9.5 Ornstein-Uhlenbeck process

Definition of Ornstein-Uhlenbeck process

The Ornstein-Uhlenbeck (OU) process is a stochastic process that describes the velocity of a massive Brownian particle under the influence of friction. Leonard Ornstein and George Eugene Uhlenbeck introduced it in 1930 as a physically motivated refinement of pure Brownian motion.

The defining feature of the OU process is mean reversion: the process tends to drift back toward a long-term average value over time. This makes it fundamentally different from a standard Wiener process, which has no such tendency. The OU process appears throughout finance (interest rate modeling), physics (particle dynamics in viscous media), and biology (neural firing models).

Mean-reverting property

Mean reversion is the tendency of a stochastic process to return over time to a long-term average value. For the OU process, random noise continually perturbs the state, but a friction-like drift term pulls it back toward the long-term mean $\mu$.

• The rate of this pull-back is governed by the parameter $\theta$. Larger $\theta$ means faster reversion.
• The restoring force is proportional to the displacement $(\mu - X_t)$, so the further the process wanders from $\mu$, the stronger the pull back.

Stationary Gaussian process

The OU process is a stationary Gaussian process, meaning its finite-dimensional distributions are multivariate normal and, once the process has reached equilibrium, these distributions are invariant under time shifts.

• At any fixed time $t$, the OU process follows a normal distribution with mean $X_0 e^{-\theta t} + \mu(1 - e^{-\theta t})$ and variance $\frac{\sigma^2}{2\theta}(1 - e^{-2\theta t})$.
• The joint distribution at any finite collection of times is multivariate normal.
• Once the transient from the initial condition has died out, the statistical properties (mean, variance, covariance structure) no longer depend on absolute time.

Continuous-time Markov process

The OU process is a continuous-time Markov process: its future evolution depends only on its current state, not on its past history.

• The transition density from state $x$ at time $t$ to state $y$ at time $t+s$ is Gaussian and can be written in closed form from the explicit solution.
• The Markov property lets you apply the full toolkit of stochastic calculus: Kolmogorov forward/backward equations, Itô's lemma, and Feynman-Kac representations all apply directly.

Mathematical formulation

The OU process is defined by the stochastic differential equation (SDE):

$dX_t = \theta(\mu - X_t)\,dt + \sigma\,dW_t$

where:

• $X_t$ is the state at time $t$
• $\theta > 0$ is the mean reversion rate
• $\mu$ is the long-term mean
• $\sigma > 0$ is the volatility parameter
• $W_t$ is a standard Brownian motion (Wiener process)

Stochastic differential equation

The SDE has two terms:

• Drift term $\theta(\mu - X_t)\,dt$: the deterministic component that pulls $X_t$ toward $\mu$.
• Diffusion term $\sigma\,dW_t$: the stochastic component that introduces Gaussian random fluctuations.

The linearity of the drift in $X_t$ is what makes the OU process analytically tractable.

Drift term

The drift $\theta(\mu - X_t)$ acts as a restoring force, analogous to a spring:

• When $X_t < \mu$, the drift is positive, pushing the process upward.
• When $X_t > \mu$, the drift is negative, pushing the process downward.
• The magnitude of the force is proportional to the displacement $|\mu - X_t|$ and to the rate constant $\theta$.

Diffusion term

The diffusion term $\sigma\,dW_t$ injects randomness into the process. The coefficient $\sigma$ controls the amplitude of these fluctuations. Since $dW_t$ has independent, normally distributed increments with variance $dt$, the instantaneous noise is Gaussian with variance $\sigma^2\,dt$.

Mean reversion rate

The parameter $\theta$ controls how quickly the process reverts to $\mu$.

• Larger $\theta$ means tighter clustering around $\mu$ and faster decay of deviations.
• The quantity $1/\theta$ is the characteristic relaxation time: roughly the time scale over which a deviation from the mean decays by a factor of $e$.

Long-term mean

The parameter $\mu$ is the equilibrium level around which the process fluctuates. If you turned off the noise ($\sigma = 0$), the process would converge deterministically to $\mu$ via exponential decay. In applications, $\mu$ is typically estimated from data.

Volatility parameter

The parameter $\sigma$ sets the scale of random fluctuations. Higher $\sigma$ produces noisier sample paths and a wider stationary distribution. In finance, $\sigma$ is often calibrated from historical data or implied from derivative prices.

Solution of Ornstein-Uhlenbeck SDE

The OU SDE is one of the few SDEs with a closed-form solution. Solving it gives you an explicit expression for $X_t$ in terms of the initial condition, the parameters, and a stochastic integral.

