SVI Parametrisation

Setup

Market Context

The implied vol surface $\sigma_{\mathrm{imp}}(K, T)$ is the complete language of the options market. Every option price, hedge ratio, and P&L scenario is rooted in this surface. Calibrating it accurately and consistently is a core task on any derivatives desk.

A vol surface must satisfy two properties to be financially meaningful:

1. Arbitrage-free: no calendar spread, butterfly, or call spread arbitrage is implied.
2. Smooth: the surface must be differentiable enough to produce well-defined local vols via the Dupire formula, and to compute stable vega ladders.

Fitting a surface point-by-point (interpolating market quotes directly) satisfies neither: it typically produces arbitrage-violating kinks and is unstable off-grid. A parametric smile model is used instead.

Conventions

Throughout this module:

• Total implied variance: $w(k, T) = \sigma_{\mathrm{imp}}^2(k, T) \cdot T$, where $k = \ln(K/F)$ is the log-moneyness and $F = S e^{(r-q)T}$ is the forward.
• Working with $w$ instead of $\sigma$ simplifies arbitrage conditions (they become inequalities on $\partial w / \partial k$ and $\partial^2 w / \partial k^2$).
• A vol slice at fixed $T$ is an implied vol smile. A collection of smiles across maturities is the vol surface.

Theory: The Raw SVI Parametrisation

Motivation: Derman-Kani and the Wings

At extreme log-moneyness $|k| \to \infty$, the implied vol smile must grow roughly linearly in $|k|$. This follows from Lee's moment formula: if the $p$-th moment of the stock price is finite under $\mathbb{Q}$, the right tail of $\sigma_{\mathrm{imp}}$ grows at most as $\sqrt{|k|/T}$ times a function of $p$. More precisely:

$\limsup_{k \to +\infty} \frac{\sigma_{\mathrm{imp}}^2(k, T) \cdot T}{|k|} \leq \psi(\tilde{p}),$

where $\tilde{p}$ is the critical moment. Any smile that grows faster than $\sqrt{|k|}$ in total variance violates moment bounds and implies arbitrage. The SVI form is designed to reproduce this linear-in-$|k$ wing behaviour exactly.

Raw SVI

Gatheral (2004) proposed the Stochastic Volatility Inspired (SVI) parametrisation. For a fixed maturity $T$, the total implied variance as a function of log-moneyness $k$ is:

$w(k) = a + b\!\left[\rho(k - m) + \sqrt{(k-m)^2 + \sigma^2}\right],$

where the five raw SVI parameters $(a, b, \rho, m, \sigma)$ satisfy:

• $a \in \mathbb{R}$: overall level of total variance (vertical shift).
• $b \geq 0$: slope of the wings; controls how fast variance grows with $|k - m|$.
• $\rho \in (-1, 1)$: correlation-like parameter; controls the asymmetry between left and right wings (skew).
• $m \in \mathbb{R}$: location of the ATM vertex (horizontal shift); usually $\approx 0$.
• $\sigma > 0$: smoothness of the vertex; larger $\sigma$ gives a wider, rounder bottom.

Geometric Interpretation

The graph of $w(k)$ is a rotated hyperbola with:

• Asymptotes: as $k \to +\infty$, $w(k) \approx a + b(1+\rho)(k-m)$ (right wing slope $b(1+\rho)$); as $k \to -\infty$, $w(k) \approx a + b(\rho-1)(k-m)$ (left wing slope $b(\rho-1) < 0$, so the wing rises).
• Vertex: at $k = m$, $w(m) = a + b\sigma$. This is the minimum of the smile (for the call side) when $\rho = 0$.
• ATM vol: $\sigma_{\mathrm{ATM}} = \sqrt{w(0)/T}$. Since $m \approx 0$ in typical calibrations, $w(0) \approx a + b\sigma$.

Note: the left wing slope $b(\rho - 1)$ is always negative (wings rise on both sides) since $b \geq 0$ and $\rho < 1$. The right wing slope $b(1 + \rho)$ is always non-negative. The asymmetry between wings is controlled by $\rho$: negative $\rho$ (typical for equities) steepens the left wing and flattens the right, consistent with the negative equity skew.

Arbitrage-Free Conditions

A slice $w(k)$ is free of static arbitrage if and only if:

Condition 1: No Butterfly Arbitrage

Butterfly arbitrage is absent if and only if the function

$g(k) = \left(1 - \frac{k\, w'(k)}{2w(k)}\right)^2 - \frac{(w'(k))^2}{4}\!\left(\frac{1}{w(k)} + \frac{1}{4}\right) + \frac{w''(k)}{2} \geq 0 \quad \text{for all } k,$

where $w' = dw/dk$ and $w'' = d^2w/dk^2$. This condition (Gatheral 2004, following Dupire) ensures the local variance implied by the Dupire formula is non-negative:

$\sigma_{\mathrm{loc}}^2(k, T) = \frac{\partial w / \partial T}{g(k)}.$

If $g(k) < 0$ for some $k$, the local variance is negative — an immediate arbitrage.

