Conditional Forecasting for the Ornstein-Uhlenbeck Process

Forecasting from the Current State

An Ornstein-Uhlenbeck (OU) process is mean-reverting, but a forecast does not simply equal its long-run mean: it also reflects the current state, with that influence decaying over time. Earlier observations can improve a forecast when the current state is not known exactly. The calculations below separate these two situations: forecasting conditional on the current state, and forecasting conditional on exact or noisy observations from earlier times.

Work on a probability space $(\Omega,\mathcal{F},\mathbb{P})$. Let $(X_t)_{t\ge0}$ be a real-valued process, with state space $S=\mathbb{R}$ and continuous time $t\ge0$. Use the OU parameters $\theta\in\mathbb{R}$, $\alpha>0$, and $\sigma>0$. Its transition law over an interval of length $h>0$, conditional on the current state, is normal with mean $\theta+e^{-\alpha h}(x-\theta)$ and variance $\frac{\sigma^2}{2\alpha}(1-e^{-2\alpha h})$ when $X_s=x$. Write $\mathcal{F}_s$ for the process's information up to time $s$; the OU Markov property says the conditional law of its future given $\mathcal{F}_s$ depends on that past information through $X_s$ alone.

Forecast from a known state: For $T>s$, conditional on $X_s=x$, the forecast $X_T$ is normal with $$ \mathbb{E}[X_T\mid X_s=x]=\theta+e^{-\alpha(T-s)}(x-\theta),\qquad \operatorname{Var}(X_T\mid X_s=x)=\frac{\sigma^2}{2\alpha}\bigl(1-e^{-2\alpha(T-s)}\bigr). $$ The same conditional law holds given $\mathcal{F}_s$, with $x$ replaced by $X_s$.

These formulas are the transition law with $h=T-s$. The conditional mean moves from $x$ toward $\theta$ as the horizon grows; the conditional variance increases from zero and approaches $\sigma^2/(2\alpha)$. At $T=s$, the forecast is exactly $X_s$, with zero conditional variance. This current-state forecast is also the benchmark for understanding what earlier observations add when $X_s$ itself is unobserved.

Worked Example: A Two-Unit Forecast from a Known State

Take $\theta=2$, $\alpha=\tfrac12$, $\sigma^2=2$, current time $s=1$, forecast time $T=3$, and current value $X_1=5$. The horizon is $T-s=2$. The conditional mean is

2+e^{-(1/2)(2)}(5-2)=2+3e^{-1}.

The conditional variance is

\frac{2}{2(1/2)}\bigl(1-e^{-2(1/2)(2)}\bigr)=2(1-e^{-2}).

Thus $X_3\mid X_1=5$ is normal with mean $2+3e^{-1}$ and variance $2(1-e^{-2})$. Check: as the horizon tends to zero, the mean formula tends to $5$ and the variance to $0$; as it tends to infinity, the mean tends to $2$ and the variance to $2=\sigma^2/(2\alpha)$.

Conditioning a Forecast on Earlier Exact Observations

Suppose the process is stationary, with mean $\theta$ and covariance $\operatorname{Cov}(X_t,X_u)=v e^{-\alpha|t-u|}$, where $v=\sigma^2/(2\alpha)$. For observation times $t_1,\dots,t_m<T$, collect the earlier observations into the column vector $Y=(X_{t_1},\dots,X_{t_m})^\mathsf{T}$. Its mean vector is $\theta\mathbf{1}$, and its covariance matrix $K$ has entries $K_{ij}=v e^{-\alpha|t_i-t_j|}$. Let $c$ be the column vector with entries $c_i=v e^{-\alpha(T-t_i)}$, the covariance between $X_T$ and $X_{t_i}$.

Theorem (Gaussian forecast update): If $K$ is invertible, then, conditional on $Y=y$, the forecast $X_T$ is normal with $$ \mathbb{E}[X_T\mid Y=y]=\theta+c^\mathsf{T}K^{-1}(y-\theta\mathbf{1}),\qquad \operatorname{Var}(X_T\mid Y=y)=v-c^\mathsf{T}K^{-1}c. $$ For continuous observations, a particular value $y$ can have probability zero. The displayed expressions specify the Gaussian conditional distribution (a version of the regular conditional law) at that value.

