Generalized Pareto Distribution - shape parameter estimation problem

Someone of you has noticed any problem with the shape parameter estimation for the Generalized Pareto Distribution? In particular when this latter is negative. I used the GPDFIT function to fit the Generalized Pareto Distribution on financial asset returns over a threshold but I found that when the shape parameter is negative enough, some values from the observations don't stay in the support of the distribution. Any idea?

[dsmalenb]

I am running into similar issues. I have an active question here:

https://www.mathworks.com/matlabcentral/answers/410467-inconsistent-results-w-fitdist-for-generalized-pareto

Take a look.