3
$\begingroup$

Let say, I have 2 stochastic processes: $$\begin{align} dS_1 &= \left( r - q_1 \right)S_1 dt + \sigma_1 S_1 dW_1 \\ dS_2 &= \left( r - q_2 \right)S_2 dt + \sigma_2 S_2 dW_2 \end{align}$$ The correlation between these 2 processes is $\rho$. Now I define 2 new processes as: $$\begin{align} x_1 = \sigma_1 \log S_2 + \sigma_2 \log S_1 \\ x_2 = \sigma_1 \log S_2 - \sigma_2 \log S_1 \end{align}$$

As per Hull's book, these processes $x_1, x_2$ are uncorrelated with standard deviation $\sigma_1 \sigma_2 \sqrt{2 \left( 1+\rho \right)}$ and $\sigma_1 \sigma_2 \sqrt{2 \left( 1-\rho \right)}$ respectively.

How can I show this result?

$\endgroup$
3
  • 1
    $\begingroup$ What have you tried? Also, the correlation $\rho$ is between the Brownian Motions $W_1$ and $W_2$ correct? $\endgroup$ – Daneel Olivaw Nov 3 '20 at 9:52
  • $\begingroup$ I am able to derive the $dt$ term, but not the volatility part. Apologies if this is very trivial question. Any pointer is appreciated. $\endgroup$ – Brian Smith Nov 3 '20 at 9:54
  • $\begingroup$ If $S_i$ is log-normal, then $\log S_i$ is normal. Have you tried computing the covariance between $x_1$ and $x_2$? $\endgroup$ – Daneel Olivaw Nov 3 '20 at 9:56
3
$\begingroup$

I have the impression the expressions in your question miss the time term $t$, though this does not change much. Define for $i,j\in\{1,2\}$: $$\begin{align} s_i(t)&=\log S_i(t) \\ y_{i,j}(t)&=\sigma_is_j(t) \end{align}$$ Then: $$\begin{align} V(y_{i,j}(t))&=\sigma_i^2V(s_j(t)) \\ &=\sigma_i^2V(\sigma_jW_j(t)) \\ &=\sigma_i^2\sigma_j^2t \\ &=V(y_{j,i}(t)) \end{align}$$ Moreover: $$\begin{align} C(y_{1,2}(t),y_{2,1}(t))&=\sigma_1\sigma_2C(s_2(t),s_1(t)) \\ &=\sigma_1^2\sigma_2^2C(W_2(t),W_1(t)) \\ &=\sigma_1^2\sigma_2^2\rho t \end{align}$$ Hence: $$\begin{align} &V(x_1(t))=\sigma_1^2\sigma^2_2t+\sigma_2^2\sigma_1^2t+2\sigma^2_1\sigma^2_2\rho t=\sigma_1^2\sigma_2^22(1+\rho)t \\ &V(x_2(t))=\sigma_2^2\sigma^2_1t+\sigma_1^2\sigma_2^2t-2\sigma^2_1\sigma^2_2\rho t=\sigma_1^2\sigma_2^22(1-\rho)t \end{align}$$ Finally, by bi-linearity and symmetry of covariance: $$\begin{align} C(x_1(t),x_2(t))&=C(y_{1,2}(t)+y_{2,1}(t),y_{1,2}(t)-y_{2,1}(t)) \\ &=V(y_{1,2}(t))-C(y_{1,2}(t),y_{2,1}(t))+C(y_{2,1}(t),y_{1,2}(t))-V(y_{2,1}(t)) \\ &=V(y_{1,2}(t))-V(y_{2,1}(t)) \\ &=0 \end{align}$$ $x_1$ and $x_2$ are uncorrelated $-$ note that for normally-distributed random variables, null correlation also implies independence.

$\endgroup$
2
  • $\begingroup$ Many thanks. Is there any way on how to generalise this for n initial correlated stochastic processed for $S_t$? $\endgroup$ – Brian Smith Nov 3 '20 at 21:22
  • $\begingroup$ Generally, you can express a specific covariance structure in your asset process as $CdW$ with $C$ the lower cholesky decomposition of the covariant empire matrix. I think that you may be able to use its inverse $C^{-1}$ to weight the assets in order to de-correlate the assets. $\endgroup$ – Kermittfrog Nov 3 '20 at 22:04

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.