Proof. The vector $(X_T,Y^\mathsf{T})^\mathsf{T}$ is jointly Gaussian because the process is Gaussian. Define the residual $R=X_T-\theta-c^\mathsf{T}K^{-1}(Y-\theta\mathbf{1})$. Its mean is zero. Using $\operatorname{Cov}(Y,X_T)=c$ and $\operatorname{Cov}(Y)=K$,

\operatorname{Cov}(R,Y) =c^\mathsf{T}-c^\mathsf{T}K^{-1}K=0.

The residual and $Y$ are jointly Gaussian, so zero cross-covariance makes them independent: the joint Gaussian characteristic function factors into the product of the two marginal characteristic functions when the cross-covariance is zero. Consequently, conditional on $Y=y$, $R$ retains its unconditional mean zero and variance. Expanding that variance gives $\operatorname{Var}(R)=v-c^\mathsf{T}K^{-1}c$. Replacing $R$ by its definition gives the stated conditional mean and variance. $\square$

When all observations are exact values of the process, the OU Markov property provides a useful simplification. If $t_m$ is the latest observation time, then the conditional distribution of $X_T$ given $X_{t_1},\dots,X_{t_m}$ is the same as its distribution given $X_{t_m}$ alone. This screening property is for forecasting strictly after the latest observation; conditioning also on a future endpoint is a different problem.

Proof. Since $T>t_m$, the Markov property at $t_m$ gives the conditional distribution of $X_T$ given all information through $t_m$ as a function of $X_{t_m}$ alone. The earlier observations $X_{t_1},\dots,X_{t_{m-1}}$ are among that past information. Conditioning this transition law on those earlier values therefore does not change it once $X_{t_m}$ is specified. $\square$

Worked Example: Exact Observations at Two Earlier Times

Let the stationary OU process have $\theta=0$, $v=1$, and $\alpha=\log 2$. Observe $X_0=1$ and $X_1=-1$, and forecast $X_2$. Since $e^{-\alpha}=1/2$, the covariance matrix of the two observations and their covariance vector with $X_2$ are

K= \begin{bmatrix} 1&1/2\\ 1/2&1 \end{bmatrix}, \qquad c= \begin{bmatrix} 1/4\\ 1/2 \end{bmatrix}.

The determinant of $K$ is $1-1/4=3/4$, so it is invertible, and

K^{-1}=\frac43 \begin{bmatrix} 1&-1/2\\ -1/2&1 \end{bmatrix}, \qquad K^{-1}c= \begin{bmatrix} 0\\ 1/2 \end{bmatrix}.

Thus the conditional mean is $0+0(1)+\tfrac12(-1)=-\tfrac12$, and the conditional variance is $1-c^\mathsf{T}K^{-1}c=1-(\tfrac14\cdot0+\tfrac12\cdot\tfrac12)=\tfrac34$. Hence $X_2\mid(X_0=1,X_1=-1)$ is normal with mean $-\tfrac12$ and variance $\tfrac34$. Check: conditioning only on the latest exact observation gives mean $e^{-\alpha}( -1)=-\tfrac12$ and variance $1-e^{-2\alpha}=1-\tfrac14=\tfrac34$, as the screening property predicts.

Noisy Observations

Now suppose the recorded values are $Y_i=X_{t_i}+\varepsilon_i$, where the errors $\varepsilon_i$ are independent of the process and of one another, Gaussian with mean zero, and have variances $\tau_i^2$. Then the observations and forecast are jointly Gaussian. The observation covariance is $K+\operatorname{diag}(\tau_1^2,\dots,\tau_m^2)$, while their covariance with $X_T$ remains $c$. The same Gaussian update applies with this adjusted observation covariance. Independent measurement errors generally remove the screening property: an earlier noisy measurement can still contain information about $X_T$ after a later noisy measurement is known.