For the raw SVI form, $g(k) \geq 0$ is not automatic: it must be checked or enforced by constraining the parameters.

Necessary condition (Lee's bound): A weaker, necessary condition is:

$0 \leq w'(k) \leq \frac{4}{k^2}\left(1 + \sqrt{1 - k w'(k)/2}\right) \quad \text{for all } k \neq 0.$

A sufficient condition for no butterfly arbitrage that is easier to check analytically: for the raw SVI form, $b(1 + |\rho|) \leq 4/T$ (Roper 2010).

Condition 2: No Calendar Spread Arbitrage

The total implied variance must be non-decreasing in maturity for each fixed log-moneyness:

$w(k, T_1) \leq w(k, T_2) \quad \text{for all } k, \quad T_1 \leq T_2.$

If this fails at any point, a calendar spread (long far-dated call, short near-dated call at the same log-moneyness) is riskless arbitrage.

Condition 3: No Call Spread Arbitrage

Total variance must satisfy:

$w'(k) > -2 \quad \text{for all } k.$

(Equivalent to the call price being non-increasing in strike.)

Jump-Wings Parametrisation

The raw SVI parameters $(a, b, \rho, m, \sigma)$ are poorly conditioned for optimisation: small changes in $m$ and $a$ can produce large changes in the smile shape, and the parameters are correlated. The jump-wings (JW) parametrisation (Gatheral and Jacquier 2014) reparametrises in terms of quantities with direct financial meaning:

$v_T = w(0)/T \quad (\text{ATM total variance divided by T, i.e., ATM var}),$ $\psi_T = \frac{1}{2\sqrt{w(0)}} \frac{dw}{dk}\bigg|_{k=0} \quad (\text{normalised ATM slope, related to skew}),$ $p_T = \frac{1}{\sqrt{w(0)}} \left.\frac{d^2w}{dk^2}\right|_{k=0}^{1/2} \cdot \tfrac{1}{2} \quad (\text{left wing slope}),$ $c_T = b(1 + \rho) \quad (\text{right wing slope}),$ $\tilde{v}_T = a/T + b\sigma/T \cdot \ldots \quad (\text{minimum variance}).$

The exact mapping is:

$b = (p_T + c_T)/2, \quad \rho = 1 - 2p_T/(p_T + c_T),$ $\beta = \rho - 2\psi_T / b, \quad m = \psi_T \cdot T / (\sqrt{w(0)} \cdot b \cdot \beta),$ $\sigma = \sqrt{\max(\beta^2 - 1, 0)} \cdot m \quad \text{(approximation for } |\beta| > 1\text{)}.$

The JW parameters have the advantage that $v_T$, $p_T$, and $c_T$ are directly related to observable quantities (ATM vol, put skew, call skew), making initial guesses intuitive and the parameter space more orthogonal.

SSVI: Extending to a Surface

Fitting each maturity slice independently with raw SVI can produce calendar spread arbitrage across slices. Surface SVI (SSVI) (Gatheral and Jacquier 2014) extends the parametrisation to the full surface:

$w(k, \theta_t) = \frac{\theta_t}{2}\left\{1 + \rho_\infty \phi(\theta_t) k + \sqrt{(\phi(\theta_t) k + \rho_\infty)^2 + (1 - \rho_\infty^2)}\right\},$

where:

• $\theta_t = \sigma_{\mathrm{ATM}}^2 \cdot t$: the ATM total variance term structure.
• $\phi: \mathbb{R}_{>0} \to \mathbb{R}_{>0}$: a smooth function (e.g., power law $\phi(\theta) = \eta / (\theta^\gamma (1 + \theta)^{1-\gamma})$).
• $\rho_\infty \in (-1, 1)$: a single surface-level correlation parameter.

Sufficient conditions for SSVI to be arbitrage-free (Gatheral-Jacquier):

1. $\partial \theta_t / \partial t > 0$ (ATM variance term structure is increasing in $t$).
2. $\partial / \partial \theta_t [\theta_t \phi(\theta_t)] \geq 0$.
3. $\theta_t \phi(\theta_t)(1 + |\rho_\infty|) < 4$.

These conditions are sufficient but not necessary. For the power-law $\phi$: conditions reduce to simple bounds on $\eta$, $\gamma$, and $\rho_\infty$, making it easy to enforce during calibration.

