1
$\begingroup$

This is my qq plot : enter image description here

Its concave-convex curve so it indicates light tails. But my mean excess plot :

enter image description here

is increases which means the tail of the distribution of my data is heavy-tailed.

I don't understand why. Some help would be appreciated.

$\endgroup$
12
  • $\begingroup$ Why do you have such huge white space around your plots? I can't read your axes $\endgroup$ – Glen_b Jun 17 '18 at 0:07
  • $\begingroup$ I used the mean excess function in R. in y-axis : The mean excess values and in x-axis : the threshold. $\endgroup$ – user44677 Jun 17 '18 at 0:10
  • $\begingroup$ I have posted versions of your plots with the excess white space removed. You may still need to add additional information to make what you did sufficiently clear. $\endgroup$ – Glen_b Jun 17 '18 at 0:12
  • $\begingroup$ Someone can explain me why are the data points clustering in the mean excess plot and in the QQ-plot too ? thanks in advance $\endgroup$ – user44677 Jun 17 '18 at 1:29
  • 1
    $\begingroup$ Those are examples of things that would give a hard cut off (either by truncation or winsorization) where otherwise the amount recorded may be larger. $\endgroup$ – Glen_b Jun 19 '18 at 9:09
0
$\begingroup$

One can define the mean-excess function to be:

$$e_X(d) = \mathbb E(X - d | X > d )$$

If a distribution has a thin tail, I'd expect this function to get smaller and smaller if I keep increasing $d$. That's exactly what's happening with your function over the threshold $X = 2e+06$.

The ME plot is meant to be interpret after a particular threshold (because you're only interested in the tail of this distribution).

$\endgroup$
1
  • $\begingroup$ So should I use GPD or GEV distribution ? What about the QQplot? thanks in advance $\endgroup$ – user44677 Jun 17 '18 at 15:47

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.