Model order selection is often a tricky business. Inspecting the Hankel singular values is one many methods. There are many such methods, arguably one as good or worse as the other.
Specifically in the case AR/ARMA models, you may look into the model residues. Specifically, if you have used a "sufficient" model orders, the residuals should look white. Note that the residuals may look white for many model order selections. There are statistical hypothesis testing procedures (e.g. Chi-2 test, etc.) for both sample auto-correlation of residuals and cross-correlation of residuals and input; Ljung's book has a lot of details on all of this.
Hope that helps.