.

Calendar arbitrage

Across maturities, a longer-dated option surface should not create an impossible decrease in option value purely from added time. In total-variance coordinates, an arbitrage-aware surface construction checks maturity consistency carefully, often requiring total variance to evolve monotonically in the appropriate forward-moneyness representation.

Fitting every maturity independently can therefore create a surface where each slice is individually smooth but neighbouring maturities cross.

SVI versus SSVI

Raw SVI fits one maturity slice. Surface SVI (SSVI) parameterizes the whole surface with a smaller cross-maturity structure and explicit sufficient conditions for absence of static arbitrage under suitable parameter constraints.

Raw SVI is flexible and local. SSVI sacrifices some independent freedom to improve global consistency.

Sequential versus global calibration

Sequential calibration fits one expiry at a time and then repairs or constrains cross-maturity inconsistencies.

Global calibration solves many maturities jointly with penalties or constraints for calendar and butterfly arbitrage.

Sequential methods are fast and easier to diagnose. Global methods can produce a more coherent surface but create a larger non-linear optimization problem.

Initialization and local minima

The SVI objective is non-linear. Different starting values can converge to different parameter sets with similar residual error.

A robust engine can use:

• multiple initial seeds;
• bounded parameters;
• warm starts from the previous day;
• coarse global search followed by local optimization;
• parameter continuity penalties across maturity.

Warm starts help — until the regime changes

Yesterday’s parameters are often excellent starting values today. But during a crash, earnings event or volatility regime shift, insisting on yesterday-like parameters can prevent the optimizer from following the market.

Warm starts should accelerate convergence, not become hidden priors that suppress real surface changes.

Inputs and outputs

An SVI calibration engine can require:

• option bid/ask or mid prices;
• strike and maturity;
• forward and discount factors;
• implied-volatility convention;
• quote-cleaning policy;
• calibration weights;
• parameter bounds;
• arbitrage constraints or penalties;
• strike/maturity grids for diagnostics.

Outputs can include raw SVI parameters per maturity, fitted total variance, fitted implied volatility, price residuals, butterfly-arbitrage flags, calendar-arbitrage flags, wing slopes and parameter-stability diagnostics.

Evidence polarity

Evidence for confidence includes fitted values inside bid/ask for liquid strikes, small price residuals, non-negative density diagnostics, no cross-maturity calendar violations, stable parameters under small quote perturbations and sensible wing slopes.

Evidence against confidence includes excellent least-squares fit but negative density in unquoted strikes, parameter jumps from tiny quote changes, calendar crossings between expiries, wings dominated by one stale quote or fitted vols outside liquid bid/ask ranges.

Counterexample: perfect interpolation can create arbitrage

A spline can pass exactly through every quoted implied volatility and still oscillate between strikes enough to produce negative butterfly density.

Exact fit is not the same as valid surface geometry.

Counterexample: smooth slices can create an invalid surface

Every maturity may have a beautiful SVI curve, but the six-month total variance can fall below the three-month curve at some log-moneyness. Slice quality does not guarantee surface quality.

Counterexample: stable parameters do not guarantee stable prices

Because parameters interact nonlinearly, a small movement in a wing-slope parameter can materially change deep-out-of-the-money option values while ATM residuals remain tiny.

Weak links

Wrong forward. Log-moneyness is calculated from stale or inconsistent forwards.

Price-bound failure. Impossible option quotes are inverted anyway.

Overweight stale wings. Illiquid quotes dominate the optimizer.

Parameter degeneracy. Several parameter sets explain sparse strikes similarly.

slice-only validation. Calendar arbitrage is never tested.

coarse arbitrage grid. A narrow butterfly violation between grid nodes is missed.

extrapolation blindness. Calibration is judged only inside quoted strikes.

Diagnostics

• bid/ask coverage: count liquid fitted points inside executable ranges.
• price replay: convert fitted vols back to option prices and compare with market prices.
• dense butterfly scan: evaluate static-arbitrage conditions well beyond quoted strikes.
• calendar scan: compare adjacent maturities over a common log-moneyness grid.
• parameter perturbation: bump quotes within bid/ask and measure parameter stability.
• seed sensitivity: rerun from multiple starting points.
• wing test: inspect asymptotic slopes and far-strike extrapolation.
• leave-one-out test: remove a quote and see whether the surface predicts it plausibly.
• regime comparison: compare current parameters with historical distributions without forcing continuity.

What would falsify confidence?

Confidence should be withdrawn if the fitted surface implies negative risk-neutral density, produces calendar arbitrage, requires extreme parameter changes for tiny quote perturbations, cannot reproduce liquid option prices within realistic spreads, or creates uncontrolled wing behaviour outside the calibration range.

Alternatives

Cubic or monotone splines are simpler but need separate arbitrage controls. SSVI gives stronger cross-maturity structure. SABR is natural for many rate smiles. Heston supplies dynamic stochastic-volatility behaviour rather than only a static surface. Machine-learning surfaces can interpolate complex data but still require explicit no-arbitrage constraints.

How this connects to the surrounding estate

Black–Scholes supplies the implied-volatility coordinate system. The implied-volatility solver turns quotes into calibration points. Heston calibration tests a dynamic alternative. The SVI surface can also become input to a Dupire local-volatility extraction.

Verification and update triggers

Preserve quote snapshots, forward/discount inputs, calibration weights, optimizer version, parameter bounds and arbitrage-test grid. Revalidate after exchange-contract changes, quote-source changes, optimizer/library upgrades, major volatility regime shifts or persistent differences between fitted and executable option prices.

Primary and high-quality references

• Jim Gatheral and Antoine Jacquier, Arbitrage-free SVI volatility surfaces, Quantitative Finance 14(1), 2014.
• Gatheral and Jacquier, Arbitrage-free SVI volatility surfaces, open preprint.
• Jim Gatheral, The Volatility Surface: A Practitioner’s Guide.
• Gaoyue Guo, Antoine Jacquier, Claude Martini and Leo Neufcourt, Generalized Arbitrage-Free SVI Volatility Surfaces, SIAM Journal on Financial Mathematics.

Educational boundary: This article explains volatility-surface fitting and static-arbitrage diagnostics. It does not recommend option positions or trading strategies.