Error terms are calculated when you already have a model, and you calculate them as predicted values - actual values. Now you want to fit MA process to predict future data points (means you still don't have any MA model, so don't have predicted values and thus don't have past forecast errors)
MA model says ...
The model will try to learn parameters theta's, but how do derive these error terms?
(Please refrain heavy mathematical equations to answer this question, if possible, please provide answer with simple example; lets say my time series is 2,5,7,9 ..how the fifth element will be calculated based on MA(1) process?)
