1
$\begingroup$

I have calculated the daily returns of a certain stock. They can be analytically described by a 3 parameter Student's $t$ distribution whose density function is

$$f(x)=\frac{\left(\frac{\nu}{\nu+\left(\frac{x-\mu}{\sigma}\right)^2}\right)^{\frac{\nu+1}{2}}}{\sqrt{\nu}\sigma B\left(\frac\nu2,\frac12\right)},$$ where $\mu=0.11$, $\sigma=0.54$ and $\nu=1.52$.

With this starting point: how can I calculate the distribution function of (compounded) returns over $n$ days?

$\endgroup$
4
  • $\begingroup$ As far as I remember, there exists no 'nice' closed form density or even characteristic function for the student distribution. There exists no 'simple' way to extend the 1day-case to an N-day case under the t-distribution, to my knowledge. $\endgroup$ – Kermittfrog Apr 15 at 9:42
  • $\begingroup$ and how about a more 'well behaved' gaussian distribution? or in general, without specifying the distribution line shape? $\endgroup$ – Luigi Apr 15 at 10:06
  • 2
    $\begingroup$ @Luigi You can simply add normally distributed log-returns using the convolution stability of that distribution, i.e. that sum is still normal. The sum of Student $t$ distribution is not Student $t$ though, see here. $\endgroup$ – Kevin Apr 15 at 10:12
  • $\begingroup$ @Kevin, so I should calculate the log-returns for 1 day and then add them n times to get the compounded returns for n days? $\endgroup$ – Luigi Apr 15 at 10:24
5
$\begingroup$

In addition to @Kevin's very helpful comment + link, if you need to calculate the distribution / density of the sum of (any) independent continuous random variables, e.g. sum of Student t distributed vars, or some other mixture, you can go two ways:

Option A: Simulation

Simply simulate the target distribution $N$ number of times. You can then estimate moments, quantiles etc. from that.

N.B: Make sure that $N$ is sufficiently large, e.g. through resampling and testing the tightness of your estimators. For distributions whose moments are not defined, (e.g. with $d.o.f.\leq 4$ for a Student t distribution, if you require first four moments) you might need lots of simulations to get sufficient results.

Option B: Numerical integration of characteristic functions

We know that the characterstic function of a random variable,

$$ \varphi_X(t)\equiv \int f(x)e^{itx}\mathrm{d}x $$

exists always, and that the characteristic function of the sum of independent random variables is the product of their individual characteristic functions. Finally, thru Gil-Pelaez or the inversion relationship,

$$ f(x)=\frac{1}{2\pi}\int e^{-itx}\phi_X(t)\mathrm{d}t, $$

or more exactly through a numerical approximation of this inversion, we can always try to recover the distribution or density of our convoluted variable.

N.B.: The numerical inversion can become quite cumbersome if your distribution is not sufficiently smooth, e.g. if it is has a strong 'bend' somewhere, implying that you need a very large region of integration to accumulate all material frequency contributions.

HTH?

$\endgroup$

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.