Worked Example: Forecasting from Two Noisy Measurements

Use the stationary process from the previous example: $\theta=0$, $v=1$, $\alpha=\log 2$. At times $0$ and $1$, observe $Y_i=X_{t_i}+\varepsilon_i$, where the independent errors are Gaussian with mean zero and variance $1$, independent of the process. Suppose $Y_0=1$, $Y_1=-1$, and forecast $X_2$. The measurement covariance matrix is the exact-observation covariance plus $1$ on each diagonal entry:

K_Y= \begin{bmatrix} 2&1/2\\ 1/2&2 \end{bmatrix}, \qquad \operatorname{Cov}(Y,X_2)= \begin{bmatrix} 1/4\\ 1/2 \end{bmatrix}.

The determinant of $K_Y$ is $4-\tfrac14=\tfrac{15}{4}$, and its inverse is $\frac4{15}\begin{bmatrix}2&-1/2\\-1/2&2\end{bmatrix}$. Multiplying gives

K_Y^{-1}\operatorname{Cov}(Y,X_2) = \begin{bmatrix} 1/15\\ 7/30 \end{bmatrix}.

The conditional mean is $\tfrac1{15}(1)+\tfrac7{30}(-1)=-\tfrac16$. The conditional variance is $1-(\tfrac14\cdot\tfrac1{15}+\tfrac12\cdot\tfrac7{30}) =1-(\tfrac1{60}+\tfrac7{60})=\tfrac{13}{15}$. Therefore $X_2\mid(Y_0=1,Y_1=-1)$ is normal with mean $-\tfrac16$ and variance $\tfrac{13}{15}$. Check: the covariance reduction is $\tfrac{2}{15}$, so the remaining variance is $1-\tfrac{2}{15}=\tfrac{13}{15}$, which is nonnegative and less than the prior variance $1$.

Assumptions and a Common Pitfall

The Gaussian update for noisy data requires joint Gaussianity, not merely a mean and variance for each error. Independence of the errors from the process and from each other justifies adding their variances to the observation covariance diagonal and leaves the cross-covariance with the target unchanged. But those covariance facts alone do not make the observations jointly Gaussian. If the errors are not Gaussian, the conditional distribution may require the full error distributions and need not be normal.

Common mistake. It is incorrect to assume that independent, mean-zero errors with variance $1$ automatically justify a Gaussian conditional forecast. For example, let $X\sim N(0,1)$ and let an independent error $\varepsilon$ equal $1$ or $-1$, each with probability $1/2$. Then $Y=X+\varepsilon$ has mean zero and variance $2$, but $X\mid Y=y$ is supported on the two values $y-1$ and $y+1$: given either error value, $X=y-\varepsilon$. It is not a normal conditional distribution. The correction is to require Gaussian errors for the normal update, or else use a method that accounts for their actual distribution.

For practical forecasting, first identify what is known at the forecast origin. If the current state is known, use the OU transition law directly. If only earlier exact observations are available, use Gaussian conditioning or, when appropriate, the latest-observation screening property. If observations are noisy, include their error variances in the observation covariance and verify the distributional assumptions before claiming a normal conditional law.

Check Your Understanding

For each question, state which conditioning information is being used and check the variance calculation.

For an OU process with parameters $\theta,\alpha,\sigma$, write the conditional mean and variance of $X_{s+h}$ given $X_s=x$.
In the exact-observation example, why does knowing $X_0$ add no forecast information once $X_1$ is known?
For two independent Gaussian measurement errors with variances $\tau_1^2$ and $\tau_2^2$, which entries of the observation covariance matrix change?
Why does specifying only mean-zero measurement errors and their variances not establish a Gaussian conditional forecast?
What condition on the observation covariance matrix is used to apply the displayed inverse-matrix update directly?

Conditional Forecasting for the Ornstein-Uhlenbeck Process

What You'll Learn