Itô's lemma

The standard approach to solving the OU SDE uses Itô's lemma. Here are the key steps:

1. Define the auxiliary process $Y_t = X_t\,e^{\theta t}$.

2. Apply Itô's lemma to $Y_t$. Because $f(t, x) = x\,e^{\theta t}$, you get: $dY_t = e^{\theta t}\,dX_t + \theta\,e^{\theta t}\,X_t\,dt$

3. Substitute the OU SDE for $dX_t$: $dY_t = e^{\theta t}\bigl[\theta(\mu - X_t)\,dt + \sigma\,dW_t\bigr] + \theta\,e^{\theta t}\,X_t\,dt = \theta\mu\,e^{\theta t}\,dt + \sigma\,e^{\theta t}\,dW_t$

4. Integrate both sides from $0$ to $t$ and solve for $X_t$.

Explicit solution

The result of the derivation above is:

$X_t = X_0\,e^{-\theta t} + \mu(1 - e^{-\theta t}) + \sigma \int_0^t e^{-\theta(t-s)}\,dW_s$

The three terms have clear interpretations:

• $X_0\,e^{-\theta t}$: the decaying influence of the initial condition.
• $\mu(1 - e^{-\theta t})$: the growing pull toward the long-term mean.
• $\sigma \int_0^t e^{-\theta(t-s)}\,dW_s$: the accumulated effect of random shocks, with recent shocks weighted more heavily than distant ones.

Initial condition

The initial value $X_0$ determines the starting point. Its influence decays exponentially: after a time of order $1/\theta$, the process has largely "forgotten" where it started.

Time-dependent mean and variance

From the explicit solution, you can read off the moments directly:

• Mean: $\mathbb{E}[X_t] = X_0\,e^{-\theta t} + \mu(1 - e^{-\theta t})$
• Variance: $\text{Var}[X_t] = \frac{\sigma^2}{2\theta}(1 - e^{-2\theta t})$

As $t \to \infty$:

• The mean converges to $\mu$.
• The variance converges to the stationary variance $\frac{\sigma^2}{2\theta}$.

Notice the variance grows from zero (if $X_0$ is deterministic) and saturates. The factor of $2\theta$ in the denominator shows that stronger mean reversion reduces the equilibrium spread.

Stationary distribution

As $t \to \infty$, the OU process converges in distribution to a unique stationary (equilibrium) distribution, regardless of the initial condition.

Gaussian distribution

The stationary distribution is Gaussian with:

• Mean: $\mu$
• Variance: $\frac{\sigma^2}{2\theta}$

The probability density function is:

$f(x) = \sqrt{\frac{\theta}{\pi\sigma^2}} \exp\left(-\frac{\theta}{\sigma^2}(x - \mu)^2\right)$

The Gaussian form follows from the linearity of the SDE and the Gaussian nature of the driving Brownian motion.

Long-term mean and variance

The two quantities $\mu$ and $\frac{\sigma^2}{2\theta}$ fully characterize the equilibrium behavior. The stationary variance reflects the balance between the noise intensity $\sigma^2$ (which spreads the distribution) and the mean reversion strength $\theta$ (which concentrates it).

Equilibrium distribution

The equilibrium distribution describes the long-run behavior of the process, independent of where it started. From any initial condition, the transient effects decay exponentially, and the distribution of $X_t$ approaches the stationary Gaussian. This convergence is exponentially fast, with rate governed by $\theta$.

Invariant measure

The stationary distribution is an invariant measure: if $X_0$ is drawn from the stationary distribution, then $X_t$ has the same distribution for all $t > 0$. This is a direct consequence of the stationarity of the OU process. The invariant measure is unique (for $\theta > 0$), which guarantees ergodicity: time averages along a single sample path converge to ensemble averages under the stationary distribution.

Properties of Ornstein-Uhlenbeck process

Beyond its basic mean-reverting Gaussian structure, the OU process has several analytically tractable properties that make it a workhorse model.

Autocorrelation function

The stationary autocorrelation function measures how correlated the process is with itself at a time lag $|t - s|$:

$R(t, s) = \frac{\sigma^2}{2\theta}\,e^{-\theta|t-s|}$

This exponential decay means the process has a "memory" that fades on a time scale of $1/\theta$. Values separated by much more than $1/\theta$ are nearly uncorrelated.

Power spectral density

Taking the Fourier transform of the autocorrelation function gives the power spectral density:

$S(\omega) = \frac{\sigma^2}{\theta^2 + \omega^2}$

This is a Lorentzian (also called a Cauchy distribution in frequency space). It peaks at $\omega = 0$ and has half-width $\theta$. Most of the process's "power" is concentrated at low frequencies, consistent with the smooth, mean-reverting behavior of sample paths.

