How to improve ARMA modeling ?

I’m currently modeling a time-series using ARMA model. I am using armax(my_signal,[2 2]).

It fits really well on high frequency but not really well on low frequency (I compared it to the signal using PSD and Allan variance).

Is there an option in armax function (or maybe another solution) for my ARMA model to focus on fitting low frequency well (even if that means fitting badly on high frequency)

Thank you all !

[Kenny Callegari]

Up