6
$\begingroup$

The data I am using can be found here: http://uploadeasy.net/upload/nwv0.rar The variable is called "alvsloss".

I want to fit a distribution to my financial data. First of all, I started with a normal distribution.

The pdf is given by:

$\frac{1}{\sqrt{2\pi\sigma^2}}\operatorname{exp}\left\{-\frac{\left(x-\mu\right)^2}{2\sigma^2}\right\}$

and the log likelihood, which is to be maximized is given by

$ \ln\mathcal{L}(\mu,\sigma^2) = \sum_{i=1}^n \ln f(x_i;\,\mu,\sigma^2) = -\frac{n}{2}\ln(2\pi) - \frac{n}{2}\ln\sigma^2 - \frac{1}{2\sigma^2}\sum_{i=1}^n (x_i-\mu)^2. $

which gives as a solution

$ \hat{\mu} = \overline{x} \equiv \frac{1}{n}\sum_{i=1}^n x_i, \qquad \hat{\sigma}^2 = \frac{1}{n} \sum_{i=1}^n (x_i - \overline{x})^2. $

I can calculate this in R via mean(alvsloss) and sd(alvsloss) and get the values -0.0004270872 and 0.02304159 or I do

library(MASS)
fitdistr(alvsloss, "normal")

and get the solution

       mean             sd      
  -0.0004270872    0.0230371244 
 ( 0.0004535429) ( 0.0003207033)

Ok, I understand this. But now my problem arises, I tried this with the t distribution. The pdf is given by:

$\textstyle\frac{\Gamma \left(\frac{\nu+1}{2} \right)} {\sqrt{\nu\pi}\,\Gamma \left(\frac{\nu}{2} \right)} \left(1+\frac{x^2}{\nu} \right)^{-\frac{\nu+1}{2}}\!$

the log likelihood can be found here: http://de.scribd.com/doc/52587265/80/STUDENT%E2%80%99S-t-DISTRIBUTION formula (17.8) (x is denoted by $\epsilon$. I am not understanding where the $h_i$ comes from in the formula, but this is not my main question.)

First question: Seems to me, that there is no closed form solution?

Second question: I do fitdistr(alvsloss, "t")

and get the output

  m               s               df      
  -0.0004919768    0.0130128873    2.6340459185 
 ( 0.0003182568) ( 0.0003453702) ( 0.1620424078)
Es gab 28 Warnungen (Anzeige mit warnings())

I am not understanding, why I have three parameters now? The df is the $\nu$, but the others? What is R doing? Since there is no closed form solution, I guess it is using a iterated algorithm? Should I bother about the 28 warnings? They are:

Warnmeldungen:

1: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
2: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
3: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
4: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
5: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
6: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
7: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
8: In log(s) : NaNs wurden erzeugt
9: In log(s) : NaNs wurden erzeugt
10: In log(s) : NaNs wurden erzeugt
11: In log(s) : NaNs wurden erzeugt
12: In log(s) : NaNs wurden erzeugt
13: In log(s) : NaNs wurden erzeugt
14: In log(s) : NaNs wurden erzeugt
15: In log(s) : NaNs wurden erzeugt
16: In log(s) : NaNs wurden erzeugt
17: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
18: In log(s) : NaNs wurden erzeugt
19: In log(s) : NaNs wurden erzeugt
20: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
21: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
22: In dt((x - m)/s, df, log = TRUE) : NaNs wurden erzeugt
23: In log(s) : NaNs wurden erzeugt
24: In log(s) : NaNs wurden erzeugt
25: In log(s) : NaNs wurden erzeugt
26: In log(s) : NaNs wurden erzeugt
27: In log(s) : NaNs wurden erzeugt
28: In log(s) : NaNs wurden erzeugt

And how can I plot this distribution with the estimated parameters?

$\endgroup$
  • 2
    $\begingroup$ Your file is 20K. Why on earth would you save it as binary and compress it as rar when you could simply upload an ascii or CSV file and save people a lot of mucking about? In fact, the best thing to do would be to grab a small subset of your data with the same issues and put the data in your question, so your question doesn't break when that file becomes inaccessible. $\endgroup$ – Glen_b -Reinstate Monica Mar 19 '13 at 1:56
7
$\begingroup$

The estimators correspond to $(\mu,\sigma,\nu)$, the parameters of a Student-$t$ distribution with location parameter $\mu\in{\mathbb R}$, scale parameter $\sigma>0$ and $\nu>0$ degrees of freedom. This density is simply given by

$$f(x;\mu,\sigma,\nu)=\dfrac{1}{\sigma}\dfrac{\Gamma \left(\frac{\nu+1}{2} \right)} {\sqrt{\nu\pi}\,\Gamma \left(\frac{\nu}{2} \right)} \left(1+\dfrac{(x-\mu)^2}{\nu \sigma^2} \right)^{-\frac{\nu+1}{2}}.$$

There seems to be no closed expression for these estimators.

The warnings in this case are harmless. You can check this by finding the MLE using the command optim as follows

# log-likelihood function
loglik <-function(par){
if(par[2]>0 & par[3]>0) return(-sum(log(dt((alvsloss-par[1])/par[2],df=par[3])/par[2])))
else return(Inf)
}

# optimisation step
optim(c(0,0.1,2.5),loglik)

The following code shows how to plot the fitted density together with a kernel density estimator.

# fitted density
param = optim(c(0,0.01,2.5),loglik)$par
fit.den <- function(x) dt((x-param[1])/param[2],df=param[3])/param[2]

curve(fit.den,-0.15,0.15)
points(density(alvsloss),type="l",col="red")
| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ So many novelists, singers and artists have been visiting this site (with good answers!) of late. It's eerie! $\endgroup$ – cardinal Mar 18 '13 at 20:43
  • $\begingroup$ @Carlos Gardel thanks for your answer, but I don't get the following: You are doing the log-likelihood maximization manually, but -sum(log(dt((alvsloss-par[1])/par[2],df=par[3])/par[2])) does not look like the log likelihood function of the t distribution? $\endgroup$ – Stat Tistician Mar 19 '13 at 7:36
  • $\begingroup$ or why do you use alvsloss-par[1])/par[2] and df=par[3])/par[2] ? $\endgroup$ – Stat Tistician Mar 19 '13 at 7:51

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.