Reader question: Option markets quote implied volatilities only at discrete strikes and expiries. How can an algorithm fit a smooth smile that is flexible enough for skewed markets but still respects the basic static-arbitrage structure required by option prices?

The Stochastic Volatility Inspired (SVI) parameterization fits total implied variance as a simple hyperbolic function of forward log-moneyness. Five raw parameters control level, slope, skew, centre and curvature for one expiry slice. The hard part is not evaluating the formula—it is calibrating and connecting slices without introducing negative risk-neutral density or calendar-spread arbitrage.

This article owns one precise computational job: SVI/SSVI-style parametric fitting and static-arbitrage diagnostics for implied total-variance smiles and surfaces. It does not own the underlying stochastic-volatility dynamics, Black–Scholes implied-vol inversion, Dupire local-vol dynamics or risk-neutral-density extraction itself.

This is public mathematical and computational education. It is not an options-trading recommendation, a forecast of future volatility or a claim that a static arbitrage-free surface has realistic smile dynamics.

1. Why fit total implied variance rather than raw volatility

For an option with maturity T and Black implied volatility σ_imp, define total implied variance:

w = σ_imp²T.

Total variance is natural because many asymptotic and no-arbitrage results are expressed directly in w(k,T), and because calendar consistency is naturally related to how variance accumulates with maturity.

The fitted object is not volatility itself. After fitting:

σ_SVI(k,T)=√[w(k,T)/T].

2. Forward log-moneyness

For strike K and expiry-matched forward F_T:

k = ln(K/F_T).

Then:

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

Using the forward rather than spot separates the smile coordinate from deterministic carry/discounting.

Falsifier: if two systems use different forward, dividend or discount conventions, they can produce different k values and therefore different SVI parameters from identical option prices.

3. Raw SVI formula

For one maturity, raw SVI is:

w(k)=a+b[ρ(k−m)+√((k−m)²+σ²)].

The raw parameters are:

a: vertical variance level;
b≥0: overall wing slope scale;
−1<ρ<1: asymmetry/skew;
m: horizontal location;
σ>0: smoothness/curvature scale.

These interpretations overlap. Parameters interact, so changing one can often be partly offset by changing another.

4. The minimum total variance

Differentiate the raw SVI function:

w′(k)=b[ρ+(k−m)/√((k−m)²+σ²)].

The minimum occurs where w′=0. Solving gives:

k_min=m−ρσ/√(1−ρ²).

The minimum total variance is:

w_min=a+bσ√(1−ρ²).

A basic admissibility condition is therefore:

a+bσ√(1−ρ²) ≥ 0.

But non-negative total variance alone is far from sufficient for arbitrage freedom.

5. Wing slopes

As k→+∞:

w(k) ~ b(1+ρ)k + constant.

As k→−∞:

w(k) ~ b(1−ρ)|k| + constant.

Thus:

right-wing slope magnitude = b(1+ρ);
left-wing slope magnitude = b(1−ρ).

For equity-index-style negative skew, ρ is often negative, making the left wing steeper than the right.

6. Lee’s moment formula constrains extreme wings

Roger Lee’s model-independent moment formula links asymptotic implied-variance wing growth to the existence of moments of the underlying distribution.

A key consequence is that arbitrage-consistent total-variance wings cannot grow arbitrarily fast; their asymptotic linear slopes obey model-independent bounds.

For raw SVI this turns b(1±ρ) into an important tail diagnostic.

But wing-slope bounds are not sufficient to rule out local butterfly arbitrage inside the quoted strike range.

7. Static butterfly arbitrage means negative convexity of call prices

At a fixed expiry, a no-arbitrage call-price curve C(K) is convex in strike:

∂²C/∂K² ≥ 0.

Under suitable regularity this second derivative is proportional to risk-neutral density.

If an implied-volatility fit produces negative strike convexity, a butterfly combination can have a negative price despite a non-negative payoff.

This is why a visually smooth SVI smile can still be invalid.

See Breeden–Litzenberger density algorithms for the density/strike-convexity owner.

8. Durrleman/Gatheral g(k) condition

For a sufficiently smooth total-variance smile w(k), define:

g(k) = [1 − k w′(k)/(2w(k))]² − [w′(k)²/4][1/w(k)+1/4] + w″(k)/2.

Under the standard forward-normalized Black setting, a non-negative risk-neutral density requires:

g(k) ≥ 0

together with appropriate tail conditions.

This is an essential post-calibration audit.

Falsifier: evaluate g(k) on a dense grid extending beyond quoted strikes. Checking only at market strikes can miss an arbitrage pocket between them.