First passage time distribution

The first passage time is the first time the process hits a specified threshold, starting from a given initial value. Unlike for a standard Brownian motion, there is no simple closed-form expression for the OU first passage time distribution.

• It can be computed numerically via the Fokker-Planck equation or by solving a boundary value problem for the backward Kolmogorov equation.
• The mean first passage time and its variance depend on $\theta$, $\sigma$, the initial condition, and the threshold level.
• For thresholds far from $\mu$, the mean reversion makes passage times much longer than they would be for a driftless process.

Ornstein-Uhlenbeck bridge

The OU bridge is the OU process conditioned on having specified values at both endpoints: $X_0 = a$ and $X_T = b$.

• It remains a Gaussian process, with time-dependent mean and variance that can be computed in closed form from the conditional distribution formulas for multivariate normals.
• OU bridges are useful for simulation and inference problems where both the start and end states are observed, such as interpolating between discrete observations of a mean-reverting time series.

Applications in finance

The OU process is one of the most widely used stochastic models in quantitative finance, primarily because mean reversion is a natural feature of interest rates, credit spreads, and commodity prices.

Vasicek model for interest rates

The Vasicek model (1977) assumes the short rate $r_t$ follows an OU process:

$dr_t = \theta(\mu - r_t)\,dt + \sigma\,dW_t$

This captures the empirical observation that interest rates tend to revert toward a long-run level rather than wandering without bound.

Calibration of parameters

To use the Vasicek model in practice, you need to estimate $\theta$, $\mu$, and $\sigma$ from data. Common approaches:

1. Maximum likelihood estimation (MLE): Fit the Gaussian transition density to observed rate changes.
2. Generalized method of moments (GMM): Match sample moments (mean, variance, autocorrelation) to their theoretical counterparts.
3. Cross-sectional calibration: Fit the model's bond pricing formula to observed yield curves.

The choice of method depends on whether you're calibrating to historical time series or to current market prices.

Bond pricing formula

A major advantage of the Vasicek model is its closed-form bond pricing formula. The price at time $t$ of a zero-coupon bond maturing at $T$ is:

$P(t, T) = A(t, T)\,e^{-B(t, T)\,r_t}$

where $A(t, T)$ and $B(t, T)$ are deterministic functions of $\theta$, $\mu$, $\sigma$, and the time to maturity $T - t$. This tractability extends to pricing interest rate derivatives and computing yield curves analytically.

Limitations of Vasicek model

The Vasicek model's simplicity comes with trade-offs:

• Negative rates: The Gaussian distribution of $r_t$ assigns positive probability to negative interest rates. (This was historically seen as a flaw, though negative rates have occurred in practice.)
• Constant volatility: The model assumes $\sigma$ is fixed, which doesn't capture the time-varying volatility observed in real rate markets.
• Single factor: Only one source of randomness drives the entire yield curve, limiting the model's ability to fit complex term structures.

Extensions like the Cox-Ingersoll-Ross (CIR) model (which ensures non-negative rates by using a $\sqrt{r_t}$ diffusion coefficient) and the Hull-White model (which allows time-dependent parameters) address some of these issues.

Applications in physics

In physics, the OU process originated as a model for the velocity of a Brownian particle experiencing friction. It remains central to statistical mechanics and stochastic thermodynamics.

Brownian motion with friction

For a particle in a viscous medium, the velocity $v_t$ satisfies:

$dv_t = -\gamma\,v_t\,dt + \sigma\,dW_t$

Here $\gamma > 0$ is the friction coefficient (playing the role of $\theta$), and the long-term mean velocity is zero ($\mu = 0$). The friction term $-\gamma\,v_t$ decelerates the particle, while the noise term $\sigma\,dW_t$ represents random thermal kicks from surrounding molecules.

The stationary velocity distribution is Gaussian with mean zero and variance $\frac{\sigma^2}{2\gamma}$. Through the fluctuation-dissipation theorem, this variance is related to the temperature of the medium: $\frac{\sigma^2}{2\gamma} = \frac{k_B T}{m}$, where $k_B$ is Boltzmann's constant, $T$ is temperature, and $m$ is the particle mass.

Langevin equation

The OU SDE is a special case of the Langevin equation, a foundational equation in statistical physics that describes a particle subject to both deterministic and random forces. In its general form, the Langevin equation can include nonlinear restoring forces and state-dependent noise, but the linear (OU) case is the one that admits full analytical treatment. The connection between the Langevin equation and the Fokker-Planck equation provides a bridge between the trajectory-level (SDE) and distribution-level (PDE) descriptions of stochastic dynamics.