0
$\begingroup$

If I am modelling my returns as $\sim N(0, \sigma^2)$, then I can evolve my spot distribution as: $$S_{t} = S_{0}e^{(\mu - \frac{1}{2}\sigma^{2})t + \sigma dW_{t}}$$ where $S_{0}$ is the spot, $\mu$ is the mean , and $\sigma$ is the returns volatility and $dW_{t}$ is the gaussian noise.

How should I amend my Spot (lognormal) distribution if I assuming my returns follow a student-t distribution $\sim t-dist(\nu)$

Thanks

Improve this question
$\endgroup$
2
  • $\begingroup$ Hi, this is a nice question. What do you think about changing the title to "return distribution with ..." ? And could you also make a reference to your desired application, e.g. derivatives valuation, asset allocation, ... ? Depending on the application, there may (not) exist a solution at all... $\endgroup$
    – Kermittfrog
    Dec 1 '20 at 11:54
  • $\begingroup$ Nevertheless, you could find some ideas in here: arxiv.org/pdf/0906.4092.pdf $\endgroup$
    – Kermittfrog
    Dec 1 '20 at 11:57
4
$\begingroup$

1. Theory

The Student $t$ distribution does not exhibit a moment generating function

$$ M_X(t)=\mathbb{E}\left(e^{tX} \right) $$ Hence, there exist no closed form solution for $M_X(t=1)=\mathbb{E}\left(e^X\right)$, i.e. the expected future spot price. Thus, at least theoretically, we are not able to pinpoint the expectation of the future asset value, thereby preventing us from finding a martingal representation.

2. Practical simulation Depending on your application (derivatives pricing, asset allocation), you can of course resort to some arcane methods. I understand that you know how to simulate from a $t$ distribution. Then, in order to properly centre your random variates $x\sim e^{t(\nu)}$, you could empirically re-scale your draws as $\tilde{x}=\frac{x}{\bar{e^x}}$. Again, that will not help you with theory-building, though...

HTH?

Improve this answer
$\endgroup$
13
  • $\begingroup$ Thanks. Wanted to understand a few things (This exercise is to understand the "range" of the FX spot price over a time horizon [t,T] )... 1. If I already have a term-structure of fwd volatility , say \sigma(t), \sigma(t+1)....\ sigma(T) of the returns (fitted using a garch(p,q) model with student-t distribution assumption), how can I use those to come up with a spot price "range" over that period? $\endgroup$
    – sumit_uk1
    Dec 1 '20 at 13:19
  • $\begingroup$ aren't the moments of the t-distribution calculable for a certain range of degrees of freedom $\nu$ $\endgroup$
    – develarist
    Dec 1 '20 at 13:20
  • $\begingroup$ You could come up with quantiles (say from 5% to 95%) of the return distribution and transform those into a price quantiles - and thereby obtain a range. would that help? $\endgroup$
    – Kermittfrog
    Dec 1 '20 at 13:23
  • 1
    $\begingroup$ @develarist The moments, yes. But The MGF itself does not exist, hence the expectation $\mathbb{E}\left(e^{X}\right)$ does not exist. Put differently: The (non-)existence of a MGF does not imply the (non-)existence of moments. $\endgroup$
    – Kermittfrog
    Dec 1 '20 at 13:24
  • $\begingroup$ @Kermittfrog If I fit a parametric distribution, such as the t-distribution, to the histogram of an empirical dataset's samples (using MLE of the parameters), why doesn't the parametric distribution's moments ($\mu, \sigma$) not match the empirical data's moments? If I can't even get the moments to match, doesn't this defeat the whole purpose of modeling real data parametrically $\endgroup$
    – develarist
    Dec 1 '20 at 13:28

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.