0
$\begingroup$

Short Version

Can I get a quantile of such an expression? \begin{equation} \sum_{k=1}^{n} A_k\exp(\mathcal{N}(t_k\mu-\sigma\sqrt{t_k}/2,\sigma))) \end{equation}

I know I can do it for one part of the summation as stated here, however I would like to know if that can be also the case for the summation.

Long Version

Let's say I have a periodic investment, in a really simplified case, we have the user deposits $A_1$, $A_2$ and $A_3$ as in the following image.

enter image description here

The deposits $A_1$, $A_2$ and $A_3$ happen at $t_1$, $t_2$ and $t_3$ respectively, being the period of time between each transaction $t$. The return of the period of time $t$ is $\mu$ and the volatility is $\sigma$.I want to know how much I have at time $t_3$ with a probability of $90\%$.

My variables are normally random distributed. This means that for example in order to project $A_1$ to the times $t_1$, $t_2$ and $t_3$ we will have the following equations:

\begin{equation} X_{A_1,t_1}\sim A_1\exp(\mathcal{N}(0t\mu-\sigma\sqrt{0t}/2,0\sigma))=A_1\\ X_{A_1,t_2}\sim A_1\exp(\mathcal{N}(t\mu-\sigma\sqrt{1t}/2,\sigma\sqrt{t}))\\ X_{A_1,t_3}\sim A_1\exp(\mathcal{N}(2t\mu-\sigma\sqrt{2t}/2,\sigma\sqrt{2t})) \end{equation}

In the first equation you have exactly the same money because your deposit was it $t_1$, the same time that we analyze. Now, we know that we can calculate a quantile for a normal distribution is $\Phi^{-1}_P=\mu+\sqrt{2}\sigma \operatorname {erf} ^{-1}(2P-1)$ where $P \in [0,1]$, in our case $.9$ because we want to verify $90\%$ and $\operatorname {erf}$ is the error function. Thus, we can calculate the quantile of $A_1$ in $t_3$ as follows:

\begin{equation} \Phi^{-1}_{A_1,t_3|P}= A_1\exp(2t\mu-\sigma\sqrt{2t}/2+\sqrt{2}\sigma\sqrt{2t} \operatorname {erf} ^{-1}(2P-1)) \end{equation}

Some reference to this equation here. My question is, can I sum all the values of the quantiles to $t_3$? or, how can I obtain the values of the quantiles at $t_3$?

I tried the following equation:

\begin{equation} \Phi^{-1}_{t_3|P}=\Phi^{-1}_{A_1,t_3|P}+\Phi^{-1}_{A_2,t_3|P}+\Phi^{-1}_{A_3,t_3|P} \end{equation}

As shown before, the last term of the quation is just $A_3$ but the others depend on the time and the quantile. This equation is not right, I have some dissimilarities with a montecarlo simulation I did to verify the results.

I'm almost certain I must not sum quantiles the way I did, but I cannot find the proper resource to verify the equations.

$\endgroup$
  • $\begingroup$ Since you have put some effort in this question I'll answer it although it appears to me as "too basic" for this site, no offense meant. $\endgroup$ – Quantuple Aug 2 '17 at 13:50
  • $\begingroup$ no offense taken, I'm new, I want to learn =) $\endgroup$ – silgon Aug 2 '17 at 14:48
2
$\begingroup$

The random variable $$X = \sum_{k=1}^{n} A_k\exp(\mathcal{N}(t_k\mu-\sigma\sqrt{t_k}/2,\sigma)))$$ emerges as a weighted sum of individual random variables that are log-normally distributed.

Unfortunately, even if we assume that the individual r.v. involved in the sum are independent, a sum of log-normals (here $X$) possesses no analytically tractable probability density function.

You will therefore not be able to find a closed-form expression for the quantiles of $X$.

A practical solution however could be to use a "Monte Carlo" simulation in the loose sense: draw a (sufficient) number $M$ of random samples from $X$, $(x_i)_{i=1,...,M}$; sort them in ascending order and infer the desired quantile

These related questions (here and there) may also help I guess.

$\endgroup$
  • $\begingroup$ thanks, both links are interesting, i'll check it out, I'm solving it with montecarlo, but I wanted an analytical solution to speed up the process $\endgroup$ – silgon Aug 2 '17 at 14:51

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.