Re: How future predictions can be made with a Time Series model in RapidMiner?
The nice thing about prediction operators like svm and neural nets is that they are multivariable.
In stock trading terms: Amplitudes of the Moving Average and trading volumes have probably a corrolation.
ARIMA is univariable but the only operator able to predict a real future.
What I am going to do to enable multivariable future predictions is:
To feed the multivariable prediction operator with real multivariable data and adjectently all of their univariable related predictions, the prediction output of an ARIMA model. I will train that model with real data. Yes, therefore I have to wait untill the future is past and I have obtained the labels to train to. Yes, I know, the resulting prediction will have a lag. The label data cannot be newer than now(). We all don't have real multivariable data from the future. But one can optimize a prediction.
What happens if q,d,p used in ARIMA change? Well, I guess that the multivariable prediction operator will get improved data to train its model untill now() with training data for the future minus the horizon. It is and will be always the future we want to predict.We have to make a guess. We ask therefore ARIMA a prediction, it's ARIMA's best guess. The multivariable prediction operator will train on it with a target label until now() aka prediction horizon minus horizon.