I’m trying to use a ML extended method to fit some data. I use a Crystal Ball for the signal and a Berstein polynomial for the background (order 4) The signal is ok well but i don’t understand why the background is flat.
It seems you run into some fit stability issues, for which I see two causes:
You use a Bernstein polynomial with forced positive coefficients. While
the ensures a positive defined pdf it does sometimes have some issues
with the final fitted shapes ends up having several coefficients consistent
with zero, which are then all at the allowed parameter boundary. If I e.g. switch
your background pdf to a Chebychev I see less problems. In your case
something like a RooExponential might also work well with your background shape,
and have none these issues typical to polynomials
If you perform a fit with a Range() command, the yields nsig and nbkg
remain expressed in the full range. In your model this stongly increases
the correlation between various parameters. An alternate way to do the
fit