Calibration Procedure

Single-Slice Calibration

For each maturity $T_j$:

1. Collect market quotes: $\{(k_i, w_i^{\mathrm{mkt}})\}_{i=1}^{N_j}$ (usually 5–10 strikes per slice).
2. Initialise parameters using JW: set $v_T = \sigma_{\mathrm{ATM}}^2$, estimate $\psi_T$ from the 25Δ skew, estimate wing slopes $p_T$, $c_T$ from risk-reversals and butterfly spreads.
3. Minimise $\sum_i (w_{\mathrm{SVI}}(k_i) - w_i^{\mathrm{mkt}})^2$ subject to $b \geq 0$, $|\rho| < 1$, $\sigma > 0$.
4. Check arbitrage conditions post-calibration; re-optimise with constraint if violated.

import numpy as np
from scipy.optimize import minimize, Bounds

def svi_total_var(k: np.ndarray, a: float, b: float, rho: float, m: float, sigma: float) -> np.ndarray:
    """Raw SVI total implied variance w(k) = a + b*(rho*(k-m) + sqrt((k-m)^2 + sigma^2))."""
    return a + b * (rho * (k - m) + np.sqrt((k - m)**2 + sigma**2))

def svi_butterfly_condition(k: np.ndarray, a, b, rho, m, sigma) -> np.ndarray:
    """
    Computes g(k) = (1 - k*w'/(2w))^2 - (w')^2/4*(1/w + 1/4) + w''/2.
    g(k) >= 0 for all k is the no-butterfly-arbitrage condition.
    """
    eps  = 1e-12
    dkm  = k - m
    sq   = np.sqrt(dkm**2 + sigma**2)
    w    = a + b * (rho * dkm + sq)
    wp   = b * (rho + dkm / sq)                        # dw/dk
    wpp  = b * sigma**2 / (sq**3)                       # d^2w/dk^2
    g    = (1 - k * wp / (2 * w + eps))**2 \
           - (wp**2) / 4 * (1 / (w + eps) + 0.25) \
           + wpp / 2
    return g

def calibrate_svi_slice(
    log_moneyness: np.ndarray,
    market_total_var: np.ndarray,
    weights: np.ndarray | None = None
) -> dict:
    """
    Calibrate raw SVI parameters (a, b, rho, m, sigma) to a single maturity slice.

    Inputs:
        log_moneyness:    k = log(K/F), array shape (N,)
        market_total_var: w_mkt = sigma_imp^2 * T, array shape (N,)
        weights:          per-instrument weights (default: uniform)
    Returns:
        dict with keys: a, b, rho, m, sigma, rms_error, converged
    """
    if weights is None:
        weights = np.ones_like(log_moneyness)

    def objective(x: np.ndarray) -> float:
        a, b, rho, m, sigma = x
        w_model = svi_total_var(log_moneyness, a, b, rho, m, sigma)
        return float(np.sum(weights * (w_model - market_total_var)**2))

    # Initialise from market observables
    w_atm  = float(np.interp(0.0, log_moneyness, market_total_var))
    skew   = float(np.gradient(market_total_var, log_moneyness)[len(log_moneyness)//2])
    x0     = [w_atm * 0.9, 0.1, -0.3, 0.0, 0.1]

    bounds = Bounds(
        lb=[-1.0,  1e-6, -0.999, -1.0,  1e-6],
        ub=[ 1.0,  2.0,   0.999,  1.0,  5.0 ],
    )

    result = minimize(
        objective, x0,
        method='L-BFGS-B', bounds=bounds,
        options=dict(ftol=1e-14, gtol=1e-10, maxiter=2000),
    )

    a, b, rho, m, sigma = result.x
    rms = np.sqrt(result.fun / len(log_moneyness))

    # Post-calibration butterfly check
    k_dense = np.linspace(log_moneyness.min() - 1.0, log_moneyness.max() + 1.0, 500)
    g_min   = float(np.min(svi_butterfly_condition(k_dense, a, b, rho, m, sigma)))

    return dict(a=a, b=b, rho=rho, m=m, sigma=sigma,
                rms_error=rms, converged=result.success,
                butterfly_g_min=g_min)

Detecting Calendar Spread Arbitrage

Check that $w(k, T_j) \leq w(k, T_{j+1})$ for each grid point $k$ across adjacent maturities:

def check_calendar_arbitrage(
    k_grid: np.ndarray,
    svi_params_by_maturity: list[dict]
) -> list[tuple[int, float]]:
    """
    Returns list of (maturity_index, log_moneyness) pairs where calendar arbitrage occurs.
    """
    violations = []
    for j in range(len(svi_params_by_maturity) - 1):
        p_near = svi_params_by_maturity[j]
        p_far  = svi_params_by_maturity[j + 1]
        w_near = svi_total_var(k_grid, **{k: p_near[k] for k in ('a','b','rho','m','sigma')})
        w_far  = svi_total_var(k_grid, **{k: p_far[k]  for k in ('a','b','rho','m','sigma')})
        bad_k  = k_grid[w_far < w_near]
        for k_val in bad_k:
            violations.append((j, float(k_val)))
    return violations

Limitations

Parameter unidentifiability at sparse strikes. With only 5 market quotes per slice (a typical single-broker strip), the raw SVI has 5 parameters and no degrees of freedom. The fit will be exact but potentially arbitrage-ridden or non-unique. Use regularisation or a lower-dimensional parametrisation (e.g., fixing $m = 0$).

