Stein-Stein (1991) Stock Price Distributions with Stochastic Volatility: An Analytic Approach

Summary

Stein and Stein derive an exact closed-form distribution for stock prices when volatility follows an arithmetic Ornstein-Uhlenbeck (OU) process, using analytic techniques related to the heat equation on the Heisenberg group. The distribution is a mixture of lognormals averaged over an integrated-variance mixing distribution. The model is applied to options pricing — yielding a volatility smile — and to characterizing fat tails in stock price distributions as power-law behavior rather than lognormal exponential decay.

Key Claims

Model: $dP = \mu P\,dt + \sigma P\,dz_1$ ; $d\sigma = -\delta(\sigma - \bar\sigma)\,dt + k\,dz_2$ , with $dz_1 \perp dz_2$ . Parameters: $\delta$ (mean-reversion speed), $\bar\sigma$ (long-run mean), $k$ (vol-of-vol).
Closed-form distribution: $S(P,t)$ is given by a single-integral formula (eq. 9–10) composed entirely of elementary functions (sinh/cosh). First closed-form SV distribution valid for any $k$ and any $\delta \neq 0$ , unlike Hull-White (1987) which required $k \approx 0$ or numerical methods.
Mixture of lognormals: $S(P,t) = \int L(\sigma)\,m_t(\sigma)\,d\sigma$ , where $L(\sigma)$ is lognormal with Root Mean Square (RMS) volatility $\sigma = (t^{-1}\int\sigma^2\,ds)^{1/2}$ and $m_t(\sigma)$ is the mixing distribution over realized RMS volatility. This representation holds for any diffusion-type Stochastic Volatility (SV) model.
Approximate distribution: $m_t(\sigma) \approx p\cdot\sigma^a\cdot\exp(-b/\sigma^2)$ matches tail asymptotics at both 0 and $\infty$ . Approximate option prices differ from exact by ~1/10 to 1/3 of Black-Scholes error; computationally 10–100 $\times$ faster than exact.
Options pricing: Stochastic volatility raises all option prices above Black-Scholes (when $\sigma_0 = \bar\sigma$ ). Implied volatility is U-shaped in strike (volatility smile): lowest at-the-money (near-linear region), higher away-from-the-money (convex region). With nonzero volatility risk premium $\phi$ , replace $\bar\sigma$ with $\tilde\sigma = \bar\sigma - \phi k/\delta$ .
Fat tails: $S(P,t) \sim P^{-\gamma}$ as $P\to\infty$ (power law). Exponent $\gamma = 3/2 + \delta/k^2 + \cdots$ depends on $\delta$ , $k$ , $t$ . Key results: (i) $\delta = 0$ : $\gamma\to 2$ as $t$ or $k\to\infty$ (variance can disappear at long horizons); (ii) $\delta > 0$ : $\lim\gamma > 2$ always (variance exists at all horizons); (iii) $t$ or $k\to 0$ : $\gamma\to\infty$ (approaches lognormal). Matches Bookstaber-McDonald (1987): fat tails at 1–5 days, near-lognormal at 250 days — indirect support for stationary (mean-reverting) volatility.
Arithmetic vs. geometric OU: Arithmetic OU chosen for tractability. Geometric OU (log-normal volatility with mean reversion) is analytically intractable. For empirically calibrated parameters ( $\bar\sigma = 0.30$ , $\delta = 16$ , $k = 0.4$ ), $P(\sigma < 0) < 2\%$ .
No leverage effect: $dz_1$ and $dz_2$ are independent; capturing leverage requires a Constant Elasticity of Variance (CEV) generalization of eq. (1) or correlated Wiener processes, but the latter was analytically intractable.
Calibration (from Stein 1989, Merville-Pieptea 1989): index $\bar\sigma \approx 0.15$ – $0.20$ , $\delta \approx 4$ – $16$ (half-life ~2 weeks to 1 month), $k \approx 0.10$ – $0.30$ ; individual stocks $\bar\sigma \approx 0.25$ – $0.35$ , $\delta \approx 14$ – $20+$ , $k \approx 0.30$ – $0.60$ .

Concepts Introduced or Extended

Stochastic Volatility

Entities Mentioned

Quotes

"We use analytic techniques (related to the heat equation for the Heisenberg group) to derive a closed-form solution for the distribution of stock prices in this case."

"The concept of the mixing distribution for $\sigma$ provides a useful heuristic device for understanding these effects."

"This observation, taken together with our analytical results, would seem to provide indirect support for the hypothesis that volatility follows a stationary process."

My Take

The mixture-of-lognormals representation (eq. 11) is the paper's deepest insight and holds well beyond the specific OU model — it identifies that the option-pricing problem reduces to characterizing the distribution of integrated realized variance, a perspective that became central to the later variance-swap and realized-variance literature. The key limitation is the zero-correlation assumption ( $dz_1 \perp dz_2$ ): the leverage effect (negative vol-return correlation) is well-documented empirically and drives the observed put skew, which this model cannot produce. Heston (1993) solved the correlated case in continuous time via a different approach (characteristic function / Fourier inversion), which superseded Stein-Stein for options pricing in practice.