Quintic polynomial from zero
1/5Fit the smile with a polynomial
Forget about choosing an SDE or a stochastic vol model. Take the total variance curve w(k) and fit it directly with a polynomial in log-moneyness. Six coefficients per slice. Done.
The idea is almost offensively simple. Total variance w(k) =σ²·T is a function of log-moneyness k = ln(K/F). Just fit it with a polynomial:
Compare this to SVI, which has five parameters with specific geometric meanings (level, slope, curvature, center, tilt). The quintic has six parameters with no inherent meaning -- they are just polynomial coefficients. What you lose in interpretation, you gain in flexibility.
Each coefficient controls a different aspect of the smile shape: a₀ sets the ATM level. a₁ controls the linear skew. a₂ controls the curvature. Higher-order terms handle asymmetry and fine structure that SVI's fixed shape cannot capture.
SVI is a shaped mold: it can only make smiles of a certain family. The quintic is soft clay: you can form more shapes, but the clay does not know what a smile should look like. You need external discipline (constraints) to prevent it from making nonsensical shapes.
Why quintic?
Degree 5 is the sweet spot. Cubic is too rigid for realistic smiles. Quartic helps but still cannot handle the asymmetry between put and call wings. Septic (degree 7) oscillates. Quintic threads the needle.
Cubic (degree 3): 4 coefficients. Can capture a tilted smile but not the independent curvature of each wing. If the left wing is steep and the right wing is flat, cubic cannot fit both without distorting the center.
Quartic (degree 4): 5 coefficients. Better -- it can handle symmetric curvature -- but still lacks an odd-power term high enough to differentiate the wings cleanly.
Quintic (degree 5): 6 coefficients. The extra fifth-degree term gives independent control over the asymmetry of the wings at the right moneyness range. Real smiles are asymmetric (put wing steeper than call wing in equities and crypto), and quintic captures this without overfitting.
Septic (degree 7) and higher: Too many degrees of freedom. The polynomial starts oscillating between data points, creating spurious bumps and wiggles that are not in the market data. This is the classic bias-variance tradeoff: more flexibility means more overfitting risk.
Look at the comparison above. Click through each degree. The cubic misses the wings. The quartic is close but stiff. The quintic matches. The septic starts wiggling. That visual is the entire argument for degree 5.
Arbitrage constraints on polynomials
Here is the fundamental problem with polynomial smile models: they grow too fast in the wings. Roger Lee's moment formula says total variance must grow at most linearly in |k| as |k| goes to infinity. A polynomial of degree 5 grows like k⁵. That is a problem.
Lee's moment formula (2004) establishes the asymptotic behavior of implied volatility:
The chart above shows the difference starkly. SVI's wings are bounded: they approach a linear slope. The quintic's wings explode. In the far wings, the polynomial quotes implied vols that imply negative butterfly spreads -- free money.
The fix: use the quintic only in the interior of the smile (say, |k| < 0.5) and blend into a wing model (linear or SVI-like) for extrapolation. This is the standard production approach: polynomial interior, controlled wings.
Alternatively, you can add explicit constraints during the fit:
1. w(k) ≥ 0 for all k (variance must be positive).
2. w(k) is convex in the interior (no butterfly arbitrage -- this is Durrleman's condition).
3. w(k)/|k| ≤ 2 at the endpoints of the fitting range.
These constraints are all linear or quadratic in the coefficients, so they can be enforced by solving a constrained least-squares problem (quadratic program) instead of unconstrained least squares.
Calibration is just linear regression
Unlike SVI's nonlinear optimization (which requires initialization, iterates, and can get stuck in local minima), fitting a polynomial is a linear least-squares problem. Set up a matrix, solve one linear system, done.
Given N observed data points (kᵢ, wᵢ), the problem is:
Drag the data points above. The fit updates instantly because it is just a matrix solve -- no iterations, no convergence issues, no initialization sensitivity. Compare this to SVI calibration, where the optimizer can take dozens of iterations and might find a different answer depending on where you start.
Adding constraints: If you add the arbitrage constraints from the previous section (positivity, convexity, wing bounds), the problem becomes a quadratic program (QP) instead of unconstrained least squares. QPs are still fast and well-studied -- solvers handle them in milliseconds. The key point: the constrained quintic is still dramatically faster to calibrate than SVI.
Numerical stability: The Vandermonde matrix can be ill-conditioned when the moneyness range is wide. Standard remedies: (1) scale k to [-1, 1] before fitting, (2) use orthogonal polynomials (Chebyshev, Legendre) instead of raw powers. These are routine numerical analysis techniques.
Quintic vs SVI
Neither one wins everywhere. The quintic is faster to fit and more flexible in the interior. SVI has bounded wings and interpretable parameters. Know which one to reach for.
Quintic wins when:
1. You need fast calibration (thousands of slices per second for a real-time surface). The linear solve is unbeatable on speed.
2. The observed smile has features that SVI's fixed shape cannot match -- local bumps, unusual curvature, asymmetric wings. Quintic is more flexible in the interior.
3. You are working in the interior of the smile (|k| < 0.3) where wing behavior does not matter and you want the tightest possible fit to observed data.
SVI wins when:
1. You need reliable wing extrapolation. SVI's asymptotic linearity in the wings is correct by construction. Quintic must be clipped or blended.
2. You want interpretable parameters for risk management. SVI's a (level), b (angle), ρ (tilt), m (center), σ (wing smoothing) map directly to observable smile features.
3. You are building a surface across expiries. SSVI extends SVI to the full surface with no-arbitrage guarantees. There is no standard "surface quintic" with the same guarantees.
The production compromise: Many desks use both. Quintic for fast interior interpolation and real-time quoting. SVI or SSVI for the official surface, wing extrapolation, and risk reports. The quintic handles the data-dense center; SVI handles the sparse wings.
The quintic polynomial is not a model of the market. It is a curve-fitting tool. It says nothing about dynamics, hedging, or why the smile has the shape it does. SVI is also a curve-fitting tool, but one with enough structure to extend to a surface. For actual dynamics, you need SABR, Heston, or a stochastic local vol model. The quintic lives in the space between raw data and a real model -- it is the fastest way to get a smooth, interpolated smile from noisy observations.
Where to go next:
SVI Parameterization -- the standard smile model with bounded wings
SSVI Surface -- SVI extended to the full surface with no-arbitrage guarantees
Interpolation Methods -- all fitting methods compared