SSVI loses slice flexibility. SSVI imposes a global correlation $\rho_\infty$ across all maturities, which may not fit markets where the skew term structure is complex. In practice, SSVI parameters are time-dependent, breaking the simple global surface structure.

Extrapolation beyond quoted strikes. SVI is well-specified by construction in the wings (linear growth in $|k|$). But the wing slope parameters $b(1 \pm \rho)$ must be consistent with moment constraints — large $b$ combined with large $|\rho|$ can violate Lee's bounds for short maturities.

Short-maturity smile. SVI was designed for maturities $T \geq 1$ month. For very short maturities ($T < 1$ week), the smile can be very steep and SVI's smooth wing structure may not capture the convexity correctly. Normalised Bachelier vols are sometimes used instead.

Calibration to bid-ask midpoints. Market implied vols are quoted with a bid-ask spread. Fitting the midpoint produces a parametric smile that may lie inside the bid-ask cone even when the fit is "good". A more robust approach is to minimise the sum of violations outside the bid-ask cone, not the midpoint residuals.

Interview Angle

L1. Write down the raw SVI formula. Explain the role of each parameter intuitively. What does negative $\rho$ imply about the shape of the vol smile, and why is this consistent with equity market stylised facts?

Raw SVI: $w(k) = a + b[\rho(k-m) + \sqrt{(k-m)^2 + \sigma^2}]$. $a$: overall variance level. $b$: wing steepness. $\rho$: skew (negative → left wing steeper than right). $m$: smile centring. $\sigma$: smile curvature at ATM.

Negative $\rho$: For equities, $\rho < 0$ corresponds to a downward-sloping smile (higher vol for lower strikes, lower vol for higher strikes) — the classical equity skew. This reflects the leverage effect (stock prices and vol are negatively correlated) and crash risk demand (OTM puts are expensive). For FX, $\rho \approx 0$ and the smile is more symmetric (higher vol on both wings).

L2. State the no-butterfly-arbitrage condition for a vol slice in terms of total variance $w(k)$. What is $g(k)$ and why must it be non-negative? Derive the connection to the Dupire local variance.

Butterfly condition: The Dupire local variance at log-moneyness $k$ and maturity $T$ is:

$\sigma_{\mathrm{loc}}^2 = \frac{\partial_T w}{g(k)},$

where $g(k) = (1 - kw'/(2w))^2 - (w')^2/4 \cdot (1/w + 1/4) + w''/2$. Since $\partial_T w \geq 0$ (calendar spread condition), $g(k) \geq 0$ is required for $\sigma_{\mathrm{loc}}^2 \geq 0$. Negative $g$ means an implied negative local variance — a butterfly arbitrage. $g(k) = 0$ corresponds to a butterfly-arbitrage boundary; the local density collapses to zero at that strike.

L3. Explain the SSVI parametrisation. What are the conditions for SSVI to be globally arbitrage-free across the full surface? How does the choice of $\phi$ function affect the smile dynamics?

SSVI: $w(k,\theta_t) = \frac{\theta_t}{2}[1 + \rho\phi(\theta_t)k + \sqrt{(\phi k + \rho)^2 + 1 - \rho^2}]$. The ATM variance $\theta_t$ varies with $T$; $\phi(\theta_t)$ controls how the smile wing slope scales with ATM level.

Arbitrage-free conditions: (1) $\partial_t \theta_t > 0$; (2) $\partial_\theta [\theta\phi(\theta)] \geq 0$; (3) $\theta\phi(\theta)(1 + |\rho|) < 4$. Condition (3) bounds the wing slope from Lee's moment formula.

$\phi$ choice: The power law $\phi(\theta) = \eta/(\theta^\gamma(1+\theta)^{1-\gamma})$ gives: for $\gamma = 1/2$, SVI smiles are self-similar under maturity scaling (sticky-moneyness dynamics). The parameter $\gamma$ controls the smile-to-vol ratio: for $\gamma \approx 1$, the smile slope is approximately independent of ATM level (sticky-strike dynamics); for $\gamma \approx 0$, the smile slope scales with $1/\sqrt{\theta_t}$ (sticky-delta). Choosing $\gamma$ is a statement about smile dynamics and should be calibrated to the empirical relationship between ATM vol level and smile curvature.