You should evaluate the quality of prediction results and perform some correlation analysis of the model residuals to verify how good a a model is.
for some information.
Use PE command to compute the prediction errors. This will show you the final prediction error of your estimated model. Interim errors (while estimation is running) is not something you can plot, but you can view a norm of this error for each iteration by turning the display on, as in armax(data, [na nc], 'display', 'on')
If you want to compare a neural network with ARMA model, compute the prediction errors for each and see whose norm is smaller. For neural network model, I believe you can call the SIM command to get the network response and then subtract it from the measured response to compute the error (provided you are using a prediction network).
