Let's say I have a normally distributed random variable $X_1$ with known standard deviation $\sigma_1$ and $E[X_1]$ is $0$. Let's say I have another variable with known standard deviation $\sigma_2$ and $E[X_2]$ is $-V$. Let's say the are correlated with $\rho = .9$.

This implies the $\beta$ (defined by $X_2 = \alpha + \beta X_1$) between $X_1$ and $X_2$ is $\beta =\rho \Sigma_2/\Sigma_1$.

This gives a graph that looks like the following.... graph

I want to know the probability that the value $[X_1]$ is less than or equal to $[X_2]$, i.e., $P(X_2 \ge X_1)$.

I tried to conceptualize this as concentric circles on an $X_1-X_2$ plane... but I'm not sure how to combine these variables with a known correlation.

  • 1
    $\begingroup$ How are $X_1$ and $X_2$ jointly distributed? If normally, then $Z=X_2-X_1$ is another normal with computable mean and variance, so you know $P(Z\ge 0)$. $\endgroup$ – A.S. Dec 23 '15 at 2:17

I think I have come up with an answer, and I'd like some feedback.

I used a cholesky decomposition to simulate the correlated random numbers, and my theoretical result appears to match my simulation.

The question $P(X_2 > X_1)$ can be turned into $P(X_2-X_1 > 0)$.

The joint distribution of $X_2-X_1$ then can be summarized as:

$$E[X_1-X_2] = w_1 E(X_1) + w_2 E(X_2)$$

$$\operatorname{Var}[X_2-X_1] = w_1^2 \sigma_1^2 + w_2^2 \sigma_2^2 + 2 w_1 w_2 \sigma_1 \sigma_2 \rho$$

where $\rho$ is the negative of the correlation of the two random variables, because if you think of a portfolio of two assets... we have $X_2 - X_1$, which is going to mean movements in $X_1$ will be the negative of the correlated movements in $X_2$.

For both $w_1$ and $w_2$, I used 1, because each of $X_1$ and $X_2$ have a coefficient of 1 in $F(X_1,X_2) = X_2-X_1$.

Then $P(X_2 - X_1 > 0) = 1-P(X_2 - X_1 < 0)$. And I can use the normal CDF with the given $E[X_2-X_1]$ and $\operatorname{Var}(X_2-X_1)$ to find that value.

Does this sound right? How about the argument for the negative for $\rho$?

  • 1
    $\begingroup$ The idea is good but it implicitly assumes $(X_1,X_2)$ have a binormal distribution, whereas your question only states that $X_2$ is "another variable." $\endgroup$ – whuber Dec 12 '17 at 15:13
  • $\begingroup$ You take $\rho$ negative because $Cor(X1,-X2) = - Cor(X1,X2)$. I think that your expressions would be more easy to follow if you either do not use the terms $w_1$ and $w_2$, or, if you do place them in your equation, then also use 'really' use them by expressing your $X_2 - X_1$ as $w_1 X_1+w_2 X_2$ with $w_2=1$ and $w_1 =-1$ $\endgroup$ – Sextus Empiricus Dec 12 '17 at 17:06

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.