I'm trying to figure out if stochastic calculus is the right approach for this problem... but I only vaguely understand it and I am trying to gauge if I need to spend the time learning measure theory etc etc for something that could be much more trivial.

I am going to make some simplifications here (e.g. normal instead of lognormal), so please bear with me. I'm just trying to get my bearings.

Let's say I have two stocks, X and Y.

Stock X is worth 100 and the price is normally distributed with $\sigma_X = 1$ per period.

Stock Y is worth 99 and the price is normally distributed with $\sigma_Y = 3.3333$ per period.

The two stocks are 90% correlated and the beta is $0.9 \cdot 3.333/1 = 3$

So what if I said to you, in one period, I'll give you stock X or stock Y, whichever has a greater value.

So the question is... what's the expected value of the max of the two stocks?

I can, of course, simulate this scenario, but is there a way to get the answer closed form?

Using Cholesky decomposition, I simulate two correlated random normals and give them the appropriate mean and standard deviation.

I approximate the value of $\max(X,Y) \approx 100.6$

  • $\begingroup$ Welcome to Quantitative Finance! Would you please add your research on the matter and what you think the answer might be? (Perhaps a tree with values of starting and ending states). $\endgroup$ – rajah9 Dec 23 '15 at 20:42
  • $\begingroup$ Is there a place I can upload my excel file? $\endgroup$ – TH4454 Dec 23 '15 at 20:45

Consider two jointly normal random variables $X_1 \sim N(u_1, \sigma_1^2)$ and $X_2 \sim N(u_2, \sigma_2^2)$. Note that, \begin{align*} \max(X_1, X_2) = X_2 + \max(X_1-X_2, \ 0). \end{align*} Moreover, $X_1-X_2$ is a normal random variable with mean $\mu=\mu_1-\mu_2$ and variance \begin{align*} \sigma^2 = \sigma_1^2+\sigma_2^2 - 2 \rho\sigma_1\sigma_2, \end{align*} where $\rho$ is the correlation. That is, \begin{align*} X_1-X_2 = \mu + \sigma \xi, \end{align*} where $\xi$ is a standard normal random variable. Therefore, \begin{align*} E\left(\max((X_1, X_2) \right) &= E(X_2) + E\left(\max(\mu + \sigma \xi, \ 0)\right)\\ &=\mu_2 + \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}\max(\mu + \sigma x, \ 0)\,e^{-\frac{x^2}{2}}\,dx\\ &=\mu_2 + \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}(\mu + \sigma x) \,\mathbb{I}_{\mu + \sigma x\geq 0}\,e^{-\frac{x^2}{2}}\, dx\\ &=\mu_2 + \frac{1}{\sqrt{2\pi}}\int_{-\frac{\mu}{\sigma}}^{\infty}(\mu + \sigma x)\,e^{-\frac{x^2}{2}} dx\\ &=\mu_2 + \mu \,N\Big(\frac{\mu}{\sigma}\Big) +\frac{\sigma}{\sqrt{2\pi}}\,e^{-\frac{\mu^2}{2\sigma^2}}, \end{align*} where $N$ is the cumulative distribution function of a standard normal random variable.

  • $\begingroup$ The end result gives me the right answer, but how did you jump from E(max(mu + sigma*epsilon,0) to the integral? $\endgroup$ – TH4454 Dec 28 '15 at 16:48
  • 1
    $\begingroup$ I added a few more steps. $\endgroup$ – Gordon Dec 28 '15 at 17:35

This took a 5 second Google search: https://stats.stackexchange.com/questions/139072/distribution-of-the-maximum-of-two-correlated-normal-variables

All you're really looking is simply to compute $E[\max\{X, Y\}]$, where $X, Y$ are two correlated normal random variables. The link there tells you the pdf of the random variable $W = \max\{X,Y\}$. From there, you can integrate and compute $E[W]$ --- although most likely you can only numerically integrate in this case.

Remark: Furthermore, if I may add, there's nothing "stochastic calculus" in your problem (at least the way that you phrase it). And also, measure theory is also unnecessary for this problem.

  • 1
    $\begingroup$ That gives the PDF, but not the expected value. Please try not to be rude, I've given considerable thought to this question. $\endgroup$ – TH4454 Dec 23 '15 at 22:11
  • $\begingroup$ Actually, I stand corrected. Looking at the reference gwern.net/docs/conscientiousness/2008-nadarajah.pdf right in the link above, there is a closed form expectation --- see equation (9) $\endgroup$ – user32416 Dec 23 '15 at 22:26
  • $\begingroup$ The paper has equations (9) and (10) for the first two moments. I can just use those for expected value and variance of the max(X,Y), right? $\endgroup$ – TH4454 Dec 23 '15 at 22:27
  • $\begingroup$ Yes. And actually with the mgf, you can get any moments. Please accept my answer if you think this solves your problem. $\endgroup$ – user32416 Dec 23 '15 at 22:28

I'd recommend following Margrabe's formula, which is a closed form Black-Scholes type formula for valuing european call option whose payoff at time T = max[0, S2(T) - S1(T)]. If you create a portfolio of S1 and such a call option, your payoff will be higher of the two stock prices - which is what you want. Plus this has a more general assumption of prices being log-normally distributed.

Margrabe's call option formula: C = S2.N(d1) - S1.k.N(d2)

For d1, d2 and other details please refer to the paper: http://www.stat.nus.edu.sg/~stalimtw/MFE5010/PDF/margrabe1978.pdf


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.