← All writing
Note · Quantitative finance

Where the volatility smile comes from

Black–Scholes prices every strike with one volatility; markets don't. Let the variance itself wander and correlate with the price, as Heston did, and the smile appears, with one parameter for its tilt and one for its curve.

Nov 20266 min readOptionsStochastic volatility

The heat-equation note priced an option with one number for volatility. Ask the market for that number by inverting its option prices, and you get a different answer at every strike: plotted against strike, implied volatility bends into a smile or slopes into a skew. A constant σ\sigma cannot produce that. The GARCH note already showed volatility is not constant; the Heston model [1] makes it random in continuous time and keeps option prices computable.

The model

Price and variance each follow their own diffusion, and the two noises are correlated:

dS=μS dt+v S dW1,dv=κ(θ−v) dt+ξv dW2,d⟨W1,W2⟩=ρ dt.\begin{aligned} dS &= \mu S\,dt + \sqrt{v}\, S\, dW_1, \\[1ex] dv &= \kappa(\theta - v)\,dt + \xi \sqrt{v}\, dW_2, \qquad d\langle W_1, W_2\rangle = \rho\, dt . \end{aligned}

The variance is pulled toward a long-run level θ\theta at speed κ\kappa, like the Ornstein–Uhlenbeck process, but its noise shrinks as v\sqrt v, so it cannot go far below zero. If 2κθ≥ξ22\kappa\theta \ge \xi^2, known as the Feller condition after Feller’s study of this diffusion [2], it never reaches zero at all. This is the square-root process of the Cox–Ingersoll–Ross interest-rate model [3], borrowed for variance.

What each parameter does to the smile

Three-month options, spot 100, current and long-run volatility both 60%, κ=2\kappa = 2, ξ=1\xi = 1. Implied volatility by strike:

Strike708090100110120130
ρ=−0.7\rho = -0.766.363.260.457.755.353.051.1
ρ=0\rho = 060.559.558.958.858.959.359.7
ρ=+0.3\rho = +0.357.257.558.259.360.461.662.7
ρ=−0.7\rho = -0.7, ξ=0.3\xi = 0.362.261.260.359.658.858.257.6

Correlation ρ\rho tilts it. With ρ<0\rho < 0, falls in price come with rises in volatility, which fattens the left tail of the return distribution; low strikes become expensive and the curve slopes down. Flip the sign and the skew flips.

Vol of vol ξ\xi bends it. It sets how much the variance moves and so how fat both tails are. Cut it from 1 to 0.3 and the same correlation gives a skew less than half as steep.

The level sits below 60%. Even with ρ=0\rho = 0, at-the-money volatility is 58.8%, not 60%. Implied volatility near the money tracks the average of v\sqrt v, and by Jensen’s inequality that is below E[v]\sqrt{\mathbb{E}[v]}.

Try it: where the smile comes from
At-the-money vol–
Skew, 90 − 110–
Wings − ATM–
Feller 2κθ/ξ²–
Spot 100, three-month options, rate 0; v₀ = θ = 0.36 (60% volatility). Left: Heston implied volatility by strike, against the flat 60% of Black–Scholes. Right: one year of simulated price and volatility (full-truncation Euler, daily steps).

Pricing without a closed form

Heston’s contribution was computability. The characteristic function of ln⁡ST\ln S_T is known in closed form, and a call price is two one-dimensional Fourier integrals of it:

C=S P1−Ke−rTP2,Pj=12+1π∫0∞Re⁡ ⁣[e−iuln⁡K fj(u)iu]du.C = S\,P_1 - K e^{-rT} P_2, \qquad P_j = \frac12 + \frac1\pi \int_0^\infty \operatorname{Re}\!\left[\frac{e^{-iu\ln K}\, f_j(u)}{iu}\right] du .

The textbook form of fjf_j has a branch-cut problem: the complex logarithm inside it can jump, and the integral silently returns a wrong price for long maturities. The rearranged form used here and in the figure is proven stable by Albrecher, Mayer, Schoutens and Tistaert [4].

heston.py
import numpy as np
from scipy.integrate import quad
def char_fn(u, S, T, r, v0, kappa, theta, xi, rho):
"""Characteristic function of ln S_T, in the form that avoids the branch cut."""
b = kappa - rho * xi * 1j * u
d = np.sqrt(b**2 + xi**2 * (1j * u + u**2))
g = (b - d) / (b + d)
e = np.exp(-d * T)
C = r * 1j * u * T + kappa * theta / xi**2 * ((b - d) * T - 2 * np.log((1 - g * e) / (1 - g)))
D = (b - d) / xi**2 * (1 - e) / (1 - g * e)
return np.exp(C + D * v0 + 1j * u * np.log(S))
def heston_call(S, K, T, r, *params):
phi = lambda u: char_fn(u, S, T, r, *params)
forward = phi(-1j) # E[S_T] = S e^{rT}
p1 = quad(lambda u: (np.exp(-1j * u * np.log(K)) * phi(u - 1j) / (1j * u * forward)).real, 1e-8, 200, limit=400)[0]
p2 = quad(lambda u: (np.exp(-1j * u * np.log(K)) * phi(u) / (1j * u)).real, 1e-8, 200, limit=400)[0]
return S * (0.5 + p1 / np.pi) - K * np.exp(-r * T) * (0.5 + p2 / np.pi)
print(heston_call(100, 90, 0.25, 0.0, 0.36, 2.0, 0.36, 1.0, -0.7)) # about 17.08

What Heston gets right and wrong

It produces skew and smile from two interpretable parameters and prices vanillas fast enough to calibrate. It struggles with very short maturities, where real smiles are steeper than a diffusion for variance can make them; that gap is one reason for adding jumps and for newer models of the variance. As a first model of why the smile exists, it is the right one; Gatheral’s book [7] is a practitioner’s tour of it and its extensions.

References

  1. S. L. Heston, “A closed-form solution for options with stochastic volatility with applications to bond and currency options,” Review of Financial Studies 6(2), 327–343 (1993). doi:10.1093/rfs/6.2.327
  2. W. Feller, “Two singular diffusion problems,” Annals of Mathematics 54(1), 173–182 (1951). doi:10.2307/1969318
  3. J. C. Cox, J. E. Ingersoll and S. A. Ross, “A theory of the term structure of interest rates,” Econometrica 53(2), 385–407 (1985). doi:10.2307/1911242
  4. H. Albrecher, P. Mayer, W. Schoutens and J. Tistaert, “The little Heston trap,” Wilmott Magazine, January 2007, 83–92.
  5. R. Lord, R. Koekkoek and D. van Dijk, “A comparison of biased simulation schemes for stochastic volatility models,” Quantitative Finance 10(2), 177–194 (2010). doi:10.1080/14697680802392496
  6. L. Andersen, “Simple and efficient simulation of the Heston stochastic volatility model,” Journal of Computational Finance 11(3), 1–42 (2008). doi:10.21314/JCF.2008.189
  7. J. Gatheral, The Volatility Surface: A Practitioner’s Guide, Wiley (2006). ISBN 978-0-471-79251-2.