9. Raw parameter bounds do not guarantee no butterfly arbitrage

Constraints such as:

b≥0, |ρ|<1, σ>0, w_min≥0

make the raw formula well behaved, but they do not guarantee g(k)≥0 for all k.

An optimizer can produce an excellent RMSE smile that violates butterfly convexity.

No-arbitrage testing must therefore be an independent acceptance gate, not assumed from “reasonable-looking” parameters.

10. Calendar-spread arbitrage

Across maturities, a longer-dated option should not become cheaper than a shorter-dated option in a way that violates the standard calendar monotonicity under consistent forwards/discounting.

In total-variance coordinates, a common surface condition is:

∂w(k,T)/∂T ≥ 0

for all relevant k under the consistent forward-moneyness setup.

If each expiry is fitted independently with raw SVI, one slice can cross another between quoted strikes even when every slice is butterfly-safe.

11. Slice-by-slice fitting is therefore incomplete

A production surface needs two gates:

within each maturity: no butterfly arbitrage;
between maturities: no calendar arbitrage.

Independent least-squares fits satisfy neither automatically.

Possible responses include constrained sequential fitting, post-fit repair, SSVI/eSSVI structures, or a global constrained optimization.

12. SSVI connects the maturities

Surface SVI (SSVI) writes a family of slices using ATM total variance θ(T) and a shape function φ(θ).

A common SSVI form is:

w(k,θ)=θ/2 [1+ρφ(θ)k + √((φ(θ)k+ρ)² + 1−ρ²)].

Gatheral and Jacquier derive simple sufficient conditions under which the resulting surface is free of static butterfly and calendar arbitrage.

The broader lesson is more important than memorising one inequality: cross-maturity structure must be designed, not left to accidental alignment of independent slices.

13. Raw SVI versus SSVI/eSSVI

Raw SVI: flexible five-parameter fit per expiry.

SSVI: links slice shape to ATM total variance through a lower-dimensional surface structure.

eSSVI and related extensions: add flexibility while retaining tractable no-arbitrage constraints across expiries.

A less flexible surface can be more robust than many independently perfect slice fits if downstream Greeks/local-volatility calculations require maturity smoothness.

14. Calibration target

Possible objectives include errors in:

total variance w;
implied volatility;
option price;
normalized option price;
bid–ask-scaled residuals.

These weight the market differently.

A one-vol-point error on a deep out-of-the-money option with tiny vega can correspond to a very small price error. Fitting raw IV equally across strikes may therefore overemphasize low-vega wings.

15. Vega weighting

For small perturbations:

ΔC ≈ Vega × Δσ.

A price-oriented objective can therefore weight implied-volatility errors roughly by vega.

Other systems divide price residuals by bid–ask spread to reflect quote confidence.

Falsifier: compare fitted price residuals as well as vol residuals. A beautiful implied-vol RMSE can hide economically large errors near ATM where vega is highest.

16. Quote cleaning is part of calibration

Raw option chains can contain:

stale bids/asks;
crossed markets;
zero bids;
very wide spreads;
bad strikes/expiries;
inconsistent forward/dividend assumptions;
last trades far from current quotes.

SVI cannot distinguish a genuine wing shape from a bad quote.

A robust fit records quote provenance and acceptance/rejection rules before optimization begins.

17. Forward estimation is a hidden weak link

Because k=ln(K/F), a wrong F shifts every x-coordinate of the smile.

In equities, F depends on rates and dividends; in futures options, the relevant forward/futures convention differs; in FX, the domestic/foreign discount curves matter.

Falsifier: infer/check the forward independently from put–call parity where reliable and compare with the forward used to construct k.

18. Parameter interaction and non-uniqueness

a,b,ρ,m,σ can compensate for one another over a finite quoted strike range.

Two parameter vectors may produce almost identical market-range smiles but very different far-wing extrapolation.

This is an identifiability issue.

Diagnostic: use multiple optimizer starts and compare both in-range fit and out-of-range wing behavior. Parameter uniqueness is less important than curve and risk stability.

19. Quasi-explicit/separable calibration ideas

Research and practitioner methods exploit partial algebraic structure to reduce the dimensional nonlinear search.

For example, some parameters can be solved conditionally or constrained analytically while only a smaller subset is optimized nonlinearly.

This improves speed and can reduce local-minimum risk.

But a fast optimizer does not replace no-arbitrage validation.

20. Multi-start optimization

For each expiry:

clean and normalize quotes;
choose parameter bounds;
run many initial seeds;
retain low-loss feasible candidates;
apply butterfly checks;
compare parameter/wing stability;
select the accepted slice;
then perform cross-maturity calendar checks.

If only one initial seed converges to an acceptable fit, the calibration should be flagged as fragile.

21. Dense-grid arbitrage checks

Market strikes may be sparse.

