2
$\begingroup$

I know how to find the mean of an SDE: write it on integral form, take derivative, solve a simple ODE.

But what to do when we want a variance?

In my case, $$X_{T + \delta t} = X_T + \int_T^{T + \delta t} \lambda ( 1 -X)dt + \int_T^{T + \delta t} \sigma \sqrt{X} dW$$ and I want the variance, conditional on $X_T$.

Is there a simple approach?

$\endgroup$
6
$\begingroup$

Here are two approaches that you could take to compute the variance of $X_t$. I am not making the conditioning explicit as it just complicates the notation but doesn't really add any additional insights.

  1. Compute $\mathbb{E} \left[ X_t \right]$ and $\mathbb{E} \left[ X_t^2 \right]$. You can then you use that

    \begin{equation} \text{Var} \left( X_t \right) = \mathbb{E} \left[ X_t^2 \right] - \mathbb{E} \left[ X_t \right]^2. \end{equation}

  2. Compute the cumulant generating function

    \begin{equation} \psi_{X_t}(\omega) = \ln \left( \mathbb{E} \left[ e^{\mathrm{i} \omega X_t} \right] \right). \end{equation}

    Then use that the $n$-th cumulant is linked to the $n$-th derivative of the cumulant generating function at $\omega = 0$

    \begin{equation} c_n \left( X_t \right) = \frac{1}{\mathrm{i}^n} \frac{\partial^n}{\partial \omega^n} \psi_{X_t}(0). \end{equation}

    The first two cumulants are equal to the mean and variance, respectively.


In what follows, I am working out the first approach. In your case, the dynamics of $X$ are given by

\begin{equation} \mathrm{d}X_t = \lambda \left( 1 - X_t \right) \mathrm{d}t + \sigma \sqrt{X}_t \mathrm{d}W_t. \end{equation}

Note that this is a special case of a square-root process with mean reversion speed $\lambda$ and a long-term mean of one. The solution here is thus a special case of the more general solution given e.g. in the context of the Cox-Ingersoll-Ross interest rate model or the Heston stochastic volatility model.

First, we remove the geometric drift term $-\lambda X_t \mathrm{d}t$ by setting $Y_t = e^{\lambda t} X_t$. Its differential is given by

\begin{eqnarray} \mathrm{d}Y_t & = & \lambda e^{\lambda t} X_t \mathrm{d}t + e^{\lambda t} \mathrm{d}X_t\\ & = & \lambda e^{\lambda t} \mathrm{d}t + \sigma e^{\lambda t} \sqrt{X_t} \mathrm{d}W_t\\ & = & \lambda e^{\lambda t} \mathrm{d}t + \sigma e^{\lambda t / 2} \sqrt{Y_t} \mathrm{d}W_t. \end{eqnarray}

Integrating yields

\begin{eqnarray} Y_t & = & Y_0 + \lambda \int_0^t e^{\lambda u} \mathrm{d}u + \sigma \int_0^t e^{\lambda u / 2} \sqrt{Y_u} \mathrm{d}W_u\\ & = & Y_0 + e^{\lambda t} - 1 + \sigma \int_0^t e^{\lambda u / 2} \sqrt{Y_u} \mathrm{d}W_u. \end{eqnarray}

Since $\mathbb{E} \left[ X_t \right] = e^{-\lambda t} \mathbb{E} \left[ Y_t \right]$, we conclude that

\begin{equation} \mathbb{E} \left[ X_t \right] = e^{-\lambda t} X_0 + \left( 1 - e^{-\lambda t} \right). \end{equation}

Next, we compute the differential of $Y_t^2$ and get

\begin{eqnarray} \mathrm{d} Y_t^2 & = & 2 Y_t \mathrm{d}Y_t + \mathrm{d} \langle Y \rangle_t\\ & = & 2 \lambda e^{\lambda t} Y_t \mathrm{d}t + 2 \sigma e^{\lambda t / 2} Y_t \sqrt{Y_t} \mathrm{d}W_t + \sigma^2 e^{\lambda t} Y_t \mathrm{d}t. \end{eqnarray}

Integrating again gives

\begin{equation} Y_t^2 = Y_0^2 + \left( 2 \lambda + \sigma^2 \right) \int_0^t e^{\lambda u} Y_u \mathrm{d}u + 2 \sigma \int_0^t e^{\lambda u / 2} Y_u \sqrt{Y_u} \mathrm{d}W_u. \end{equation}

The expectated value is

\begin{eqnarray} \mathbb{E} \left[ Y_t^2 \right] & = & Y_0^2 + \left( 2 \lambda + \sigma^2 \right) \int_0^t e^{\lambda u} \mathbb{E} \left[ Y_u \right] \mathrm{d}u\\ & = & Y_0^2 + \left( 2 \lambda + \sigma^2 \right) \int_0^t e^{\lambda u} \left( Y_0 + e^{\lambda u} - 1 \right) \mathrm{d}u\\ & = & Y_0^2 + \left( 2 \lambda + \sigma^2 \right) \left( \frac{1}{\lambda} \left( Y_0 - 1 \right) \left( e^{\lambda t} - 1 \right) + \frac{1}{2 \lambda} \left( e^{2 \lambda t} - 1 \right) \right) \end{eqnarray}

Since $\mathbb{E} \left[ X_t^2 \right] = e^{-2 \lambda t} \mathbb{E} \left[ Y_t^2 \right]$, we get

\begin{eqnarray} \mathbb{E} \left[ X_t^2 \right] & = & e^{-2 \lambda t} X_0^2 + \left( 2 \lambda + \sigma^2 \right) \left( \frac{1}{\lambda} \left( X_0 - 1 \right) \left( e^{-\lambda t} - e^{-2 \lambda t} \right) + \frac{1}{2 \lambda} \left( 1 - e^{-2 \lambda t} \right) \right)\\ & = & e^{-2 \lambda t} X_0^2 + 2 X_0 \left( e^{-\lambda t} - e^{-2 \lambda t} \right) + \frac{\sigma^2}{\lambda} X_0 \left( e^{-\lambda t} - e^{-2 \lambda t} \right)\\ & & + \left( 1 - e^{-\lambda t} \right)^2 + \frac{\sigma^2}{2 \lambda} \left( 1 - e^{-\lambda t} \right)^2. \end{eqnarray}

Since

\begin{eqnarray} \left( \mathbb{E} \left[ X_t \right] \right)^2 = e^{-2 \lambda t} X_0^2 + 2 X_0 \left( e^{-\lambda t} - e^{-2 \lambda t} \right) + \left( 1 - e^{-\lambda t} \right)^2, \end{eqnarray}

it follows that the variance of $X_t$ is given by

\begin{equation} \text{Var} \left( X_t \right) = \frac{\sigma^2}{\lambda} X_0 \left( e^{-\lambda t} - e^{-2 \lambda t} \right) + \frac{\sigma^2}{2 \lambda} \left( 1 - e^{-\lambda t} \right)^2. \end{equation}

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.