Let say i have a single time series of N observations. I'm wondering how informative are ACF and PACF functions of this series. As we know, they can be used to infer orders of AR and MA part of ARIMA model - there is visual technique to do this.
My question is: Do thease functions (together) have all necesary information to define (with no uncertainty) best ARIMA model (a native model)?
Let say i have some ARMA procesess with different (p,q) randomly generated. Can i train a multi layer perceptron which gets ACF and PACF values as input and outputs p and q in the last layer?
How much is this theoreticaly justified or allowed?