A fitted SVI curve is continuous and defines prices at every intermediate strike.

Therefore evaluate:

w(k)>0;
g(k)≥0;
call-price monotonicity/convexity;
calendar ordering;
wing-slope bounds

on a grid wider and finer than the quote set.

The model is responsible for the unquoted region it creates.

22. Extrapolation is where parameter differences become visible

Inside liquid strikes, many fits can look equally good.

Far from the money, raw SVI becomes asymptotically linear in k.

Thus b and ρ determine the implied total-variance wings.

If wing data are sparse, the extrapolated tail can be driven more by parameter bounds than market evidence.

Falsifier: compare several near-optimal fits and report wing uncertainty rather than a single precise far-tail number.

23. Short maturities are difficult

As T becomes small:

total variance is small;
bid–ask noise can be large relative to w;
discrete dividends/events can distort the smile;
one quote can dominate curvature;
calendar ordering becomes numerically delicate.

Short-dated surfaces need tighter data-quality diagnostics and should not be forced to look like stable long-dated slices.

24. Event expiries

An earnings announcement or other discrete event can create a local term-structure jump in variance.

A surface smoother that assumes gentle maturity evolution may spread event variance into adjacent expiries and create biased prices.

Falsifier: compare forward variance between expiries and known event dates. Large concentrated event variance should be represented explicitly or at least flagged.

25. SVI is not a stochastic-volatility process

Despite the name “Stochastic Volatility Inspired,” raw SVI is primarily a parametric implied-volatility shape.

It does not specify how spot and the smile co-evolve through time.

A surface can be perfectly static-arbitrage-free today and still generate unrealistic hedging P&L if tomorrow’s smile dynamics are modeled poorly.

26. SVI versus Black–Scholes

Black–Scholes option-pricing algorithms own the constant-volatility pricing/inversion framework that defines implied volatility.

SVI takes those strike-by-strike implied volatilities and fits a smooth cross-sectional total-variance smile.

27. SVI versus Heston

A Heston-type model specifies stochastic variance dynamics and produces a smile through risk-neutral pricing.

SVI directly parameterizes the smile without claiming that its five parameters are Heston parameters.

Heston is a dynamics/pricing model; SVI is a surface representation.

28. SVI versus splines

A cubic spline can fit quoted volatility points very closely but may overshoot between knots and violate option-price convexity.

SVI imposes a structured asymptotic shape and offers tractable arbitrage diagnostics.

But SVI may fit less tightly if the market smile contains local features outside its five-parameter shape.

The right comparison is not RMSE alone; it is fit plus arbitrage, stability and downstream risk.

29. SVI versus Carr–Madan/COS

Carr–Madan FFT and Fourier-COS price options from a model characteristic function.

SVI does the inverse-style market representation job: it fits the observed implied-volatility smile, regardless of whether a closed-form characteristic function model sits underneath.

30. Risk-neutral density diagnostics

Once an arbitrage-free SVI slice is converted to option prices, its strike derivatives imply a risk-neutral density.

Useful diagnostics include:

density non-negativity;
density integrates approximately to one under correct normalization;
forward mean consistency;
tail mass sensitivity to wing parameters.

A density spike near a sparse wing quote can reveal an otherwise hidden overfit.

31. Local volatility is derivative-hungry

Dupire local volatility differentiates an option-price surface in both strike and maturity.

Small SVI slice inconsistencies can therefore become large local-volatility artifacts.

A surface intended for local-vol extraction needs:

butterfly safety;
calendar smoothness;
stable maturity interpolation;
controlled wings.

A stack of independently fitted raw SVI slices may be inadequate even if every slice has low RMSE.

32. Inputs and outputs

Inputs can include:

option bid/ask/mid prices or implied vols;
strikes and expiries;
forward per expiry;
discount factors;
day-count/time-to-expiry;
quote-quality flags;
calibration objective/weights;
parameter bounds;
optimizer starts;
arbitrage-grid range/resolution;
cross-maturity fitting rule.

Outputs can include:

raw SVI parameters by expiry;
fitted w(k) and σ(k);
price/vol/variance residuals;
minimum g(k);
wing slopes;
calendar-crossing diagnostics;
risk-neutral-density checks;
parameter stability;
wing uncertainty;
accepted/rejected publish status.

33. Evidence polarity

Evidence for confidence includes:

market price residuals sit within bid–ask noise;
w(k) stays non-negative;
g(k) is non-negative on a dense extended grid;
wing slopes obey theoretical bounds;
calendar checks pass between every expiry pair;
multiple starts produce similar accepted smiles;
parameters/wing shapes are stable across nearby market snapshots;
derived densities are non-negative and well behaved.

Evidence against confidence includes:

excellent RMSE but g(k)<0 between strikes;
slice crossings across maturities;
fit driven by stale or zero-bid quotes;
parameters hit bounds frequently;
near-identical in-range fits have radically different wings;
small quote changes cause large parameter/Greek jumps;
density becomes negative or spiky;
local-vol extraction becomes unstable.

34. Counterexample: perfect quote fit with butterfly arbitrage

Five market strikes are fit almost exactly.

Between two strikes, the implied-vol curve bends so that call prices become locally concave.

Falsifier: compute g(k) and option-price convexity on a dense grid. Market-point RMSE alone cannot detect this failure.

35. Counterexample: independent slices create a calendar crossing

The one-month and two-month slices each pass butterfly tests.

At one deep-out-of-the-money k, the one-month total variance exceeds the two-month total variance.

Falsifier: compare w(k,T) across all maturities on a common dense k grid. Independent slice validity does not imply surface validity.

36. Counterexample: wrong forward creates fake skew change

Spot is unchanged, but the dividend estimate used in F changes incorrectly.

Every strike’s k coordinate shifts and the calibrated m,ρ parameters move sharply.

Falsifier: recompute forwards from independent put–call parity and compare parameter changes after correcting F.

37. Counterexample: one illiquid wing quote controls extrapolation

A far OTM put has a huge bid–ask spread but is given equal weight to ATM options.

The fit steepens the entire left wing to pass close to that one midpoint.

Falsifier: refit using bid–ask-aware/vega-aware weights and perform leave-one-wing-quote-out tests.

38. Counterexample: parameter stability mistaken for surface stability

Two parameter sets differ greatly but produce nearly identical w(k) over liquid strikes.

Falsifier: evaluate fitted curves, Greeks and wings directly. Raw parameter distance is not the correct loss when the parameterization is weakly identified.

39. Counterexample: static arbitrage-free surface, bad dynamics

Today’s SVI surface is perfectly arbitrage-free. Tomorrow the system simply keeps parameters fixed while spot moves, producing implausible sticky behavior and hedging errors.

Falsifier: backtest smile dynamics and hedge P&L. Static consistency is necessary but not sufficient for dynamic risk modeling.

40. Alternatives

SABR: stochastic-volatility-inspired asymptotic smile model widely used in rates/FX, with different parameter/dynamics interpretation.

Heston: explicit stochastic-volatility dynamics and characteristic-function pricing.

Splines with convexity constraints: more local flexibility but different tail behavior.

SSVI/eSSVI: surface-level SVI families with stronger cross-maturity structure.

Direct arbitrage-free price interpolation: fit call prices/densities rather than implied vols.

42. What would falsify confidence?

Confidence should be withdrawn if any accepted slice produces negative density/butterfly arbitrage; if maturities cross in total-variance/price consistency; if recalibration is highly seed-dependent; if wing behavior is dominated by one unreliable quote; if parameter/Greek jumps are not supported by market movement; or if downstream local-vol/density calculations become unstable.

43. Verification and update triggers

Preserve raw quotes, timestamps, forward/discount inputs, time-to-expiry convention, calibration weights, parameter bounds, optimizer starts, residuals, g(k) grid, wing slopes and calendar-check results.

Revalidate when:

forward/dividend methodology changes;
new expiries/strikes appear;
market liquidity deteriorates;
parameters hit bounds;
butterfly/calendar margins approach zero;
wing quotes disappear;
event risk changes the term structure;
surface is reused for local volatility or exotic pricing.

44. Primary and high-quality references

Jim Gatheral and Antoine Jacquier, Arbitrage-Free SVI Volatility Surfaces, Quantitative Finance, 2014.
Roger W. Lee, The Moment Formula for Implied Volatility at Extreme Strikes, Mathematical Finance, 2004.
Gaoyue Guo, Antoine Jacquier, Claude Martini and Leo Neufcourt, Generalized Arbitrage-Free SVI Volatility Surfaces, SIAM Journal on Financial Mathematics.
Zeliade Systems, Quasi-Explicit Calibration of Gatheral’s SVI Model, calibration note.
Jim Gatheral, The Volatility Surface: A Practitioner’s Guide, for the SVI framework and volatility-surface context.

Educational boundary: SVI is a compact language for option-smile geometry. Its real quality is determined by market-data normalization, constrained calibration, butterfly/calendar checks and honest wing uncertainty—not by how smooth the plotted curve looks.

Small Group Tutorials

How SVI Volatility-Smile Algorithms Fit Option Surfaces: Total Variance, Log-Moneyness, Wing Slopes, Butterfly/Calendar Arbitrage, Calibration and Extrapolation Failure